Morgan
Cosponsors:
Bank of America
Bank of Montreal
BZW
Deutsche Morgan Grenfell
KMV Corporation
Swiss Bank Corporation
Union Bank of Switzerland
CreditMetrics
™
– Technical Document
• A valueatrisk (VaR) framework applicable to all institutions worldwide that carry
credit risk in the course of their business.
• A full portfolio view addressing credit event correlations which can identify the costs of
over concentration and benefits of diversification in a marktomarket framework.
• Results that drive: investment decisions, riskmitigating actions, consistent riskbased
credit limits, and rational riskbased capital allocations.
This Technical Document describes CreditMetrics™, a framework for quantifying credit risk
in portfolios of traditional credit products (loans, commitments to lend, financial letters of
credit), fixed income instruments, and marketdriven instruments subject to counterparty
default (swaps, forwards, etc.). This is the first edition of what we intend will be an ongoing
refinement of credit risk methodologies.
Just as we have done with RiskMetrics™, we are making our methodology and data
available for three reasons:
1. We are interested in promoting greater transparency of credit risk. Transparency is the
key to effective management.
2. Our aim is to establish a benchmark for credit risk measurement. The absence of a com
mon point of reference for credit risk makes it difficult to compare different approaches
to and measures of credit risk. Risks are comparable only when they are measured with
the same yardstick.
3. We intend to provide our clients with sound advice, including advice on managing their
credit risk. We describe the CreditMetrics™methodology as an aid to clients in under
standing and evaluating that advice.
Both J.P. Morgan and our cosponsors are committed to further the development of
CreditMetrics™as a fully transparent set of risk measurement methods. This broad sponsor
ship should be interpreted as a signal of our joint commitment to an open and evolving stan
dard for credit risk measurement. We invite the participation of all parties in this continuing
enterprise and look forward to receiving feedback to enhance CreditMetrics™as a bench
mark for measuring credit risk.
CreditMetrics™is based on, but differs significantly from, the risk measurement methodolo
gy developed by J.P. Morgan for the measurement, management, and control of credit risk in
its trading, arbitrage, and investment account activities. We remind our readers that no
amount of sophisticated analytics will replace experience and professional judgment in
managing risks. CreditMetrics™is nothing more than a highquality tool for the profes
sional risk manager in the financial markets and is not a guarantee of specific results.
The benchmark for understanding credit risk
New York
April 2, 1997
CreditMetrics™ — Technical Document
CreditMetrics™—Technical Document
Copyright © 1997 J.P. Morgan & Co. Incorporated.
All rights reserved.
CreditMetrics
™
is a trademark of J.P. Morgan in the United States and in other countries. It is written
with the symbol
™
at its ﬁrst occurance in the publication, and as CreditMetrics thereafter.
iii
This book
Authors:
Greg M. Gupton
Morgan Guaranty Trust Company
Risk Management Research
(1212) 6488062
gupton_greg@jpmorgan.com
Christopher C. Finger
Morgan Guaranty Trust Company
Risk Management Research
(1212) 6484657
ﬁnger_christopher@jpmorgan.com
Mickey Bhatia
Morgan Guaranty Trust Company
Risk Management Research
(1212) 6484299
bhatia_mickey@jpmorgan.com
This is the reference document for CreditMetrics™. It is meant to serve as an introduc
tion to the methodology and mathematics behind statistical credit risk estimation, as well
as a detailed documentation of the analytics that generate the data set we provide.
This document reviews:
• the conceptual framework of our methodologies for estimating credit risk;
• the description of the obligors’ credit quality characteristics, their statistical descrip
tion and associated statistical models;
• the description of credit exposure types across “marketdriven” instruments and the
more traditional corporate ﬁnance credit products; and
• the data set that we update periodically and provide to the market for free.
In the interest of establishing a benchmark in a ﬁeld with as little standardization and
precise data as credit risk measurement, we have invited ﬁve leading banks, Bank of
America, BZW, Deutsche Morgan Grenfell, Swiss Bank Corporation, and Union Bank of
Switzerland, and a leading credit risk analytics ﬁrm, KMV Corporation, to be cospon
sors of CreditMetrics. All these ﬁrms have spent a signiﬁcant amount of time working on
their own credit risk management issues, and we are pleased to have received their input
and support in the development of CreditMetrics. With their sponsorship we hope to
send one clear and consistent message to the marketplace in an area with little clarity to
date.
We have also had many fruitful dialogues with professionals from Central Banks, regula
tors, competitors, and academics. We are grateful for their insights, help, and encourage
ment. Of course, all remaining errors and omissions are solely our responsibility.
How is this related to RiskMetrics
™
?
We developed CreditMetrics to be as good a methodology for capturing counterparty
default risk as the available data quality would allow. Although we never mandated dur
ing this development that CreditMetrics must resemble RiskMetrics, the outcome has
yielded philosophically similar models. One major difference in the models was driven
by the difference in the available data. In RiskMetrics, we have an abundance of daily
liquid pricing data on which to construct a model of conditional volatility. In Credit
Metrics, we have relatively sparse and infrequently priced data on which to construct a
model of unconditional volatility.
What is different about CreditMetrics?
Unlike market risks where daily liquid price observations allow a direct calculation of
valueatrisk (VaR), CreditMetrics seeks to
construct
what it cannot directly
observe
: the
volatility of value due to credit quality changes. This constructive approach makes
CreditMetrics less an exercise in ﬁtting distributions to observed price data, and more an
exercise in proposing models which explain the changes in credit related instruments.
iv Preface
CreditMetrics™—Technical Document
And as we will mention many times in this document, the models which best describe
credit risk do not rely on the assumption that returns are normally distributed, marking a
signiﬁcant departure from the RiskMetrics framework.
In the end, we seek to balance the best of all sources of information in a model which
looks across broad historical data rather than only recent market moves and across the
full range of credit quality migration — upgrades and downgrades — rather than just
default.
Our framework can be described in the diagram below. The many sources of informa
tion may give an impression of complexity. However, we give a stepbystep introduc
tion in the ﬁrst four chapters of this book which should be accessible to all readers.
One of our fundamental techniques is
migration analysis
, that is, the study of changes in
the credit quality of names through time. Morgan developed transition matrices for this
purpose as early as 1987. We have since built upon a broad literature of work which
applies migration analysis to credit risk evaluation. The ﬁrst publication of transition
matrices was in 1991 by both Professor Edward Altman of New York University and sep
arately by Lucas & Lonski of Moody’s Investors Service. They have since been pub
lished regularly (see Moody’s Carty & Lieberman [96a]
1
and Standard & Poor’s
Creditweek
[15Apr96]) and are also calculated by ﬁrms such as KMV.
Are RiskMetrics and CreditMetrics comparable?
Yes, in brief, RiskMetrics looks to a horizon and estimates the
valueatrisk
across a dis
tribution of historically estimated realizations. Likewise, CreditMetrics looks to a hori
zon and constructs a distribution of historically estimated credit outcomes (rating
migrations including potentially default). Each credit quality migration is weighted by
its likelihood (transition matrix analysis). Each outcome has an estimate of change in
value (given by either credit spreads or studies of recovery rates in default). We then
aggregate volatilities across the portfolio, applying estimates of correlation. Thus,
although the relevant time horizon is usually longer for credit risk, with CreditMetrics
we compute credit risk on a comparable basis with market risk.
1
Bracketed numbers refer to year of publication.
Credit Rating Seniority Credit Spreads
Value at Risk due to Credit
bond revaluation
Present value
quality changes for a single exposure
Standard Deviation of value due to credit
Ratings series,
Equities series
Correlations
Models (e.g.,
correlations)
Portfolio Value at Risk due to Credit
Exposures
Market
volatilities
Exposure
distributions
User
Portfolio
Rating migration
likelihoods
Joint credit
rating changes
in default
Recovery rate
Preface v
What CreditMetrics is not
We have sought to add value to the market’s understanding of credit risk estimation, not
by replicating what others have done before, but rather by ﬁlling in what we believe is
lacking. Most prior work has been on the estimation of the relative likelihoods of default
for individual ﬁrms; Moody’s and S&P have long done this and many others have started
to do so. We have designed CreditMetrics to accept as an input any assessment of default
probability
2
which results in ﬁrms being classiﬁed into discrete groups (such as rating
categories), each with a deﬁned default probability. It is important to realize, however,
that these assessments are only inputs to CreditMetrics, and not the ﬁnal output.
We wish to estimate the
volatility of value
due to changes in credit quality, not just the
expected loss
. In our view, as important as default likelihood estimation is, it is only one
link in the long chain of modeling and estimation that is necessary to fully assess credit
risk (volatility) within a portfolio. Just as a chain is only as strong as its weakest link, it
is also important to diligently address: (i) uncertainty of exposure such as is found in
swaps and forwards, (ii) residual value estimates and their
uncertainties
, and (iii) credit
quality
correlations
across the portfolio.
How is this document organized?
One need not read and fully understand the details of this entire document to understand
CreditMetrics. This document is organized into three parts that address subjects of par
ticular interest to our diverse readers.
Part I Risk Measurement Framework
This section is for the general practitioner. We provide a practicable
framework of how to think about credit risk, how to apply that thinking in
practice, and how to interpret the results. We begin with an example of a
single bond and then add more variation and detail. By example, we
apply our framework across different exposures and across a portfolio.
Part II Model Parameters
Although this section occasionally refers to advanced statistical analysis,
there is content accessible to all readers. We ﬁrst review the current aca
demic context within which we developed our credit risk framework. We
review the statistical assumptions needed to describe discrete credit
events; their mean expectations, volatilities, and correlations. We then
look at how these credit statistics can be estimated to describe what hap
pened in the past and what can be projected in the future.
Part III Applications
We discuss two implementations of our portfolio framework for estimat
ing the
volatility of value due to credit quality changes
. The ﬁrst is an
analytic calculation of the mean and standard deviation of value changes.
The second is a simulation approach which estimates the full distribution
of value changes. These both embody the same modeling framework and
2
These assessments may be agency debt ratings, a user’s internal ratings, the output of a statistical default predic
tion model, or any other approach.
vi Preface
CreditMetrics™—Technical Document
produce comparable results. We also discuss how the results can be used
in portfolio management, limit setting, and economic capital allocation.
Future plans
We expect to update this
Technical Document
regularly. We intend to further develop
our methodology, data and software implementation as we receive client and academic
comments.
CreditMetrics has been developed by the Risk Management Research Group at
J.P. Morgan. Special mention must go to Greg M. Gupton who conceived of this project
and has been working on developing the CreditMetrics approach at JPMorgan for the last
four years. We welcome any suggestions to enhance the methodology and adapt it fur
ther to the changing needs of the market. We encourage academic studies and are pre
pared to supply data for wellstructured projects.
Acknowledgments
We would like to thank our cosponsors for their input and support in the writing and
editing of this document. In particular, we thank the KMV Corporation, which has been
a pioneer in developing portfolio approaches to credit risk, and whose work has inﬂu
enced many of the methods presented here.
We thank numerous individuals at J.P. Morgan who participated in this project, as well
as professionals at other banks and academic institutions who offered input at various
levels. Also, this document could not have been produced without the contributions of
our consulting editor, Margaret Dunkle. We apologize for any omissions.
vii
Table of Contents
Part I Risk Measurement Framework
Chapter 1. Introduction to CreditMetrics 5
1.1 The portfolio context of credit 5
1.2 Types of risks modeled 8
1.3 Modeling the distribution of portfolio value 8
1.4 Different credit risk measures 15
1.5 Exposure type differences 17
1.6 Data issues 20
1.7 Advanced modeling features 21
Chapter 2. Standalone risk calculation 23
2.1 Overview: Risk for a standalone exposure 23
2.2 Step #1: Credit rating migration 24
2.3 Step #2: Valuation 26
2.4 Step #3: Credit risk estimation 28
2.5 Choosing a time horizon 31
Chapter 3. Portfolio risk calculation 35
3.1 Joint probabilities 36
3.2 Portfolio credit risk 38
3.3 Marginal risk 40
Chapter 4. Differing exposure types 41
4.1 Receivables 42
4.2 Bonds and loans 43
4.3 Loan commitments 43
4.4 Financial letters of credit (LCs) 46
4.5 Marketdriven instruments 47
Part II Model Parameters
Chapter 5. Overview of credit risk literature 57
5.1 Expected losses 57
5.2 Unexpected losses 60
5.3 A portfolio view 63
Chapter 6. Default and credit quality migration 65
6.1 Default 65
6.2 Credit quality migration 66
6.3 Historical tabulation 67
6.4 Longterm behavior 70
Chapter 7. Recovery rates 77
7.1 Estimating recovery rates 77
7.2 Distribution of recovery rate 80
Chapter 8. Credit quality correlations 81
8.1 Finding evidence of default correlation 81
viii Table of contents
CreditMetrics™ —Technical Document
8.2 Direct estimation of joint credit moves 83
8.3 Estimating credit quality correlations through bond spreads 84
8.4 Asset value model 85
8.5 Estimating asset correlations 92
Part III Applications
Chapter 9. Analytic portfolio calculation 107
9.1 Threeasset portfolio 107
9.2 Marginal standard deviation 110
Chapter 10. Simulation 113
10.1 Scenario generation 113
10.2 Portfolio valuation 116
10.3 Summarizing the results 117
Chapter 11. Portfolio example 121
11.1 The example portfolio 121
11.2 Simulation results 122
11.3 Assessing precision 125
11.4 Marginal risk measures 129
Chapter 12. Application of model outputs 133
12.1 Prioritizing risk reduction actions 133
12.2 Credit risk limits 135
12.3 Economic capital assessment 138
12.4 Summary 140
Appendices
Appendix A. Analytic standard deviation calculation 147
Appendix B. Precision of simulationbased estimates 149
Appendix C. Derivation of the product of N random variables 153
Appendix D. Derivation of risk across mutually exclusive outcomes 155
Appendix E. Derivation of the correlation of two binomials 157
Appendix F. Inferring default correlations from default volatilities 159
Appendix G. International bankruptcy code summary 161
Appendix H. Model inputs 163
Appendix I. Indices used for asset correlations 166
Reference
Glossary of terms 173
Bibliography 183
Index 191
ix
List of Tables
Table 1.1 Probability of credit rating migrations in one year for a BBB 9
Table 1.2 Calculation of yearend values after credit rating migration from BBB ($) 10
Table 1.3 Distribution of value of a BBB par bond in one year 11
Table 1.4 Yearend values after credit rating migration from singleA ($) 12
Table 1.5 All possible 64 yearend values for a twobond portfolio ($) 12
Table 1.6 Probability of credit rating migrations in one year for a singleA 13
Table 1.7 Yearend joint likelihoods (probabilities) across 64 different states (%) 14
Table 1.8 Oneyear transition matrix (%) 20
Table 2.1 Oneyear transition matrix (%) 25
Table 2.2 Recovery rates by seniority class (% of face value, i.e., “par”) 26
Table 2.3 Example oneyear forward zero curves by credit rating category (%) 27
Table 2.4 Possible oneyear forward values for a BBB bond plus coupon 28
Table 2.5 Calculating volatility in value due to credit quality changes 28
Table 3.1 Joint migration probabilities with zero correlation (%) 36
Table 3.2 Joint migration probabilities with 0.30 asset correlation (%) 38
Table 4.1 Fee on undrawn portion of commitment (b.p.) 43
Table 4.2 Average usage of commitments to lend 45
Table 4.3 Example estimate of changes in drawdown 45
Table 4.4 Revaluations for $20mm initially drawn commitment 46
Table 4.5 Value of swap at the risk horizon in each rating state 51
Table 5.1 Moody’s corporate bond average cumulative default rates (%) 58
Table 5.2 Credit quality migration likelihoods for a BBB in one year 60
Table 5.3 Volatility of historical default rates by rating category 61
Table 6.1 Moody’s Investors Service: Oneyear transition matrix 68
Table 6.2 Standard & Poor’s oneyear transition matrix – adjusted for removal of N.R. 69
Table 6.3 KMV oneyear transition matrices as tabulated from expected default frequencies70
Table 6.4 Average cumulative default rates (%) 71
Table 6.5 Imputed transition matrix which best replicates default rates 72
Table 6.6 Resulting cumulative default rates from imputed transition matrix (%) 73
Table 6.7 “BB barrier” probabilities calculated from
Table 6.6
matrix (%) 73
Table 6.8 “BB barrier” probabilities calculated from
Table 6.6
matrix (%) 74
Table 6.9 Imputed transition matrix with default rate rank order constraint 74
Table 6.10 Estimate of debt market proﬁle across credit rating categories 75
Table 6.11 Achieving a closer ﬁt to the longterm steady state proﬁle 76
Table 7.1 Recovery statistics by seniority class 78
Table 8.1 Inferred default correlations with conﬁdence levels 82
Table 8.2 Historically tabulated joint credit quality comovements 84
Table 8.3 Historically tabulated joint credit quality comovement (%) 84
Table 8.4 One year transition probabilities for a BB rated obligor 87
Table 8.5 Threshold values for asset return for a BBB rated obligor 88
Table 8.6 Transition probabilities and asset return thresholds for A rating 89
Table 8.7 Joint rating change probabilities for BB and A rated obligors (%) 90
Table 8.8 Countries and respective index families 94
Table 8.9 Industry groupings with codes 95
Table 8.10 Countryindustry index availability 96
Table 8.11 Volatilities and correlations for countryindustry pairs 98
Table 9.1 Transition probabilities (%) 107
Table 9.2 Instrument values in future ratings ($mm) 108
Table 9.3 Values of a twoasset portfolio in future ratings ($mm) 109
Table 10.1 Transition probabilities (%) 114
Table 10.2 Asset return thresholds 114
Table 10.3 Correlation matrix for example portfolio 115
x List of tables
CreditMetrics™—Technical Document
Table 10.4 Scenarios for standardized asset returns 115
Table 10.5 Mapping return scenarios to rating scenarios 116
Table 10.6 Valuation of portfolio scenarios ($mm) 117
Table 11.1 Example portfolio 121
Table 11.2 Asset correlations for example portfolio 122
Table 11.3 Percentiles of future portfolio values ($mm) 125
Table 11.4 Portfolio value statistics with 90% conﬁdence levels ($mm) 126
Table 11.5 Standard deviation of value change 130
Table G.1 Summary of international bankruptcy codes 161
Table H.1 Required inputs for each issuer 165
Table H.2 Required inputs for each exposure type 165
xi
List of Charts
Chart 1.1 Comparison of distribution of credit returns and market returns 7
Chart 1.2 Distribution of value for a 5year BBB bond in one year 11
Chart 1.3 Distribution of value for a portfolio of two bonds 14
Chart 2.1 Our ﬁrst “road map” of the analytics within CreditMetrics 23
Chart 2.2 Examples of credit quality migrations (oneyear risk horizon) 24
Chart 3.1 Our second “road map” of the analytics within CreditMetric
s 35
Chart 3.2 Model of firm value and its default threshold 37
Chart 3.3 Model of ﬁrm value and generalized credit quality thresholds 37
Chart 4.1 Our ﬁnal “road map” of the analytics within Credit
Metrics 41
Chart 5.1 Credit migration 60
Chart 5.2 Construction of volatility across credit quality categories 63
Chart 6.1 Model of ﬁrm value and migration 67
Chart 6.2 Achieving a closer ﬁt to the longterm steady state proﬁle 75
Chart 7.1 Distribution of bank facility recoveries 79
Chart 7.2 Example beta distributions for seniority classes 80
Chart 8.1 Credit rating migration driven by underlying BB ﬁrm asset value 86
Chart 8.2 Distribution of asset returns with rating change thresholds 88
Chart 8.3 Probability of joint defaults as a function of asset return correlation 91
Chart 8.4 Translation of equity correlation to default correlation 92
Chart 10.1 Frequency plot of portfolio scenarios 118
Chart 11.1 Histogram of future portfolio values – upper 85% of scenarios 123
Chart 11.2 Histogram of future portfolio values – scenarios
between 95th and 65th percentiles 123
Chart 11.3 Histogram of future portfolio values – lower 5% of scenarios 124
Chart 11.4 Evolution of confidence bands for portfolio mean ($mm) 126
Chart 11.5 Evolution of confidence bands for standard deviation ($mm) 127
Chart 11.6 Evolution of confidence bands for 5th percentile ($mm) 127
Chart 11.7 Evolution of conﬁdence bands for 1st percentile ($mm) 128
Chart 11.8 Evolution of confidence bands for 0.5 percentile ($mm) 128
Chart 11.9 Evolution of confidence bands for 0.1 percentile ($mm) 129
Chart 11.10 Marginal risk versus current value for example portfolio 131
Chart 12.1 Risk versus size of exposures within a typical credit portfolio 134
Chart 12.2 Possible risk limits for an example portfolio 135
xii List of charts
CreditMetrics™—Technical Document
1
Part I
Risk Measurement Framework
2
CreditMetrics™—Technical Document
3
Part I: Risk Measurement Framework
Overview of Part I
This section describes the risk measurement framework used in CreditMetrics. We
emphasize the basic ideas and illustrate them by means of simple examples. Later in
Parts II and III, we give a more detailed treatment of CreditMetrics including methodol
ogy and data issues.
Part I is organized into the following four chapters:
•
Chapter 1: Introduction to CreditMetrics.
In this chapter, we discuss the merits
and challenges of pursuing a quantitative portfolio approach to measuring credit risk.
We give a summary of what we hope to achieve and the scope of our application.
Using simple examples of one and twobond portfolios, we explain the ideas, meth
odology and data requirements of CreditMetrics.
•
Chapter 2: Standalone risk calculation.
In this chapter, we provide details of
how CreditMetrics estimates credit risk for a single bond. We discuss how we
directly calculate the standard deviation of value due to credit quality changes.
•
Chapter 3: Portfolio risk calculation.
In this chapter, we extend the credit risk
calculation to a portfolio containing two bonds and introduce the notion of
correla
tions,
which will be central to our treatment of risk at the portfolio level. Again, we
illustrate the credit risk calculation for this portfolio with the help of a simple exam
ple. This twobond “portfolio” will serve to illustrate all the methodology we need
to calculate credit risk across a portfolio of any size.
•
Chapter 4: Differing exposure types.
For simplicity, we have limited our discus
sion in the previous two chapters to bonds. In this chapter, we discuss how Credit
Metrics addresses other instruments such as: receivables, loans, commitments to
lend, ﬁnancial letters of credit, swaps and forwards. We emphasize that our risk
modeling framework is general, and we discuss the data necessary to extend it to
other exposure types.
4
CreditMetrics™—Technical Document
5
Chapter 1. Introduction to CreditMetrics
CreditMetrics is a tool for assessing portfolio risk due to changes in debt value caused by
changes in obligor credit quality. We include changes in value caused not only by possi
ble default events, but also by upgrades and downgrades in credit quality. Also, we
assess the valueatrisk (VaR) – the volatility of value – not just the expected losses.
Importantly, we assess risk within the full context of a portfolio. We address the correla
tion of credit quality moves across obligors. This allows us to directly calculate the
diversiﬁcation beneﬁts or potential overconcentrations across the portfolio.
For example, suppose we invest in a bond. Credit risk arises because the bond’s value in
one year can vary depending on the credit quality of its issuer. In general, we know that
the value of this bond will decline with a downgrade or default of its issuer – and appre
ciate if the credit quality of the obligor improves. Value changes will be relatively small
with minor up(down)grades, but could be substantial – 50% to 90% are common – if
there is a default. This is far from the more
normally distributed
market risks that VaR
models typically address.
In this chapter, we step through our CreditMetrics methodology and data in a survey
fashion to give the broad picture of what we hope to achieve. Speciﬁcally, we will:
• establish the link between the process of credit quality migration and the resulting
changes in debt value;
• illustrate the resulting risk assessment with the simple example of a single bond;
• discuss the beneﬁts and challenges to a portfolio approach and use a twobond exam
ple to show how we address a full portfolio;
• extend our focus to speciﬁc credit instruments other than bonds; and
• summarize the data required for any credit instrument.
The result of our efforts will be measures of valueatrisk due to credit quality changes.
These measures will assist in the evaluation, deployment and management of credit risk
taking across both a portfolio and marginal transactions. These measures are consistent
with the – perhaps more familiar – valueatrisk models which are used for market risks.
1.1 The portfolio context of credit
Credit risk has become perhaps the key risk management challenge of the late 1990s.
Globally, institutions are taking on an increasing amount of credit risk. As credit expo
sures have multiplied, the need for more sophisticated risk management techniques for
credit risk has also increased.
Of course, credit risk can be managed – as it has been – by more rigorous enforcement of
traditional credit processes such as stringent underwriting standards, limit enforcement
and counterparty monitoring. However, risk managers are increasingly seeking to quan
tify and integrate the overall credit risk assessment within a VaR statement which cap
tures exposure to market, rating change, and default risks.
6 Chapter 1. Introduction to CreditMetrics
CreditMetrics™—Technical Document
In the end, a better understanding of the credit portfolio will help portfolio managers to
better identify pockets of concentration and opportunities for diversiﬁcation. Over time,
positions can be taken to best utilize risktaking capacity – which is a scarce and costly
resource. Managers can then make risk versus return tradeoffs with knowledge of not
only the expected credit losses, but also the uncertainty of loss.
1.1.1 The need for a portfolio approach
The primary reason to have a quantitative portfolio approach to credit risk management
is so that we can more systematically address
concentration risk
. Concentration risk
refers to additional portfolio risk resulting from increased exposure to one obligor or
groups of correlated obligors (e.g., by industry, by location, etc.).
Traditionally, portfolio managers have relied on a qualitative feel for the concentration
risk in their credit portfolios. Intuitive – but arbitrary – exposurebased credit limits
have been the primary defense against unacceptable concentrations of credit risk. How
ever, ﬁxed exposure limits do not recognize the relationship between risk and return.
A more quantitative approach such as that presented here allows a portfolio manager to
state credit lines and limits in units of marginal portfolio volatility. Furthermore, such a
model creates a framework within which to consider concentrations along almost any
dimension (industry, sector, country, instrument type, etc.).
Another important reason to take a portfolio view of credit risk is to more rationally and
accountably address portfolio diversiﬁcation. The decision to take on ever higher expo
sure to an obligor will meet ever higher marginal risk – risk that grows geometrically
with the concentration on that name. Conversely, similar additional exposure to an
equally rated obligor who has relatively little existing exposure will entail less risk.
Indeed, such names may be individually risky, but offer a relatively small marginal con
tribution to overall portfolio risk due to diversiﬁcation beneﬁts.
Finally, by capturing portfolio effects (diversiﬁcation beneﬁts and concentration risks)
and recognizing that risk accelerates with declining credit quality, a portfolio credit risk
methodology can be the foundation for a rational riskbased capital allocation process.
1
There are also more practical reasons for a more quantitative approach to credit risk:
• Financial products have become more complex. The growth of derivatives activity
has created uncertain and dynamic counterparty exposures that are signiﬁcantly
more challenging to manage than the static exposures of more traditional instruments
such as bonds or loans. Endusers and providers of these instruments need to identify
such exposures and understand their credit, as well as related market, risks.
1
A capital measure reﬂecting these economic factors is a fundamental departure from the capital adequacy mea
sures mandated for bank regulation by the Bank for International Settlements ("BIS"). The BIS riskbased capital
guidelines do not distinguish high quality and welldiversiﬁed portfolios from low quality and concentrated port
folios. Some bank regulators, recognizing that the BIS regulatory capital regime can create uneconomic decision
incentives and misleading presentation of the level of a bank's risk, are increasingly looking to internal economic
models for a better understanding of a bank's credit risk.
Sec. 1.1 The portfolio context of credit 7
Part I: Risk Measurement Framework
• The proliferation of credit enhancement mechanisms: thirdparty guarantees, posted
collateral, margin arrangements, and netting, makes it increasingly necessary to
assess credit risk at the portfolio level as well as at the individual exposure level.
• Improved liquidity in secondary cash markets and the emergence of credit deriva
tives have made possible more active management of credit risk based on rational
pricing. Proper due diligence standards require that institutions thoroughly review
existing risks before hedging or trading them.
• Innovative new credit instruments explicitly derive value from correlation estimates,
or credit events such as upgrades, downgrades or default. We can best understand
these in the context of a portfolio model that also explicitly accounts for credit qual
ity migrations.
Above, we discussed why a
portfolio
approach to credit risk is necessary. In the follow
ing section, we discuss why estimating portfolio credit risk is a much harder problem
than estimating portfolio market risk.
1.1.2 Challenges in estimating portfolio credit risk
Modeling portfolio risk in credit portfolios is neither analytically nor practically easy.
For instance, modern portfolio theory has taken enormous strides in its application to
equity price risks. However, fundamental differences between credit risks and equity
price risks make equity portfolio theory problematic when applied to credit portfolios.
There are two problems.
The ﬁrst problem is that equity returns are relatively symmetric and are well approxi
mated by normal or Gaussian distributions. Thus, the two statistical measures – mean
(average) and standard deviation of portfolio value – are sufﬁcient to help us understand
market risk and quantify percentile levels for equity portfolios. In contrast, credit returns
are highly skewed and fattailed (see
Chart 1.1).
Thus, we need more than just the mean
and standard deviation to fully understand a credit portfolio’s distribution.
Chart 1.1
Comparison of distribution of credit returns and market returns
This long downside tail of the distribution of credit returns is caused by defaults. Credit
returns are characterized by a fairly large likelihood of earning a (relatively) small proﬁt
through net interest earnings (NIE), coupled with a (relatively) small chance of losing a
0 Losses Gains
credit returns
market returns
Typical
Typical
8 Chapter 1. Introduction to CreditMetrics
CreditMetrics™—Technical Document
fairly large amount of investment. Across a large portfolio, there is likely to be a blend
of these two forces creating the smooth but skewed distribution shape above.
The second problem is the difﬁculty of modeling correlations. For equities, the correla
tions can be directly estimated by observing highfrequency liquid market prices. For
credit quality, the lack of data makes it difﬁcult to estimate any type of credit correlation
directly from history. Potential remedies include either: (i) assuming that credit correla
tions are uniform across the portfolio, or (ii) proposing a model to capture credit quality
correlations that has more readily estimated parameters.
In summary, measuring risk across a credit portfolio is as necessary as it is difﬁcult.
With the CreditMetrics methodology, we intend to address much of this difﬁculty.
1.2 Types of risks modeled
A distinction is often drawn between “market” and “credit” risk. But increasingly, the
distinction is not always clear (e.g., volatility of credit exposure due to FX moves). The
ﬁrst step, then, is to state exactly what risks we will be treating.
CreditMetrics estimates portfolio risk due to credit
events
. In other words, it measures
the uncertainty in the forward value of the portfolio at the risk horizon caused by the pos
sibility of obligor credit quality changes – both up(down)grades and default.
In addition, CreditMetrics allows us to capture certain market risk components in our
risk estimates. These include the marketdriven volatility of credit exposures like swaps,
forwards, and to a lesser extent, bonds. For these instruments, volatility of value due to
credit quality changes is increased by this further volatility of credit exposure.
2
Typi
cally, market volatilities are estimated over a daily or monthly risk horizon. However,
since credit is generally viewed over a larger horizon, marketdriven exposure estimates
should match the longer credit risk horizon.
1.3 Modeling the distribution of portfolio value
In this section, we begin to introduce some key modeling components: speciﬁcation of
which rating categories
3
to employ, probabilities of migrations between these categories,
revaluation upon an up(down)grade, and valuation in default.
For this section, we will be satisﬁed to obtain the distribution of outcomes; we will leave
until
Section 1.4
the calculation of standard deviations and percentile levels.
2
As a matter of implementation, the estimation of marketdriven exposure is performed in a J.P. Morgan software
product called FourFifteen™ which uses RiskMetrics data sets of market volatility and correlation to analyze mar
ket risk. However, the software implementation of CreditMetrics, CreditManager™, can accept marketdriven
exposures from any source.
3
By “rating categories”, we mean any grouping of ﬁrms of similar credit quality. This includes, but is in no way
limited to, the categories used by rating agencies. Groups of ﬁrms which KMV has assigned similar expected
default frequencies could just as easily be used as “rating categories.”
Sec. 1.3 Modeling the distribution of portfolio value 9
Part I: Risk Measurement Framework
1.3.1 Obtaining a distribution of values for a single bond
To begin, let us use S&P’s rating categories. Consider a single BBB rated bond which
matures in ﬁve years. For the purposes of this example, we make two choices. The ﬁrst
is to utilize the Standard & Poor’s rating categories and corresponding transition matri
ces.
4
The second is to compute risk over a one year horizon. Of course, other risk hori
zons may certainly be appropriate. Refer to
Section 2.5
for a discussion of how to
choose a risk horizon.
Our risk horizon is one year; therefore we are interested in characterizing the range of
values that the bond can take at the end of that period. Let us ﬁrst list all possible credit
outcomes that can occur at the end of the year due to credit events:
• the issuer stays at BBB at the end of the year;
• the issuer migrates up to AAA, AA, or A or down to BB, B, or CCC; or
• the issuer defaults.
Each outcome above has a different likelihood or probability of occurring. We derive
these from historical rating data, which we will discuss at the end of the chapter. For
now, we assume that the probabilities are known. That is, for a bond starting out as
BBB, we know precisely the probabilities that this bond will end up in one of the seven
rating categories (AAA through CCC) or defaults at the end of one year. These probabil
ities are shown in
Table 1.1
.
Table 1.1
Probability of credit rating
migrations in one year for a BBB
Note that there is a 86.93% likelihood that the bond stays at the original rating of BBB.
There is a smaller likelihood of a rating change (e.g., 5.95% for a rating change to
singleA), and a 0.18% likelihood of default.
So far we have speciﬁed: (i) each possible outcome for the bond’s yearend rating, and
(ii) the probabilities of each outcome. Now we must obtain the value of the bond under
4
Throughout Part I, we will consistently follow one set of credit quality migration likelihoods to aid clarity of expo
sition. This set of migration likelihoods happens to be taken from Standard & Poor’s. There are however a variety
of data providers and we in no way wish to give the impression that we endorse one over any other.
Yearend rating Probability (%)
AAA 0.02
AA 0.33
A 5.95
BBB 86.93
BB 5.30
B 1.17
CCC 0.12
Default 0.18
10 Chapter 1. Introduction to CreditMetrics
CreditMetrics™—Technical Document
each of the possible rating scenarios. What value will the bond have at yearend if it is
upgraded to singleA? If it is downgraded to BB?
To answer these questions, we must ﬁnd the new present value of the bond’s remaining
cash ﬂows at its new rating. The discount rate that enters this present value calculation is
read from the forward zero curve that extends from the end of the risk horizon to the
maturity of the bond. This zero curve is different for each forward rating category.
To illustrate, consider our ﬁveyear BBB bond. Say the face value of this bond is $100
and the coupon rate is 6%. We want to ﬁnd the value of the bond at yearend if it
upgrades to singleA. Assuming annual coupons for our example, at the end of one year
we receive a coupon payment of $6 from holding the bond. Four coupon payments ($6
each) remain, as well as the principal payment of $100 at maturity.
To obtain the value of the bond assuming an upgrade to singleA, we discount these ﬁve
cash ﬂows (four coupons and one principal) with interest rates derived from the forward
zero singleA curve. We leave aside the details of this calculation until the next chapter.
Here we just note that the calculations result in the following values at yearend across
all possible rating categories.
In
Table 1.2
, in the nondefault state, we show the coupon payment received, the forward
bond value, and the total value of the bond (sometimes termed the
dirty price
of the
bond). In the default state, the total value is due to a
recovery rate
(51.13% in this exam
ple), which we discuss in detail in the next chapter. Note that as expected, the value of
the bond increases if there is a rating upgrade. Conversely, the value decreases upon rat
ing downgrade or default. There is also a rise in value as the BBB remains BBB which is
commonly seen when the credit spread curve is upward sloping.
Table 1.2
Calculation of yearend values after credit rating migration from BBB ($)
Let us summarize what we have achieved so far. First, we have obtained the probabili
ties or likelihoods for the original BBB bond to be in any given rating category in one
year (
Table 1.1
). Further, we have also obtained the values of the bond in these rating
categories (
Table 1.2
). The information in
Tables
1.1
and
1.2
is now used to specify the
distribution of value of the bond in one year, as shown in
Table 1.3
.
Rating Coupon Forward Value Total Value
AAA 6.00 103.37 109.37
AA 6.00 103.10 109.19
A 6.00 102.66 108.66
BBB 6.00 101.55 107.55
BB 6.00 96.02 102.02
B 6.00 92.10 98.10
CCC 6.00 77.64 83.64
Default – 51.13 51.13
Sec. 1.3 Modeling the distribution of portfolio value 11
Part I: Risk Measurement Framework
Table 1.3
Distribution of value of a BBB par bond in one year
The value distribution is also shown graphically in
Chart 1.2
. Note that in the chart the
horizontal axis represents the value, and the vertical axis represents the probability. The
distribution of value tells us the possible values the bond can take at yearend, and the
probability or likelihood of achieving these numbers.
Chart 1.2
Distribution of value for a 5year BBB bond in one year
In the next section, we deﬁne the credit risk estimate for this value distribution. First,
however, we will discuss how we can generalize this probability distribution to a portfo
lio with more than just one instrument.
1.3.2 Obtaining a distribution of values for a portfolio of two bonds
So far we have illustrated the treatment of a standalone ﬁveyear BBB bond. Now we
will add a singleA three year bond to this portfolio. This bond pays annual coupons at
the rate of 5%. We want to obtain the distribution of values for this twobond portfolio
in one year. Just as in the onebond case, to characterize the distribution of values, we
need to specify the portfolio’s possible yearend values and the probability of achieving
these values. Now, we already know that the BBB bond can have one of the eight values
at yearend, as shown in
Table 1.2
. Similarly, we can calculate the corresponding year
Yearend rating Value ($) Probability (%)
AAA 109.37 0.02
AA 109.19 0.33
A 108.66 5.95
BBB 107.55 86.93
BB 102.02 5.30
B 98.10 1.17
CCC 83.64 0.12
Default 51.13 0.18
0.000
0.025
0.050
0.075
0.100
0.900
Default
CCC
B
BB
BBB
AAA
A
AA
Revaluation at risk horizon
Frequency
50 70 100 110 90 80 60
12 Chapter 1. Introduction to CreditMetrics
CreditMetrics™—Technical Document
end values for the singleA rated bond. Again, we leave the details of the calculation to
the next chapter, but simply state the results in
Table 1.4
.
Table 1.4
Yearend values after credit rating migration from singleA ($)
Next, we combine the possible values for the individual bonds (
Tables 1.2
and
1.4
) to
obtain the yearend values for the portfolio as a whole. Since either of the bonds can
have any of eight values in one year as a result of rating migration, the portfolio can take
on 64 (that is, 8 • 8) different values. We obtain the portfolio’s value at the risk horizon
in each of the 64 states by simply adding together the values for the individual bonds.
Thus, as an example, consider the top left cell in
Table 1.5
, which reads 215.96. This cell
corresponds to the outcome that both the BBB and singleA bonds upgrade to AAA at the
end of the year. From
Table 1.2
, the yearend value of the original BBB bond is $109.37
if it upgrades to AAA. Further, from
Table 1.4
, the yearend value of the original
singleA bond is $106.59 if it upgrades to AAA. Thus the portfolio as a whole has a
value of $215.96 (= $109.37 + $106.59) in the ﬁrst of 64 states. By similarly calculating
the values of the portfolio in the other states, we obtain the results shown in
Table 1.5
.
Table 1.5
All possible 64 yearend values for a twobond portfolio ($)
So
Table 1.5
shows the portfolio taking on 64 possible values at the end of a year
depending on the credit rating migration of the two bonds. These values range from
$102.26 (when both bonds default) to $215.96 (when both bonds are upgraded to AAA).
Yearend rating Coupon Forward Value Total Value
AAA 5.00 101.59 106.59
AA 5.00 101.49 106.49
A 5.00 101.30 106.30
BBB 5.00 100.64 105.64
BB 5.00 98.15 103.15
B 5.00 96.39 101.39
CCC 5.00 73.71 88.71
Default – 51.13 51.13
Obligor #1
(BBB)
Obligor #2 (
singleA
)
AAA AA A BBB BB B CCC Default
106.59 106.49 106.30 105.64 103.15 101.39 88.71 51.13
AAA 109.37 215.96 215.86 215.67 215.01 212.52 210.76 198.08 160.50
AA 109.19 215.78 215.68 215.49 214.83 212.34 210.58 197.90 160.32
A 108.66 215.25 215.15 214.96 214.30 211.81 210.05 197.37 159.79
BBB 107.55 214.14 214.04 213.85 213.19 210.70 208.94 196.26 158.68
BB 102.02 208.61 208.51 208.33 207.66 205.17 203.41 190.73 153.15
B 98.10 204.69 204.59 204.40 203.74 201.25 199.49 186.81 149.23
CCC 83.64 190.23 190.13 189.94 189.28 186.79 185.03 172.35 134.77
Default 51.13 157.72 157.62 157.43 156.77 154.28 152.52 139.84 102.26
Sec. 1.3 Modeling the distribution of portfolio value 13
Part I: Risk Measurement Framework
We have illustrated the different possible values for the portfolio at the end of the year.
To obtain the portfolio value distribution, the remaining piece in the puzzle is the likeli
hood or probability of observing these values. So we must estimate the likelihood of
observing each of the 64 states of
Table 1.5
in one year.
Those 64 “joint” likelihoods must reconsile with each set of eight likelihoods which we
have seen for the bonds on a standalone basis. In
Table 1.1
we showed the eight likeli
hoods for the BBB bond to be in each rating category in one year. Similarly, the corre
sponding likelihoods for the singleA rated bond are displayed in
Table 1.6
.
Table 1.6
Probability of credit rating migrations
in one year for a singleA
Again, we derive these likelihoods from historical rating data. We will brieﬂy touch on
the data issues at the end of the chapter, but leave aside the details for later (see
Chapter 6).
Here we just note the numbers as they are given to us.
We must now estimate the 64 joint likelihoods
5
so that we can calculate the volatility of
value in our twobond example. These joint likelihoods must satisfy the constraint of
summing to the standalone likelihoods in
Tables 1.1
and
1.6
. While doing this, we can
also specify that they reﬂect some desired correlation (i.e., a correlation equal to 0.0).
This is simple if the rating outcomes on the two bonds are independent of each other. In
this case the joint likelihood is simply a product of the individual likelihoods from
Tables 1.1
and
1.6
. Thus, for example, assuming independence, the likelihood that both
bonds will maintain their original rating at yearend is simply equal to the product of
86.93% (the probability of BBB bond staying BBB from
Table 1.1
) and 91.05% (the
probability of a singleA bond staying as singleA from
Table 1.6
) which is equal to
79.15%.
Unfortunately, this picture is simplistic. In reality, the rating outcomes on the two bonds
are not independent of each other, because they are affected at least in part by the same
macroeconomic factors. Thus, it is extremely important to account for correlations
between rating migrations in an estimation of the risk on a portfolio. We introduce our
model for correlations in
Chapter 3
and describe the model in detail in
Chapter 8.
5
By joint likelihoods, we mean the chance that the two obligors undergo a given pair of rating migrations, for
example, the ﬁrst obligor downgrades to BB while the second obligor remains at A.
Yearend rating Probability (%)
AAA 0.09
AA 2.27
A 91.05
BBB 5.52
BB 0.74
B 0.60
CCC 0.01
Default 0.06
14 Chapter 1. Introduction to CreditMetrics
CreditMetrics™—Technical Document
Here, we simply assume a correlation equal to 0.3 and take the resulting joint likelihoods
as given. For example, for the case mentioned above where the two bonds maintain their
original ratings at the end of the year, the actual joint likelihood value is 79.69%. For
other portfolio states, the joint likelihood values are as shown in
Table 1.7
below.
Table 1.7
Yearend joint likelihoods (probabilities) across 64 different states (%)
We now have all the data with which to specify the portfolio value distribution. Speciﬁ
cally, from
Table 1.5
we know all the different 64 values that the portfolio can have at the
end of a year. From
Table 1.7
we know the likelihoods of achieving each of these 64 val
ues. By plotting the likelihoods in
Table 1.7
and the values in
Table 1.5
on the same
graph, we obtain the portfolio value distribution shown in
Chart 1.3
.
Chart 1.3
Distribution of value for a portfolio of two bonds
1.3.3 Obtaining a distribution of values for a portfolio of more than two bonds
In our examples of one and two bond portfolios, we have been able to specify the entire
distribution of values for the portfolio. We remark that this becomes inconvenient, and
ﬁnally impossible, to do this in practice as the size of the portfolio grows. Noting that for
Obligor #1
(BBB)
Obligor #2 (singleA)
AAA AA A BBB BB B CCC Default
0.09 2.27 91.05 5.52 0.74 0.26 0.01 0.06
AAA 0.02 0.00 0.00 0.02 0.00 0.00 0.00 0.00 0.00
AA 0.33 0.00 0.04 0.29 0.00 0.00 0.00 0.00 0.00
A 5.95 0.02 0.39 5.44 0.08 0.01 0.00 0.00 0.00
BBB 86.93 0.07 1.81 79.69 4.55 0.57 0.19 0.01 0.04
BB 5.30 0.00 0.02 4.47 0.64 0.11 0.04 0.00 0.01
B 1.17 0.00 0.00 0.92 0.18 0.04 0.02 0.00 0.00
CCC 0.12 0.00 0.00 0.09 0.02 0.00 0.00 0.00 0.00
Default 0.18 0.00 0.00 0.13 0.04 0.01 0.00 0.00 0.00
102.3 172.4 203.4 211.8 215.7
0%
10%
70%
80%
Revaluation at risk horizon
Probability
Sec. 1.4 Different credit risk measures 15
Part I: Risk Measurement Framework
a three asset portfolio, there are 512 (that is, 8 times 8 times 8) possible joint rating
states. For a ﬁve asset portfolio, this number jumps to 32,768, and in general, for a port
folio with N assets, there are
8
N
possible joint rating states.
Because of this exponential growth in the complexity of the portfolio distribution, for
larger portfolios, we utilize a simulation approach. A simulation is very much like the
preceding example except that outcomes are sampled at random across all the possibile
joint rating states. The result of such an approach is an estimate of the portfolio distribu
tion, which for large portfolios looks more like a smooth curve and less like the collec
tions of a few discrete points in
Charts 1.2
and
1.3
.
We remark that it is always possible to compute some portfolio risk measures analyti
cally, regardless of the portfolio size, and discuss these in the following section.
1.4 Different credit risk measures
CreditMetrics can calculate two measures commonly used in risk literature to character
ize the credit risk inherent in a portfolio:
standard deviation
and
percentile level
. Both
measures reﬂect the portfolio value distribution and aid in quantifying credit risk. Nei
ther is “best.” They both contribute to our understanding of the risk.
We emphasize that the credit risk model underlying both of these risk measures is the
same. Therefore, the two risk measures reﬂect potential losses from the same portfolio
distribution. However, they are different measures of credit risk.
The credit risk in a portfolio arises because there is variability in the value of the portfo
lio due to credit quality changes. Therefore, we expect any credit risk measure to reﬂect
this variability. Loosely speaking, the greater the dispersion in the range of possible val
ues, the greater the absolute amount at credit risk. With this background, we next pro
vide two alternative measures of credit risk that we use in CreditMetrics.
1.4.1 Credit risk measure #1: standard deviation
The
standard deviation
is a symmetric measure of dispersion around the average portfo
lio value. The greater the dispersion around the average value, the larger the standard
deviation, and the greater the risk. If the portfolio values are expressed in dollars, this
standard deviation calculation also results in a dollar amount.
To illustrate the standard deviation calculation, we again refer to our twobond portfolio.
For this portfolio, the likelihoods of each state are shown in
Table 1.7
, and the values
corresponding to these states are displayed in
Table 1.5
. To calculate the standard devia
tion, we ﬁrst must obtain the
mean
value for the portfolio. This is obtained by multiply
ing the values with the corresponding probabilities and then adding the resulting values.
Mathematically, the average value can be written as:
[1.1]
where
p
1
refers to the probability or likelihood of being in State 1 at the end of the risk
horizon, and
V
1
refers to the value in State 1.
Mean p
1
V
1
p
2
V
2
⋅ … p
64
V
64
⋅ + + + ⋅ =
16 Chapter 1. Introduction to CreditMetrics
CreditMetrics™—Technical Document
Performing this simple calculation for our portfolio with the data from Table 1.5 and
Table 1.7, we ﬁnd that the average value for the portfolio is $213.63. Now the standard
deviation, which measures the dispersion between each potential migration value (V’s)
and this average value, is calculated as:
[1.2]
(Standard Deviation)
2
= p
1
· (V
1
–Mean)
2
+ p
2
· (V
2
–Mean)
2
+...+ p
64
· (V
64
–Mean)
2
Note that the above expression yields the squared standard deviation value, which is also
known as the “variance.” The squareroot of this value is the standard deviation. The
individual terms in the expression are of the form (V
i
–Average). This is consistent with
our earlier comment that the standard deviation measures the dispersion of the individual
values around the average value. Carrying out the above calculation for our example
portfolio, we ﬁnd that the portfolio standard deviation is $3.35.
The interpretation of standard deviation is difﬁcult here because credit risk is not nor
mally distributed. Thus, it is not possible to look up distribution probabilities in a normal
table. The distribution of credit value is likely to have a long tail on the “loss” side and
limited “gains” (see Chart 1.1). The length of this downside tail could be characterized
by its length in standard deviations. For instance, the 99% tail is 1.70 standard deviation
below the average (the 99.75% tail is 7.90 standard deviation below the average). By
comparison, these distances for a normal distribution are 2.33 and 2.81 standard devia
tions respectively.
Because the standard deviation statistic is a symmetric measure of dispersion, it does not
itself distinguish in our example between the gains side versus the losses side of the dis
tribution. It cannot, for instance, distinguish in our example that the maximum upside
value is only 0.70 standard deviations above the average while the maximum downside
value is 33.25 standard deviations below the average.
To calculate the standard deviation we do not have to specify the entire distribution of
portfolio values. Rather, we can operate pairwise across all pairs in the portfolio. We
discuss this pairwise calculation in the remainder of this section. For now, simply note
that since we do not have to rely on simulation to obtain the distribution of portfolio val
ues, the standard deviation calculation is computationally simple and efﬁcient.
1.4.2 Credit risk measure #2: percentile level
We deﬁne this second measure of risk as a speciﬁed percentile level of the portfolio
value distribution. The interpretation of the percentile level is much simpler than the
standard deviation: the lowest value that the portfolio will achieve 1% of the time is the
1
st
percentile.
Therefore, once we have calculated the 1
st
percentile level, the likelihood that the actual
portfolio value is less than this number is only 1%. Thus the 1
st
percentile level number
provides us with a probabilistic lower bound on the yearend portfolio value. Of course,
there is no particular percentile level that is “best” (5%, 1%, 0.5%, etc.). The particular
level used is the choice of the portfolio manager, and depends mostly on how the risk
measure will be applied.
Sec. 1.5 Exposure type differences 17
Part I: Risk Measurement Framework
For normal distributions (or any other known distribution which is completely character
ized by its mean and standard deviation), it is possible to calculate percentile levels from
knowledge of the standard deviation. Unfortunately, normal distributions are mostly a
characteristic of market risk.
6
In contrast, credit risk distributions are not typically sym
metrical or bellshaped. In particular, the distributions display a much fatter lower tail
than a standard bellshaped curve, as illustrated in Chart 1.1. Since we cannot assume
that credit portfolio distributions are normal, nor can we characterize them according to
any other standard distribution (such as the lognormal or Studentt), we must estimate
percentile levels via another approach.
To calculate a percentile level, we must ﬁrst specify the full distribution of portfolio val
ues. For portfolios consisting of more than two exposures, this requires a simulation
approach, which may be timeconsuming. Our approach will be to generate possible
portfolio scenarios at random according to a Monte Carlo framework. While the genera
tion of scenarios may be time consuming, once we obtain these scenarios, the calculation
of the 1
st
percentile level is simple. To do this, we ﬁrst sort the portfolio values in
ascending order. Given these sorted values, the 1
st
percentile level is the one below
which there are exactly 1% of the total values. So if the simulation generates 10,000
portfolio values, the 1
st
percentile level is the 100th largest among these.
Percentile levels may have more meaning for portfolios with many exposures, where the
portfolio can take on many possible values. We may still consider our example portfolio
with two bonds, however. For this portfolio, we estimate the 1
st
percentile to be
$204.40. Note that this amount is $9.23 (= $213.63 – $204.40) less than the mean port
folio value. Thus, using the 1
st
percentile, we estimate the amount at credit risk to be
$9.23, while using (one) standard deviation, we estimate this value at $3.35. Thus we
see that the two measures give different values and so must be interpreted differently.
These different computational requirements introduce a tradeoff between using the stan
dard deviation and using the percentile level. The percentile level is intuitively appeal
ing to use, because we know precisely what the likelihood is that the portfolio value will
fall below this number. On the other hand, it is often much faster to compute the stan
dard deviation. Users should evaluate this tradeoff carefully and use the risk measure
that best ﬁts their purpose. Further discussion of this issue is presented in Chapter 12.
1.5 Exposure type differences
Up to this point our examples have used bonds, but the concepts that we have described
in this chapter are equally applicable to other exposure types. The other exposure types
we consider are receivables, loans, commitments to lend, ﬁnancial letters of credit and
marketdriven instruments such as swaps and forwards.
Recall from Section 1.3 that we derive both of our credit risk measures from the portfolio
value distribution. Two components characterize this distribution. The ﬁrst is the likeli
hood of being in any possible portfolio state. The second is the value of the portfolio in
each of the possible states. Only the calculation of future values is different for different
instrument categories. The likelihoods of being in each credit quality state are the same
for all instrument categories since these are tagged to the obligor rather than to each of its
6
See, for example, RiskMetrics™—Technical Document, 4th Edition, 1996.
18 Chapter 1. Introduction to CreditMetrics
CreditMetrics™—Technical Document
obligations. In the remainder of this section, we brieﬂy discuss the different exposure
types in CreditMetrics. We provide a more detailed treatment in Chapter 4.
1.5.1 Receivables
Many commercial and industrial ﬁrms will have credit exposure to their customers
through receivables, or trade credit. We suggest that the risk in such exposures be
addressed within this same framework. It will commonly be the case that receivable will
have a “maturity” which is shorter than the risk horizon (e.g., one year or less). This
would simplify matters in that there would be no need to revalue the exposure upon
up(down)grade. But even if revaluing is necessary, the credit risk is – in concept – no
different than the risk in a comparable bond issued to the customer, and so it can be
revalued accordingly.
1.5.2 Bonds and loans
For bonds, as we discussed in Section 1.3, the value at the end of the risk horizon is the
present value of the remaining cash ﬂows. These cash ﬂows consist of the remaining
coupon payments and the principal payment at maturity. To discount the cash ﬂows, we
use the discount rates derived from the forward zero curve for each speciﬁc rating cate
gory. This forward curve is calculated as of the end of the risk horizon.
We treat loans in the same manner as bonds, revaluing in each future rating state by dis
counting future cash ﬂows. This revaluation accounts for the change in the value of a
loan which results from the likelihood changing that the loan will be repaid fully.
1.5.3 Commitments
A loan commitment is a facility which gives the obligor the option to borrow at his own
discretion. In practice, this essentially means both a loan (equal to the amount currently
drawn on the line) and an option to increase the amount of the loan up to the face amount
of the facility. The counterparty pays interest on the drawn amount, and a fee on the
undrawn amount in return for the option to draw down further. For these exposures three
factors inﬂuence the revaluation in future rating states:
• the amount currently drawn;
• expected changes in the amount drawn that are due to credit rating changes; and
• the spreads and fees needed to revalue both the drawn and undrawn portions.
All of these factors may be affected by covenants speciﬁc to a particular commitment.
The details of commitment revaluation and typical covenants are discussed in
Section 4.3.
Sec. 1.5 Exposure type differences 19
Part I: Risk Measurement Framework
1.5.4 Financial letters of credit
A ﬁnancial or standby letter of credit is treated as an off balance sheet item until it is
actually drawn. When it is drawn down its accounting treatment is just like a loan.
However, the obligor can draw down at his discretion and the lending institution typi
cally has no way to prevent a drawdown even during a period of obligor credit distress.
Thus, for risk assessment purposes, we argue that the full nominal amount should be con
sidered “exposed.” This means that we suggest treating a ﬁnancial letter of credit –
whether or not any portion is actually drawn – exactly as a loan.
Note that there are other types of letters of credit which may be either securitised by a
speciﬁc asset or project or triggered only by some infrequent event. The unique features
of these types of letters of credit are not currently addressable within the current speciﬁ
cation of CreditMetrics.
1.5.5 Marketdriven instruments
For instruments whose credit exposure depends on the moves of underlying market rates,
such as swaps and forwards, revaluation at future rating states is more difﬁcult. The
complexity for these instruments comes from the fact that if a swap, for example, is
marked to market and is currently outofthemoney to us, then a default by the counter
party does not inﬂuence the swap’s value, since we will still make the payments we owe
on the swap.
7
On the other hand, if the swap is inthemoney to us, then we expect pay
ments, and do not receive the full amount in the case of a counterparty default. So in
general, the credit exposure at any time to a marketdriven instrument is the maximum of
the transaction’s net present value or zero.
The methodology we propose for marketdriven instruments is applicable to single
instruments, such as swaps or forwards, or to groups of swaps, forwards, bonds, or other
instruments whose exposures can be netted. Thus, any set of cash ﬂows which are set
tled together (typically, these will all be exposures to the same counterparty) can be con
sidered as one marketdriven instrument.
In cases of default, we estimate the future value of marketdriven instruments using the
expected exposure of the instrument at the risk horizon. This expected exposure depends
both on the current market rates and their volatilities. In nondefault states, the revalua
tion consists of two parts: the present value of future cashﬂows, and the amount we
might lose if the counterparty defaults at some future time. The second part, the
expected loss, depends on the average marketdriven exposure over the remaining life of
the instrument (which is estimated in a similar fashion to the expected exposure men
tioned above), the probability that the counterparty will default over the same time
(which is determined by the credit rating at the risk horizon), and the recovery rate in
default.
Details of this methodology and a discussion of the various exposure calculations appear
in Section 4.5.
7
The exact settlement will depend on the covenants particular to the swap, but this is a reasonable assumption for
explanatory purposes.
20 Chapter 1. Introduction to CreditMetrics
CreditMetrics™—Technical Document
1.6 Data issues
Given a choice of which rating system (that is, what groupings of similar credits) will be
used, CreditMetrics requires three types of data:
• likelihoods of credit quality migration, including default likelihoods;
• likelihoods of joint credit quality migration; and
• valuation estimates (e.g. bonds revalued at forward spreads) at the risk horizon given
a credit quality migration.
Together, these data types result in the portfolio value distribution, which determines the
absolute amount at risk due to credit quality changes.
1.6.1 Data required for credit migration likelihoods
We showed these individual likelihoods for BBB and singleA rating separately in
Tables 1.1 and 1.6 respectively, but this information is more compactly represented in
matrix form as shown below in Table 1.8. We call this table a transition matrix. The rat
ings in the ﬁrst column are the starting or current ratings. The ratings in the ﬁrst row are
the ratings at the risk horizon. For example, the likelihoods in Table 1.8 corresponding
to an initial rating of BBB are represented by the BBB row in the matrix. Further, note
that each row of the matrix sums to 100%.
Table 1.8
Oneyear transition matrix (%)
Source: Standard & Poor’s CreditWeek (15 April 96)
Transition matrices can be calculated by observing the historical pattern of rating change
and default. They have been published by S&P and Moody’s rating agencies, and can be
computed based on KMV’s studies, but any provider’s matrix is welcome and usable
within CreditMetrics.
8
The transition matrix should, however, be estimated for the same
time interval as the risk horizon over which we are interested in estimating risks. For
instance, a semiannual risk horizon would use a semiannual rather than oneyear transi
tion matrix.
8
As we discuss later in Chapter 6, adjustments due to limited historical data may sometimes be desirable.
Initial
rating
Rating at yearend (%)
AAA AA A BBB BB B CCC Default
AAA 90.81 8.33 0.68 0.06 0.12 0 0 0
AA 0.70 90.65 7.79 0.64 0.06 0.14 0.02 0
A 0.09 2.27 91.05 5.52 0.74 0.26 0.01 0.06
BBB 0.02 0.33 5.95 86.93 5.30 1.17 0.12 0.18
BB 0.03 0.14 0.67 7.73 80.53 8.84 1.00 1.06
B 0 0.11 0.24 0.43 6.48 83.46 4.07 5.20
CCC 0.22 0 0.22 1.30 2.38 11.24 64.86 19.79
Sec. 1.7 Advanced modeling features 21
Part I: Risk Measurement Framework
1.6.2 Data required for joint likelihood calculations
Individual likelihoods are just one component of the portfolio joint likelihood. Earlier
we stated that the joint likelihood is not simply the product of the individual likelihoods.
This is because using the product as joint likelihood implicitly assumes that the pairwise
rating outcomes are independent of each other, which is generally not true.
Therefore, to have joint likelihoods, we need to do one of two things. First, we may his
torically tabulate joint credit rating moves just as we historically tabulated single credit
rating moves in the transition matrix. Second, we may propose a model for how the
credit ratings of multiple names evolve together, and estimate the requisite correlation
parameters for the model. We discuss several approaches to estimating joint likelihoods
in Chapter 8.
1.6.3 Data required for portfolio value calculation
Each instrument type requires sufﬁcient data to calculate the change in value for each
possible credit quality migration. These have been detailed in Section 1.5. In general,
there are three generic types:
1. Coupon rate and term of maturity are required for: receivables, loans, letters of
credit, and bonds in order to revalue them.
2. In addition to (1) above, we require the drawn and undrawn portions for a loan
commitment and the spread/fees for both portions.
3. Marketdriven instruments, including swaps, forwards, and to a lesser extent
bonds, require an examination of exposures which is detailed in Section 4.5.
It is interesting to note that an obligor’s exposures across instruments can be estimated
on a netted basis.
1.7 Advanced modeling features
CreditMetrics incorporates provisions to model additional parameters that make the
credit risk estimate more precise. One such provision is for cases such as swaps and for
wards, where the amount subject to credit risk is itself driven by market rates. Thus, not
only is it uncertain in these cases whether a counterparty will default or experience a
change in credit quality, but it is also uncertain what the loss will be in the event of a
default. Estimates of the exposures in these cases rely on market rates and volatilities.
Thus, as mentioned before, our software implementation of CreditMetrics, CreditMan
ager™, takes marketdriven exposures as an import from an external source, such as the
current version of J.P. Morgan’s FourFifteen™.
Separately, recoveries in the event of default are notoriously uncertain. Thus, we allow
for the treatment of recoveries as random quantities. We present mean and standard
deviation estimates for recoveries in Chapter 7. Standard deviation estimates of recov
ery value also are available from public research; these are provided in the CreditMetrics
data set.
22 Chapter 1. Introduction to CreditMetrics
CreditMetrics™—Technical Document
23
Chapter 2. Standalone risk calculation
This chapter illustrates the methodology used by CreditMetrics for calculating the credit
risk for a single or standalone exposure. In
Chapter 1
, we summarized this methodol
ogy with the help of a BBB rated bond. Here, we discuss in detail each of the steps out
lined in
Chapter 1
, using the same BBB example for illustration. Speciﬁcally:
• we describe an individual obligor and how his credit rating implies both a default
likelihood and the likelihoods for possible credit quality migrations;
• we describe a credit exposure and how its seniority standing implies a loss rate (that
is, loss in the event of default);
• we describe credit spreads over the default free yield and their implication for the
bond value upon up(down)grade in credit quality; and
• we assemble all of these pieces to estimate volatility of value due to credit quality
changes.
2.1 Overview: Risk for a standalone exposure
There are three steps to calculating the credit risk for a “portfolio” of one bond, as illus
trated in
Chart 2.1
below:
•
Step 1:
The senior unsecured credit rating of the bond’s issuer determines the
chance of the bond either defaulting or migrating to any possible credit quality state
at the risk horizon.
•
Step 2:
The seniority of the bond determines its recovery rate in the case of default.
The forward zero curve for each credit rating category determines the value of the
bond upon up(down)grade. Both of these aid revaluation of the bond.
•
Step 3:
The likelihoods from Step 1 and the values from Step 2 then combine in our
calculation of volatility of value due to credit quality changes.
Chart 2.1
Our ﬁrst “road map” of the analytics within CreditMetrics
Credit Rating Seniority Credit Spreads
Value at Risk due to Credit
bond revaluation
Present value
quality changes for a single exposure
Standard Deviation of value due to credit
Rating migration
likelihoods in default
Recovery rate
24 Chapter 2. Standalone risk calculation
CreditMetrics™—Technical Document
Readers who are familiar with RiskMetrics will see that the framework for credit risk
shown above is different from the market risk framework. This is because the quality
and availability of credit data are generally much different. Therefore, we
construct
what we cannot directly
observe
. In the process, we model the mechanisms of changes
in value rather than try to observe value changes.
In the following sections, we detail each step used in CreditMetrics to quantify the risk of
a standalone exposure. We illustrate these steps with our senior unsecured 5year BBB
rated bond. This bond pays an annual coupon at the rate of 6%. We also include this
example in the CHAP01.XLS Excel spreadsheet, which is available on our web site loca
tion (
http://www.jpmorgan.com
).
The calculations performed in this chapter assume a risk horizon of one year. This choice
is somewhat arbitrary. However, at the end of this chapter we discuss some of the issues
surrounding the choice of this risk horizon.
2.2 Step #1: Credit rating migration
In our model, risk comes not only from default but also from changes in value due to
up(down)grades. Thus, it is important for us to estimate not only the likelihood of
default but also the chance of
migrating
to any possible credit quality state at the risk
horizon. So we view default as just one of several “states of the world” that may exist for
this credit one period from now.
The likelihood of any credit rating migration in the coming period is conditioned on the
senior unsecured credit rating of the obligor.
1
Chart 2.2
shows the credit quality migra
tion likelihoods for obligors currently rated A, AAA, and BBB. For our BBB bond, the
rightmost diagram is applicable.
Chart 2.2
Examples of credit quality migrations (oneyear risk horizon)
1
There are some academic studies which condition the estimation of default likelihood upon not only the current
credit rating but also whether the speciﬁc debt issue is new: see for instance Altman [89]. While this is sound and
has its applications, we believe that many users will have dealings with established – not just new – obligors.
Chart 2.2
says, for example, that there is a 5.30% chance that a BBB rated credit will
downgrade to a BB rating within one year. There are several common patterns among
the three examples. Intuitively, we see that the most likely credit rating one year from
now is the current credit rating. The next most likely ratings are one letter grade above
or below. The only absolute rule about credit quality migrations is that the likelihoods
Credit
Rating
Rating
migration
Currently A rated. Currently AAA rated. Currently BBB rated.
0.09% AAA AAA 90.81% AAA 0.02% AAA
2.27% AA 8.33% AA 0.33% AA
A 91.05% A 0.68% A 5.95% A
5.52% BBB 0.06% BBB BBB 86.93% BBB
0.74% B B 0.12% B B 5.30% B B
0.26% B 0.00% B 1.17% B
0.01% CCC 0.00% CCC 0.12% CCC
0.06% D 0.00% D 0.18% D
100.00% 100.00% 100.00%
Sec. 2.2 Step #1: Credit rating migration 25
Part I:. Risk Measurement Framework
must sum to 100% since these are all the “states of the world” that are possible. Rather
than showing each rating’s credit quality migration likelihoods separately, it is often con
venient to think of them in a square table, or transition matrix, as shown below in
Table 2.1
.
Table 2.1
Oneyear transition matrix (%)
Source: Standard & Poor’s CreditWeek (15 April 96)
To read this table, ﬁnd today’s credit rating on the left and follow along that row to the
column which represents the rating at the risk horizon. For instance, the leftmost bottom
ﬁgure of 0.22% says that there is a 0.22% likelihood that a CCC rated credit will migrate
to AAA at the end of one year.
We derived the transition matrix in
Table 2.1
from rating migration data published by
S&P. Thus the leftmost bottom ﬁgure of 0.22% means that 0.22% of the time (over the
15year history from which this data was tabulated) a CCCrated credit today migrated to
AAA in one year. Of course, migrating from CCC to AAA within one year is highly
unusual and likely represents only one instance in the historical data.
2
This presents a
practical problem: results based on limited data are subject to estimation errors. Later, in
Chapter 6
, we discuss the anticipated longterm behavior of credit migrations, which
would tend to mitigate this estimation noise.
As we have mentioned, it is possible to create transition matrices for any system of
grouping similar credits. Again, we refer to these groupings loosely as rating categories.
Regardless of how the rating categories are constructed and of how many categories
there are, it is necessary to specify the default likelihood for each category, and the like
lihoods that ﬁrms in one category migrate to any other. In addition, as we will see in the
following section, for the purposes of revaluation, it is necessary to provide a credit
spread to correspond to each category as well.
2
The only adjustment we made to S&P’s data was for the “nolongerrated” migrations. The CCC row in this tran
sition matrix, sourced from S&P, is based upon 561 ﬁrm/years worth of observation with 79 occurrences of a tran
sition to “no longer rated.” Across all rows in this transition matrix, there are more than 25,000 ﬁrm/years worth
of observation, with most being in the BBBtoAA rows.
Initial
Rating
Rating at yearend (%)
AAA AA A BBB BB B CCC
Default
AAA 90.81 8.33 0.68 0.06 0.12 0 0 0
AA 0.70 90.65 7.79 0.64 0.06 0.14 0.02 0
A 0.09 2.27 91.05 5.52 0.74 0.26 0.01 0.06
BBB 0.02 0.33 5.95 86.93 5.30 1.17 0.12 0.18
BB 0.03 0.14 0.67 7.73 80.53 8.84 1.00 1.06
B 0 0.11 0.24 0.43 6.48 83.46 4.07 5.20
CCC 0.22 0 0.22 1.30 2.38 11.24 64.86 19.79
26 Chapter 2. Standalone risk calculation
CreditMetrics™—Technical Document
2.3 Step #2: Valuation
In Step 1, we determined the likelihoods of migration to any possible credit quality states
at the risk horizon. In Step 2, we determine the values at the risk horizon for these credit
quality states. Value is calculated once for each migration state; thus there are (in this
example) eight revaluations in our simple onebond example.
These eight valuations fall into two categories. First, in the event of a default, we esti
mate the recovery rate based on the seniority classiﬁcation of the bond. Second, in the
event of up(down)grades, we estimate the change in credit spread that results from the
rating migration. We then perform a present value calculation of the bond’s remaining
cash ﬂows at the new yield to estimate its new value.
2.3.1 Valuation in the state of default
If the credit quality migration is into default, the likely residual value net of recoveries
will depend on the seniority class of the debt. In CreditMetrics
,
we offer several histori
cal studies of this dependence.
3
Table 2.2
below summarizes the recovery rates in the
state of default as reported by one of the available studies.
Table 2.2
Recovery rates by seniority class (% of face value, i.e., “par”)
Source: Carty & Lieberman [96a] —Moody’s Investors Service
In this table, we show the mean recovery rate (middle column) as well as the standard
deviation of the recovery rate (last column). Our example BBB bond is senior unse
cured. Therefore, we estimate its mean value in default to be 51.13% of its face value –
which in this case we have assumed to be $100. Also from
Table 2.2
, the standard devia
tion of the recovery rate is 25.45%.
2.3.2 Valuation in the states of up(down)grade
If the credit quality migration is to another letter rating rather than to default, then we
must revalue the exposure by other means.
To obtain the values at the risk horizon corresponding to rating up(down)grades, we per
form a straightforward present value bond revaluation. This involves the following
steps:
3
There is also a recent study (see Altman & Kishore [96]) which conditions recovery rates on industry participa
tions of the obligor in addition to seniority class.
Seniority Class Mean (%) Standard Deviation (%)
Senior Secured 53.80 26.86
Senior Unsecured 51.13 25.45
Senior Subordinated 38.52 23.81
Subordinated 32.74 20.18
Junior Subordinated 17.09 10.90
Seniority
Recovery
rate in
default
PV bond
revaluation
Credit
Spreads
Sec. 2.3 Step #2: Valuation 27
Part I:. Risk Measurement Framework
1. Obtain the forward zero curves for each rating category. These forward curves are
stated as of the risk horizon and go to the maturity of the bond.
2. Using these zero curves, revalue the bond's remaining cash ﬂows at the risk hori
zon for each rating category.
Let us illustrate the above steps with the help of our BBB bond example. Recall that this
bond has a ﬁveyear maturity, and pays annual coupons at the rate of 6%. Assume that
the forward zero curves for each rating category has been given to us. We show an
example in
Table 2.3
below.
Table 2.3
Example oneyear forward zero curves by credit rating category (%)
First, let us determine the cash ﬂows which result from holding the bond position.
Recall that our example bond pays an annual coupon at the rate of 6%. Therefore,
assuming a face value of $100, the bond pays $6 each at the end of the next four years.
At the end of the ﬁfth year, the bond pays a cash ﬂow of face value plus coupon, which
equals $106 in this case.
Now, let us calculate the value
V
of the bond at the end of one year assuming that the
bond upgrades to singleA. This calculation is described by the formula below:
[2.1]
In the above formula, we use the forward zero rates for the singleA rating category from
Table 2.3
. To calculate the value of the bond in a rating category other than singleA, we
would substitute the appropriate zero rates from the table. After completing these calcu
lations for different rating categories, we obtain the values in
Table 2.4.
Category Year 1 Year 2 Year 3 Year 4
AAA 3.60 4.17 4.73 5.12
AA 3.65 4.22 4.78 5.17
A 3.72 4.32 4.93 5.32
BBB 4.10 4.67 5.25 5.63
BB 5.55 6.02 6.78 7.27
B 6.05 7.02 8.03 8.52
CCC 15.05 15.02 14.03 13.52
V 6
6
1 3.72% + ( )

6
1 4.32% + ( )
2

6
1 4.93% + ( )
3

6
1 5.32% + ( )
4
 108.66 = + + + + =
28 Chapter 2. Standalone risk calculation
CreditMetrics™—Technical Document
Table 2.4
Possible oneyear forward values for a BBB bond plus coupon
2.4 Step #3: Credit risk estimation
We now have all the information that we need to estimate the volatility of value due to
credit quality changes for this one exposure on a standalone basis. That is, we know the
likelihood of all possible outcomes – all up(down)grades plus default – and the distribu
tion of value within each outcome. These likelihoods and values, which we obtain from
Steps 1 and 2 respectively, are shown in
Table 2.5
below.
Table 2.5
Calculating volatility in value due to credit quality changes
The ﬁgures in the ﬁrst two columns – likelihoods of migration and value in each state –
have been discussed in
Sections
2.2,
and
2.3,
respectively. Here we use these two col
umns to calculate the risk estimate.
2.4.1 Calculation of standard deviation as a measure of credit risk
Recall from Chapter 1 that there are two useful measures of credit risk that one can use:
standard deviation and percentile level. First we consider the calculation of the standard
deviation. For this, we have to ﬁrst obtain the average value (the mean).
Yearend rating Value ($)
AAA 109.37
AA 109.19
A 108.66
BBB 107.55
BB 102.02
B 98.10
CCC 83.64
Default 51.13
Yearend
rating
Probability
of state (%)
New bond
value plus
coupon ($)
Probability
weighted
value ($)
Difference of
value from
mean ($)
Probability
weighted difference
squared
AAA 0.02 109.37 0.02 2.28 0.0010
AA 0.33 109.19 0.36 2.10 0.0146
A 5.95 108.66 6.47 1.57 0.1474
BBB 86.93 107.55 93.49 0.46 0.1853
BB 5.30 102.02 5.41 (5.06) 1.3592
B 1.17 98.10 1.15 (8.99) 0.9446
CCC 0.12 83.64 1.10 (23.45) 0.6598
Default 0.18 51.13 0.09 (55.96) 5.6358
Mean = $107.09 Variance = 8.9477
Standard deviation = $2.99
Standard deviation of value due to credit
changes for a single exposure
Sec. 2.4 Step #3: Credit risk estimation 29
Part I:. Risk Measurement Framework
Note that the mean is just the probabilityweighted average of the values across all rating
categories including default. As shown in
Table 2.5
, this is estimated at $107.09, which
includes the $6.00 coupon in all nondefault states. The standard deviation then mea
sures the dispersion between the individual values and this mean. After completing the
calculations, we observe that the standard deviation of value changes due to credit is
$2.99. This (or some scale of this) is one measure of the absolute amount that is at credit
risk.
In the above calculation of standard deviation, we used a recovery value of $51.13 for the
case of default. (This is the expected recovery rate for a senior unsecured bond from
Table 2.2
.) While discussing the results presented in this table, we pointed out that there
is an uncertainty or standard deviation associated with this recovery rate. This uncer
tainty adds to the overall credit risk of holding the bond position. Before we describe
how to account for this recovery rate uncertainty, let us present the standard deviation
calculation in a manner that will help us later to incorporate recovery rate uncertainty.
Let
p
i
be the probability of being in any given state and
µ
i
be the value within each state
(the ﬁrst and second columns of
Table 2.5
respectively). Given this, we calculate the
mean
µ
and the standard deviation
σ
using the formulae below:
[2.2]
The above formula is overly simple in that it allows the bond to only take on a mean
value within each state. In general, the bond can take on a distribution of values within
each state. In particular, there is well documented uncertainty surrounding the recovery
rate in default. We incorporate this added uncertainty as follows:
µ
Total
p
i
µ
i
i 1 =
s
∑
=
0.02% 109.37 + ⋅
0.33% 109.19 + ⋅
5.95% 108.66 + ⋅
86.93% 107.55 + ⋅
5.30% 102.02 + ⋅
1.17% 98.10 + ⋅
0.12% 83.64 + ⋅
0.18% 51.13 ⋅ ¸ ,
¸ _
=
107.09 =
σ
Total
p
i
µ
i
2
µ
Total
2
–
i 1 =
s
∑
=
0.02% 109.37
2
+ ⋅
0.33% 109.19
2
+ ⋅
5.95% 108.66
2
+ ⋅
86.93% 107.55
2
+ ⋅
5.30% 102.02
2
+ ⋅
1.17% 98.10
2
+ ⋅
0.12% 83.64
2
+ ⋅
0.18% 51.13
2
⋅ ¸ ,
¸ _
107.09
2
– =
2.99 =
30 Chapter 2. Standalone risk calculation
CreditMetrics™—Technical Document
[2.3]
Note that the expected value or mean calculation remains the same as before. The only
difference is in the standard deviation calculation, where we add a component
σ
i
repre
senting the uncertainty in recovery value in the defaulted state
i
= 8.
For a derivation of this formula for the standard deviation, refer to
Appendix D
. This
inclusion of the uncertainty in recovery rates increases the standard deviation from $2.99
to $3.18 (a 6.32% increase).
Finally, note the zero values in the standard deviation formula. These zeros represent the
uncertainty of value in the up(down)grade states. Just as there is an uncertainty in value
in the default state, we expect an uncertainty in value in the other rating up(down)grade
states. This would be caused by the uncertainty of credit spreads within each credit rat
ing category. For now, we have set this credit spread uncertainty to zero since it is
unclear what portion of it is systematic versus diversiﬁable. If we ever have sufﬁcient
data to resolve this issue, we hope to allow credit spread volatility in future versions of
CreditMetrics.
2.4.2 Calculation of percentile level as a measure of credit risk
Standard deviation is just one of two useful credit risk measures. The other risk measure
is the percentile level.
Say we are interested in determining the 1
st
percentile level for our bond. This is the
level below which our portfolio value will fall with probability 1%. Again, 1% is not the
only percentile level we advocate for the reader. There are good reasons why different
users should use different percentile levels. For the sake of illustration, however, we
concentrate on the calculation of the 1
st
percentile level.
As we have mentioned before, percentile levels are more meaningful statistics for large
portfolios, where the portfolio can take on many different portfolio values. It is also the
µ
Total
p
i
µ
i
i 1 =
s
∑
=
0.02% 109.37 + ⋅
0.33% 109.19 + ⋅
5.95% 108.66 + ⋅
86.93% 107.55 + ⋅
5.30% 102.02 + ⋅
1.17% 98.10 + ⋅
0.12% 83.64 + ⋅
0.18% 51.13 ⋅ ¸ ,
¸ _
=
107.09 =
σ
Total
p
i
µ
i
2
σ
i
2
+ ( ) µ –
Total
2
i 1 =
s
∑
=
0.02% 109.37
2
0
2
+ ( ) + ⋅
0.33% 109.19
2
0
2
+ ( ) + ⋅
5.95% 108.66
2
0
2
+ ( ) + ⋅
86.93% 107.55
2
0
2
+ ( ) + ⋅
5.30% 102.02
2
0
2
+ ( ) + ⋅
1.17% 98.10
2
0
2
+ ( ) + ⋅
0.12% 83.64
2
0
2
+ ( ) + ⋅
0.18% 51.13
2
25.45
2
+ ( ) ⋅ ¸ ,
¸ _
107.09
2
– =
3.18 =
Sec. 2.5 Choosing a time horizon 31
Part I:. Risk Measurement Framework
case that for these large portfolios (in fact, for any portfolio with much more than two
assets), it is necessary to perform simulations to compute percentile levels. Nonetheless,
in order to provide an example, we compute percentile levels for our single bond.
Table 2.5 displays the likelihood that our bond will be in any given credit rating at the
risk horizon and the value at each credit rating. We start from the bottom of the table, the
state of default, and move upwards towards the AAA rating state. We keep a running
total of the likelihoods as we move up. The value at which this running total ﬁrst
becomes equal to or greater than 1% is the 1
st
percentile level.
Let us go through the procedure: The likelihood of being in the defaulted state is 0.18%.
This is less than 1%, so we move up to the CCC state. The combined likelihood of being
in default or CCC state is 0.30% (sum of 0.18% and 0.12%). This is also less than 1%;
so we move up again, this time to B rating state. The combined likelihood of being in
default, CCC, or B is now equal to 2.17% (sum of 0.30% and 1.17%). This now exceeds
1%. We therefore stop here and read off the corresponding value from the B row. This
value, which is equal to $98.10, is the 1
st
percentile level value. This is $8.99 below the
mean value.
So far we have used an arbitrary risk horizon of one year. Below, we discuss issues sur
rounding the choice of risk horizon.
2.5 Choosing a time horizon
Much of the academic and credit agency data is stated on an annual basis. This is a con
vention rather than a requirement. It is important to note, that there is nothing about the
CreditMetrics methodology that requires a oneyear horizon. Indeed, it is difﬁcult to
support the argument than any one particular risk horizon is best. Illiquidity, credit rela
tionships, and common lack of credit hedging instruments can all lead to prolonged risk
mitigating actions.
The choice of risk horizon raises two practical questions:
• Should a practitioner use only one risk horizon or many?
• Is there any ﬁrm basis for saying that any one particular horizon is best?
2.5.1 Should there be one horizon or many?
The choice of time horizon for risk measurement and risk management is not clear
because there is no explicit theory to guide us. However, the one thing that is clear is
that comparisons between alternatives must be made at the same risk horizon.
Many different security types bear credit risk. One of the common arguments in favor of
multiple credit risk horizons is that they allow us to calculate risk at horizons tailored to
each credit security type. For instance, it may be that interest rate swaps are more liquid
than loans. The managers for each securitytype (e.g., loans versus swaps) may wish to
see their security type calculated at their own risk horizon. However, the risk estimates
for these different subportfolios cannot be aggregated if there is a mismatch in time hori
zons.
32 Chapter 2. Standalone risk calculation
CreditMetrics™—Technical Document
2.5.2 Which horizon might be “best”?
Almost any risk measurement system is better at stating relative risk than it is at stating
absolute risk. Since relative risk measurements will likely drive decisions, the choice of
risk horizon is not likely to make an appreciable difference. The key element to any risk
information system is the resulting riskmitigating actions; any given risk horizon is
likely to lead to the same qualitative decisions.
Although these actions may differ among institutions, the risk horizon is not likely to be
signiﬁcantly less than a quarter for a bank with loans, commitments, ﬁnancial letters of
credit, etc. On the other side, the natural turnover due to the ongoing maturity and rein
vestment of positions provides appreciable room for riskmitigating action even for
highly illiquid instruments. Thus, using as a convention a one year risk horizon – not
unlike the convention of annualized interest rates – is common.
Even if riskmitigating actions are performed daily, recalculating risk at a longer horizon
can still provide guidance to changes in relative risk. An analogy is driving a car. A
car’s instrument panel serves perfectly well when reporting speed at kilometers per hour
even though driving decisions are made far more often, perhaps every second and every
meter. So too, risk stated over the coming year can guide riskmitigating actions.
2.5.3 Computing credit risk on different horizons
Two CreditMetrics modeling parameters must change to address different risk horizons:
• the credit instrument revaluation formulas change to perform the revaluation compu
tation for the alternate time horizon; and
• the likelihoods of credit quality migration, as shown in the transition matrix, must be
restated to the new risk horizon.
One way of doing the latter is simply to multiply the shorthorizon transition matrices to
obtain the transition matrix for a longer horizon. (For example, a twoyear transition
matrix could be obtained by multiplying the oneyear transition matrix with itself.)
Unfortunately, however, this methodology ignores the issue of autocorrelation in the
credit quality changes over multiple time horizons. A nonzero autocorrelation would
indicate that successive credit quality moves are not statistically independent between
adjoining periods.
This issue of autocorrelation surfaces for market risk calculations also. For instance,
some markets tend to exhibit mean reversion (that is, a tendency for prices to return to
some longterm stable level), autocorrelation prevents us from translating daily volatili
ties to monthly or yearly volatilities in a simple way.
Regrettably, the issue of time period interdependencies can also arise for credit quality
migrations. For instance, Altman & Kao [92b] ﬁnd that there is positive autocorrelation
in S&P downgrades, so a downgrade implies a higher likelihood of a downgrade in the
following period. We conﬁrm this, looking at the S&P rating data, and also ﬁnd that an
upgrade tends to lead to a “quiet” period. Note, however, that this ﬁnding applies in par
ticular to the S&P rating system, and other credit assessment approaches are not neces
Sec. 2.5 Choosing a time horizon 33
Part I:. Risk Measurement Framework
sarily subject to this problem. We discuss these and other issues surrounding transition
matrices in Chapter 6.
In this chapter we have discussed the essence of the CreditMetrics methodology. The
next two chapters extend our framework across a portfolio of exposures and across dif
ferent exposure types beyond a simple bond.
34 Chapter 2. Standalone risk calculation
CreditMetrics™—Technical Document
35
Chapter 3. Portfolio risk calculation
In
Chapter 2
, we explained the methodology used by CreditMetrics to obtain the credit
risk for a standalone exposure. Here, we extend our methodology to a “portfolio” of
two exposures. The chapter is organized as follows:
• we elaborate on the joint likelihoods in the credit quality comovements;
• we extend our credit risk calculation for standalone exposure (discussed in
Chapter 2
) to the multiple exposure case; and
• we discuss the calculation of marginal risk estimation, which identiﬁes overconcen
trations within a portfolio and thus suggests potential riskmitigating actions.
For clarity, we discuss the required steps to calculate credit risk across a portfolio with
an example portfolio consisting of the following two speciﬁc bonds:
•
Bond #1:
BBB rated, senior unsecured, 6% annual coupon, ﬁveyear maturity
•
Bond #2:
A rated, senior unsecured, 5% annual coupon, threeyear maturity
This example portfolio is the same as the one that we considered in
Chapter 1
when we
were highlighting the steps in the calculation of portfolio credit risk. Here, we discuss
the same steps as in
Chapter 1,
but in greater detail. Also, we point out that Bond #1 is
the one for which we estimated the credit risk on a standalone basis in
Chapter 2
.
We now update the “road map” for CreditMetrics in
Chart 3.1
to show the additional
work needed to address a portfolio. The reader can compare this chart with the prior
chapter’s corresponding
Chart 2.1
for a standalone exposure. Note that there is one sig
niﬁcant addition. We must now estimate the contribution to risk brought by the effects
of nonzero credit quality
correlations
. Thus, we must estimate
joint likelihoods
in the
credit quality
comovements
.
Chart 3.1
Our second “road map” of the analytics within CreditMetric
s
Understanding joint likelihoods will allow us to properly account for the portfolio diver
siﬁcation effects. Correlation will, for example, determine how often losses occur in
multiple exposures at the same time. Our volatility of value – our risk – will be lower if
Credit Rating Seniority Credit Spreads
Value at Risk due to Credit
bond revaluation
Present value
quality changes for a single exposure
Standard Deviation of value due to credit
Ratings series,
Equities series
Correlations
Models(e.g.,
correlations)
Rating migration
likelihoods
Joint credit
rating changes
in default
Recovery rate
36 Chapter 3. Portfolio risk calculation
CreditMetrics™—Technical Document
the correlation between credit events is lower. However, we do not elaborate here on the
connection between correlation and joint likelihoods. Rather, we assume that the joint
likelihoods are given to us. Later, in
Chapter 8
, we will discuss several different meth
ods for determining joint likelihoods of credit quality migrations.
3.1 Joint probabilities
We have seen that, with the major ratings from AAA to CCC, there are eight possible
outcomes for an obligor’s credit quality in one year. Now we are interested in two obli
gors considered together. For us to estimate this joint risk, we need to consider all possi
ble combinations of states between the two obligors. There are eight times eight or
sixtyfour possible states to which the two credits might migrate at the risk horizon.
The simplest way of obtaining the joint likelihoods is to just assume that these are the
product of the individual likelihoods. Thus, as shown in
Table 3.1
below, the joint likeli
hood that the two obligors maintain their initial ratings is equal to 79.15%. This is the
product of 86.93% (the likelihood that the BBB rated bond remains a BBB) and 91.05%
(the likelihood that the singleA rated bond remains a singleA).
By repeating this type of calculation for all the 64 states we then ﬁll the joint likelihood
table shown below. However, calculation will be true only for the simplest case where
the two obligors’ credit rating changes are statistically independent.
Table 3.1
Joint migration probabilities with zero correlation (%)
Assuming a zero correlation like this is too simplistic. It is unrealistic since these credit
movements are affected in part by the same macro economic variables. In order to cap
ture this effect, we will introduce in
Chapter 8
a model which links ﬁrm asset value to
ﬁrm credit rating. We touch brieﬂy on this model here.
In
Chart 3.2
we illustrate a framework for thinking about default as a function of the
underlying (and volatile) value of the ﬁrm. This framework was ﬁrst proposed by Robert
Merton (see Merton [74]), and is often referred to as the
option theoretic
valuation of
79.15% 86.93% 91.05% ⋅ =
¹ ¹ ' ¹ ¹ ¹ ¹ ' ¹ ¹ ¹ ¹ ' ¹ ¹
Chance both
retain current rating
Chance a BBB
remains at BBB
Chance an A
remains at A
Obligor #1
(BBB)
Obligor #2 (singleA)
AAA AA A BBB BB B CCC Default
0.09 2.27 91.05 5.52 0.74 0.26 0.01 0.06
AAA 0.02 0.00 0.00 0.02 0.00 0.00 0.00 0.00 0.00
AA 0.33 0.00 0.01 0.30 0.02 0.00 0.00 0.00 0.00
A 5.95 0.01 0.14 5.42 0.33 0.04 0.02 0.00 0.00
BBB 86.93 0.08 1.98 79.15 4.80 0.64 0.23 0.01 0.05
BB 5.30 0.00 0.12 4.83 0.29 0.04 0.01 0.00 0.00
B 1.17 0.00 0.03 1.06 0.06 0.01 0.00 0.00 0.00
CCC 0.12 0.00 0.00 0.11 0.01 0.00 0.00 0.00 0.00
Default 0.18 0.00 0.00 0.16 0.01 0.00 0.00 0.00 0.00
Sec. 3.1 Joint probabilities 37
Part I: Risk Measurement Framework
debt. It builds upon Black and Scholes option pricing model by stating that the credit
risk component of a ﬁrm’s debt can be valued like a put option on the value of the under
lying assets of the ﬁrm. Under the Merton model, underlying ﬁrm value is random with
some distribution. If the value of assets should happen to decline so much that the value
is less than amount of liabilities outstanding, which we refer to as the default threshold,
then it will be impossible for the ﬁrm to satisfy its obligations and it will thus default.
Chart 3.2
Model of ﬁrm value and its default threshold
We do not suggest here that default likelihoods must be estimated based on the volatility
of underlying ﬁrm value. CreditMetrics assumes that each obligor will be labeled with a
credit rating, which in turn will be associated with a default likelihood. It is unimportant
to CreditMetrics
how
default likelihoods are estimated. We treat them as input
parameters.
The Merton model can be easily extended to include rating changes. The generalization
involves stating that in addition to the default threshold, there are credit rating
up(down)grade thresholds as well. The ﬁrm’s asset value relative to these thresholds
determines its future rating, as illustrated in
Chart 3.3
.
Chart 3.3
Model of ﬁrm value and generalized credit quality thresholds
Value of the firm
Higher Lower
Default
Scenarios
Default
threshold
Default
CCC
B
BB
Firm remains
A
AA
AAA
Value of BBB firm at horizon date
BBB
Higher Lower
BBB
38 Chapter 3. Portfolio risk calculation
CreditMetrics™—Technical Document
In the end, we have a link between the underlying ﬁrm value and the ﬁrm’s credit rating,
and can build the joint probabilities for two obligors from both this and a knowledge of
the correlation between the two obligors’ ﬁrm values. Again, this approach is developed
in detail in
Section 8.4
. In
Table 3.2
, we present joint likelihoods which result from an
application of this model.
Table 3.2
Joint migration probabilities with 0.30 asset correlation (%)
There are at least four interesting features in the joint likelihood table above:
1. The probabilities across the table necessarily sum to 100%.
2. The most likely outcome is that both obligors simply remain at their current credit
ratings. In fact, the likelihoods of joint migration become rapidly smaller as the
migration distance grows.
3. The effect of correlation is generally to increase the joint probabilities along the
diagonal drawn through their current joint standing (in this case, through BBBA).
4. The sum of each column or each row must equal the chance of migration for that
obligor standing alone. For instance, the sum of the last row must be 0.18%,
which is the default likelihood for Obligor #1 (BBB) in isolation.
With this discussion of the joint likelihood, we turn our attention next to the credit risk
calculation for our example twobond portfolio. Speciﬁcally, we show how we extend
our credit risk calculation from the standalone exposure case to the multiple exposure
portfolio case.
3.2 Portfolio credit risk
As mentioned in
Chapter 1
, to calculate the volatility of value due to credit quality
changes, we need two types of information for each of the 64 joint states between two
obligors: joint likelihoods and revaluation estimates. In the previous section, we covered
the joint likelihoods across comovements in credit quality. Here, we discuss the revalu
ation of the two exposures – given the credit quality migration – for each of the 64 states.
We will see that these data are then combined for two obligors in a fashion very similar
to what we have already shown for one obligor.
Obligor #1
(BBB)
Obligor #2 (singleA)
AAA AA A BBB BB B CCC Default
0.09 2.27 91.05 5.52 0.74 0.26 0.01 0.06
AAA 0.02 0.00 0.00 0.02 0.00 0.00 0.00 0.00 0.00
AA 0.33 0.00 0.04 0.29 0.00 0.00 0.00 0.00 0.00
A 5.95 0.02 0.39 5.44 0.08 0.01 0.00 0.00 0.00
BBB 86.93 0.07 1.81 79.69 4.55 0.57 0.19 0.01 0.04
BB 5.30 0.00 0.02 4.47 0.64 0.11 0.04 0.00 0.01
B 1.17 0.00 0.00 0.92 0.18 0.04 0.02 0.00 0.00
CCC 0.12 0.00 0.00 0.09 0.02 0.00 0.00 0.00 0.00
Default 0.18 0.00 0.00 0.13 0.04 0.01 0.00 0.00 0.00
Sec. 3.2 Portfolio credit risk 39
Part I: Risk Measurement Framework
We now must determine the 64 possible values of the portfolio at the risk horizon. This
is easy since the value that the portfolio takes on in each pairwise credit rating class is
simply the sum of the individual values. The portfolio values in each possible joint rat
ing state are given in
Table 1.5
.
It is again noteworthy that the greatest potential for changes in value are on the down
side. Indeed, the value is relatively ﬂat across most of
Table 1.5
. It is only when either
(or both) obligors suffer a downturn that the change in value becomes great. This same
data was illustrated in
Chart 1.3
showing the frequency distribution of values.
We next focus on the calculation of the two risk measures for the portfolio, namely the
standard deviation and the percentile level. As far as the portfolio standard deviation is
concerned, we use the same formula in the standalone exposure case of
Chapter 2
. The
only difference is that now we have 64 possible states rather than just eight in the stand
alone case. We illustrate this standard deviation calculation for the twobond portfolio
below:
[3.1]
The probabilities and the values that enter the standard deviation calculation are read off
Tables 3.2
and
1.5
respectively. Also, for simplicity, the above calculation ignores the
additional contribution to the portfolio risk from the uncertainty (i.e., the standard devia
tion) in the recovery rate value. Note from the calculation that the mean and standard
deviation for the portfolio are $213.63 and $3.35 respectively.
Recall from
Chapter 2
that the mean and standard deviation of our BBB bond were
$107.09 and $2.99 respectively. For the singleA bond the comparable statistics are a
mean of $106.55 and a standard deviation of $1.49. So we see that the means or
expected values sum directly, but the risk – as measured by standard deviations – is
much less than the summed individuals due to diversiﬁcation.
At this point, although we have only discussed the calculation of standard deviation for a
twoasset portfolio, we have presented all of the components necessary to calculate the
standard deviation for any portfolio. For an arbitrary portfolio, we ﬁrst identify all pairs
of assets. We then consider each pair of assets as a subportfolio and compute its variance
using the methods described in this section. Finally, we combine these variances with
the variances for individual assets and arrive at a portfolio standard deviation. The
details of this calculation are discussed in
Chapter 9
and
Appendix A
.
After having calculated the portfolio standard distribution, we next calculate a second
measure of credit risk, that is, the percentile level. Assume that we are interested in cal
culating the 1
st
percentile level. Again, we point out that there is no ﬁxed rule to prefer
any given percentile level over another.
Mean: µ
Total
p
i
µ
i
213.63 =
i 1 =
S 64 =
∑
=
Variance: σ
Total
2
p
i
µ
i
2
µ
Total
2
11.22 {std. dev. is 3.35} = –
i 1 =
S 64 =
∑
=
40 Chapter 3. Portfolio risk calculation
CreditMetrics™—Technical Document
Recall from
Chapter 1
and
Chapter 2
that this calculation is quite simple. All we have to
do is to ﬁnd the portfolio value such that the likelihoods of all the values less than this
sum to 1%. Since in this case, the portfolio has no more than two assets, we may simply
examine the probabilities and values shown in
Table 3.2
, and obtain a 1
st
percentile level
number of $204.40. This is $9.23 below the mean value. Of course, for larger portfo
lios, it is not possible to calculate percentile levels analytically – we would have to per
form a simulation.
This ﬁnishes our discussion of the calculation of the credit risk measures for the example
portfolio of two bonds. In the next section, we introduce the concept of marginal risk.
This concept enables us to understand where the risks are concentrated in the portfolio,
and with which exposures we beneﬁt due to diversiﬁcation.
3.3 Marginal risk
We saw in
Chapter 2
how the credit risk can be calculated for an individual bond on a
stand alone basis. However, the decision to hold a bond or not is likely to be made
within the context of some existing portfolio. Thus, the more relevant calculation is the
marginal increase to the portfolio risk that would be created by adding a new bond to it.
Let us ﬁrst illustrate the calculation of marginal risk by using the standard deviation as a
risk measure. Recall that the standard deviation of our onebond (BBBrated) portfolio
in
Chapter 2
was $2.99. The portfolio standard deviation increased to $3.35 once we
added the second, singleA rated bond. The marginal standard deviation of this second
bond is therefore equal to $0.36, which represents the difference between $3.35 and
$2.99. Note that this marginal standard deviation is much smaller than the standalone
standard deviation of the second bond, which is $1.49. This is because of the diversiﬁca
tion effect that is in turn caused by the fact that the yearend values of the individual
bonds are not perfectly correlated.
We describe next how we extend our marginal risk calculation to percentile levels, again
with the caveat that this approach is most appropriate for large portfolios. Recall that the
BBBrated bond had a mean value of $107.09 and a 1
st
percentile level value of $98.10.
This percentile level is therefore $8.99 below the mean. Once the singleA rated bond is
added, the twobond portfolio has a mean of $213.63 and a 1
st
percentile level of 204.40.
This percentile level is $9.23 below the mean. We can now calculate the marginal risk of
the singleA rated bond as the difference between $9.23 and $8.99, which is equal to
$0.24. On a standalone basis, the singleA rated bond has a 1
st
percentile level value of
$103.15, which is $3.39 below the mean value of $106.55. This difference between the
marginal risk ($0.24) and the standalone risk ($3.39) is again due to diversiﬁcation.
We remark that marginal risk statistics are sometimes deﬁned in a slightly different way.
Where we deﬁne marginal risk to be the contribution of one asset to the total portfolio
risk, others deﬁne it to be the marginal impact on portfolio risk of increasing an exposure
by some small amount. While the two deﬁnitions do differ, they both serve the purpose
of measuring an exposure’s risk contribution to a portfolio, accounting for the effects of
diversiﬁcation, or lack thereof.
This concludes our discussion of marginal risk. In the next chapter, we show how
CreditMetrics treats other asset types: receivables, loans, loan commitments, ﬁnancial
letters of credit and marketdriven instruments such as swaps and forwards.
41
Chapter 4. Differing exposure types
So far we have demonstrated the methodology used in CreditMetrics with the help of
portfolios consisting of only bonds. In
Chapter 2,
we discussed the case of a standalone
bond exposure. In
Chapter 3
, we extended this standalone credit risk methodology to a
portfolio of two bonds.
As discussed in
Chapter 1,
however, CreditMetrics is not limited to bonds. Rather,
CreditMetrics is capable of estimating most any credit risk type limited only by the data
available to revalue exposures upon up(down)grade and default. As a matter of imple
mentation, we have included the following generic exposure types:
1. noninterest bearing receivables (trade credit);
2. bonds and loans;
3. commitments to lend;
4. ﬁnancial letters of credit; and
5. marketdriven instruments (swaps, forwards, etc.)
This chapter explains how CreditMetrics addresses different exposure types. Recall that
for bonds, the analysis consisted of two steps: (i) specifying the likelihoods and joint
likelihoods of obligors experiencing a credit quality change, and (ii) calculating the new
values given each possible rating change at the risk horizon. For each of the additional
exposure types covered in this chapter, the ﬁrst step is identical as for bonds. Thus, here
we only need to describe how to revalue each new exposure type in the event of a rating
change or default. Thus, the goal of this chapter is to provide the reader with methods
for constructing versions of
Table 1.2
for a variety of exposure types.
We show below in
Chart 4.1
the ﬁnal “road map” for credit risk analytics within Credit
Metrics. Note that this map now includes the provision for different exposure types.
Chart 4.1
Our ﬁnal “road map” of the analytics within CreditMetrics
Before we describe how we treat different exposure types in CreditMetrics, let us sum
marize the CreditMetrics analytics shown in the chart above.
Credit Rating Seniority Credit Spreads
Value at Risk due to Credit
bond revaluation
Present value
quality changes for a single exposure
Standard Deviation of value due to credit
Ratings series,
Equities series
Correlations
Models(e.g.,
correlations)
Portfolio Value at Risk due to Credit
Exposures
Market
volatilities
Exposure
distributions
User
Portfolio
Rating migration
likelihoods
Joint credit
rating changes
in default
Recovery rate
42 Chapter 4. Differing exposure types
CreditMetrics™—Technical Document
• The
credit exposure
is the amount subject to either changes in value upon credit
quality up(down)grade or loss in the event of default. In practice, we must estimate
this amount for: (i) commitments, where the exposure may change due to additional
drawdowns, and (ii) marketdriven instruments (swaps, forwards, etc.), where the
exposure depends on the movement of market variables (e.g., FX or interest rates).
• The obligor’s current longterm senior unsecured credit rating indicates its likeli
hood of credit quality migration that will be applied to all the obligor’s obligations.
• The seniority standing and instrument type of each transaction indicates the recovery
rate of that exposure in default.
• The transaction details and the forward zero rates for each credit rating category
determine the changes in value of each transaction upon obligor up(down)grade.
• The joint likelihoods of credit quality movements for any pair of obligors is esti
mated through a treatment of the correlation between obligors.
• The portfolio standard deviation of changes in value due to credit quality changes
are then calculated directly – in
closed form
.
In the following sections, we discuss how CreditMetrics treats different exposure types.
4.1 Receivables
Corporations which do not hold loans or bonds as exposures may still be subject to credit
risk through payments due from their customers. Non interest bearing receivables (also
called
trade credit
), are at risk to changes in credit quality of their customers. It is neces
sary, then, to consider such receivables on a comparable basis with any other risky credit
instrument.
In concept, we treat receivables in the same way as we treated the bonds in the previous
chapters. For receivables which become due beyond the risk horizon, we treat the cash
ﬂow as if it were a zero coupon bond paying on the receivable date, and revalue the cash
ﬂow based on the bond spreads in each rating category. If there are more applicable
spreads available, speciﬁc to receivables, it would certainly be reasonable to use these in
place of the bond spreads for the purposes of this revaluation.
Often, a receivable will be due before the risk horizon. In this situation, it is not even
necessary to revalue in different rating categories. Either the payment is made, and we
“revalue” at the receivable’s face amount, or there is a default, and we revalue based on
some recovery rate. Thus, in the case of a $1mm receivable due in nine months, where
the risk horizon is one year and the recovery rate is, say 30%, we revalue the exposure at
$1mm in each nondefault state, and at $300,000 (= 30% times $1mm) in default.
We know of no systematic study of recovery rate experience for corporate receivables
and so suggest that users take senior unsecured bond recovery experience as a guide.
Sec. 4.2 Bonds and loans 43
Part I: Risk Measurement Framework
4.2 Bonds and loans
In
Chapter 2
, we described how to revalue bonds in each future rating state using the for
ward interest rate curves for each credit rating. We generated a forward curve based on
the credit spread curve for that rating, and discounted the future cash ﬂows of the bond.
In the case of default, the bond’s value was taken to be a recovery fraction multiplied by
the face value of the bond.
In concept, we treat a loan as a par bond, revaluing the loan using loan forward curves
upon up(down)grade and applying a loan recovery rate to the principal amount in the
case of default. This revaluation upon up(down)grade accounts for the decreasing likeli
hood that the full amount of the loan will be repaid as the obligor undergoes rating down
grades, and the increasing likelihood of repayment if the obligor is upgraded.
As an alternative, bonds may be treated in the same way as the marketdriven instru
ments described in
Section 4.4
. In that section, we discuss the differences between the
two approaches.
4.3 Loan commitments
A loan commitment is composed of a drawn and undrawn portion. The drawdown on the
loan commitment is the amount currently borrowed. Interest is paid on the drawn por
tion, and a fee is paid on the undrawn portion. Typical fees on the undrawn portion are
presented in
Table 4.1
. When we revalue a loan commitment given a credit rating
change, we must therefore account for the changes in value to both portions. The drawn
portion is revalued exactly like a loan. To this we add the change in value of the
undrawn portion. As a practical matter, each lending institution is the best judge of its
own pricing for loan commitments. Thus, it is quite appropriate for each institution to
utilize its own pricing for this revaluation rather than using generic spreads in a down
loadable data set.
Table 4.1
Fee on undrawn portion
of commitment (b.p.)
Because loan commitments give the obligor the option of changing the size of a loan,
loan commitments can dynamically change the portfolio composition. The amount
drawn down at the risk horizon is closely related to the credit rating of the obligor (see
Table 4.2
). For example, if an obligor deteriorates, it is likely to draw down additional
funds. On the other hand, if its prospects improve, it is unlikely to need the extra bor
rowings.
Yearend
rating
Fee: undrawn
portion
AAA 3
AA 4
A 6
BBB 9
BB 18
B 40
CCC 120
44 Chapter 4. Differing exposure types
CreditMetrics™—Technical Document
Note that it is not uncommon for loan commitments to have covenants that can reduce
the credit risk. For example, if the loan rate not only ﬂoats with interest rate levels but
also has credit spreads which change upon up(down)grade – a repricing grid – then the
value of the facility will remain essentially unchanged across all up(down)grade catego
ries. Thus, the risk will have been reduced because the only volatility of value remaining
will be the potential loss in the event of default.
The worst possible case for a commitment is that the counterparty draws down the full
amount and then defaults. It is intuitive, then, to treat a commitment as if it were a loan,
with principal equal to the full commitment line. This is certainly the simplest approach
to commitments, and from a risk perspective, the most conservative.
In practice, it has been seen that commitments are not always fully drawn in the case of
default, and hence, that the risk on a commitment is less than the risk of a fully drawn
loan. In order to model commitments more accurately, it is necessary to estimate not only
the amount of the commitment which will be drawn down in the case of default, but also
the amount which will be drawn down (or paid back) as the counterparty undergoes
credit rating changes. For purposes of explanation in this section, we will rely on one
study of commitment usage, but this is certainly an area where each ﬁrm should apply its
own experience.
In the remainder of this section, we present a framework to calculate, given one source of
data on commitment usage, the change in value of a commitment in each possible credit
rating migration. We do this via an example.
Consider a threeyear $100mm commitment to lend at a ﬁxed rate (on the drawn portion)
of 6% to a currently A rated obligor. For ease of illustration, we will assume the credit
spreads of
Section 1.3.2
, so that the change in value of any drawn amount due only to
changes in credit quality (and neglecting changes in commitment usage) will be the same
as for bond example in that section. Our example will have $20mm currently drawn
down with the remaining $80mm undrawn and charged at a fee of 6 b.p.
In each credit rating, the commitment’s revaluation estimate at the risk horizon will
depend on both:
• estimates of the change in amount drawn due to credit quality changes; and
• estimates of the change in value for both the drawn and undrawn portions.
We will address each of these in turn. Our ﬁrst task is to estimate changes in drawdown
given each possible credit rating change. Part of this estimation, the drawdown in
default, can be directly taken from a published study. Asarnow & Marker (A&M) [95]
have examined the average drawdown (of normally unused commitment) in the event of
default (see
Table 4.2
). We may use this information to estimate how much of the
undrawn amount of our commitment will be drawn in the case of a default. Thus, if our
counterparty defaults, it will draw an additional 71% of the $80mm undrawn amount, or
$56.8mm. Added to the current drawn amount of $20mm, this results in an estimated
drawdown in default of $76.8mm.
Sec. 4.3 Loan commitments 45
Part I: Risk Measurement Framework
Table 4.2
Average usage of commitments to lend
Source: Asarnow & Marker [95]
We also expect that there will be a credit rating related change in drawdown in all non
default states. One suggestion of this behavior is also evidenced in the A&M study. We
see that the average commitment usage varies directly with credit rating. How to best
apply this information is open to question. We offer the following method as a sugges
tion to initiate discussion rather than as a deﬁnitive result. We fully anticipate that, as a
matter of implementation, each institution will substitute its own study based on its
unique experience.
The following table estimates the changes in drawdown attributable to credit rating
changes. As an example, consider the change to BBB from our initial rating of singleA.
A&M found that the average draw increased from 4.6% to 20.0% between singleA and
BBB. In other words, the undrawn portions moved from 95.4% to 80.0% or a reduction
of 16.1% (= 100%  [80.0%/95.4%]). Thus, with our example initial drawdown of 20%,
we estimate that an additional 12.9% (=16.1% * 80%) will be drawn down in the case of
a migration to BBB. We present estimates of change in drawdown for all possible migra
tions in
Table 4.3
. Note that in cases of upgrades, we actually estimate a negative change
in drawdown, corresponding to the counterpart paying back some amount.
Table 4.3
Example estimate of changes in drawdown
It is against this estimate of the new drawdown amounts at the risk horizon that we now
apply our revaluation estimates – to both the drawn and undrawn portions. Referring to
Table 4.4
, we see for instance that when the obligor downgrades to BB, the change in
Credit
rating
Average
commitment
usage
Usage of the normally
unused commitment in
the event of default
AAA 0.1% 69%
AA 1.6% 73%
A 4.6% 71%
BBB 20.0% 65%
BB 46.8% 52%
B 63.7% 48%
CCC 75.0% 44%
Yearend
rating
Current
Drawdown
Change in
Drawdown
Estimate of
New Drawdown
AAA 20.0 19.6 0.4
AA 20.0 13.0 7.0
A 20.0 0.0 20.0
BBB 20.0 12.9 32.9
BB 20.0 35.4 55.4
B 20.0 49.6 69.6
CCC 20.0 59.0 79.0
Default 20.0 56.8 76.8
46 Chapter 4. Differing exposure types
CreditMetrics™—Technical Document
value is negative 2.7%, or $1.5mm due to credit spread widening on the ﬁxed rate dawn
portion.
There is also a change in the value of the fees collected on the undrawn amount. Recall
that at present, we are collecting 6 b.p. on the undrawn amount of $80mm. As the
undrawn amount changes, so too do the fees we collect. Thus in the case of downgrade to
BB, we collect fees on $35.4mm less. (This $35.4mm corresponds to the additional draw
in this case, as seen in
Table 4.3
.) In default, we lose the all of the fees which we were
receiving, the full 6 b.p. on the full $80mm. The change in value of the fees
1
is also given
in
Table 4.4
, as well as the total change in value for the commitment, for all possible rat
ing migrations.
Table 4.4
Revaluations for $20mm initially drawn commitment
Several observations are in order:
• the expected percentage drawn down in default is the most important factor;
• fees have a relatively small impact on the revaluations;
• it is possible to have negative revaluations greater than the current drawdown; and
• covenants that reset the drawdown spread upon an up(down)grade would reduce the
volatility of value in all nondefault states – keeping value close to par.
Thus we have obtained the revaluations estimates in each future credit rating state. The
calculation of credit risk for the commitment is now simply a matter of applying the
techniques of the previous two chapters.
4.4 Financial letters of credit (LCs)
There are times when an obligor may desire to have the option to borrow even if there is
no immediate need to borrow. In such a case, an outright loan would be inefﬁcient since
its proceeds may sit in the obligor’s hands underemployed. What is typically needed in
this case is a
ﬁnancial letter of credit
. With this type of off balance sheet access to
1
To be precise, the change in fees should also account for a new discount function corresponding to the new rating,
but we ignore this effect for this example.
Yearend
rating
Drawdown
at Yearend
Change in
Value (%)
Change in
Value ($mm)
Change in
fees ($mm)
Total Value
Change ($mm)
AAA 0.4 0.6 0.0 0.01 0.01
AA 7.0 0.5 0.0 0.01 0.01
A 20.0 0.3 0.1 0.00 0.10
BBB 32.9 0.3 0.1 0.01 0.11
BB 55.4 2.7 1.5 0.02 1.52
B 69.6 4.3 3.0 0.03 3.03
CCC 79.0 16.3 12.9 0.04 12.94
Default 76.8 51.8 39.8 0.05 39.85
Sec. 4.5 Marketdriven instruments 47
Part I: Risk Measurement Framework
funds, there is assurance that funds will be available even when other sources of funding
may dry up due perhaps to credit quality deterioration.
We argue that such an exposure is comparable in risk to an outright term loan Indeed,
the market typically prices these instruments comparable to loans. The obligor will
almost assuredly drawdown the LC as it approaches credit distress. Thus, in all the cases
where there can be a default, there will also be full exposure just like a loan. We suggest
that LCs be treated identically to loans, including the use of the credit spread curves and
recovery rates that have been estimated for loans.
Note that a
ﬁnancial
letter of credit is distinguished from either
performance
or a
trade
letter of credit. Performance LCs are typically secured by the income generating ability
of a particular project and trade LCs are triggered only infrequently by non credit related
events. Both of these would have smaller risk than ﬁnancial LCs.
4.5 Marketdriven instruments
We have so far described how CreditMetrics calculates the credit risk for receivables,
bonds, loans, and commitments. In this section, we explain how CreditMetrics treats
derivative instruments that are subject to counterparty default (swaps, forwards, etc.).
Our discussion focuses on the example of swaps.
Throughout this document, we refer to these types of exposures as
marketdriven
instru
ments. In these transactions, credit risk and market risk components are intimately cou
pled because of an inherent optionality. This optionality stems from the fact that we face
a loss on the transaction if the counterparty defaults
only
if we are inthemoney (i.e., the
obligor owes us money on a net present value basis). This complicates CreditMetrics’
handling of derivatives exposures.
We remark that to treat products like swaps in full detail, it would be necessary to pro
pose an integrated model of credit and market risk. Such a model would describe both
the correlations of swap exposures across a portfolio (capturing, for instance, that swaps
based on the same interest rate would tend to go in or outofthemoney together), and
the correlations between credit and market moves (for instance, that swap counterparties
might be more likely to default in one interest regime than in another). Our goal here is
not to provide a fully integrated model, but to capture the most crucial inﬂuences of mar
ket volatilities to the credit risks of marketdriven instruments.
4.5.1 Credit risk calculation for swaps
Swaps are treated within CreditMetrics consistent with the way bonds and loans are
treated. However, the revaluation of swaps in each credit quality state at the risk horizon
is much more complicated than that of either bonds or loans. Credit loss occurs when
both of the following two conditions are satisﬁed:
1. The counterparty undergoes a credit quality change.
2. The swap transaction is outofthemoney for the counterparty, that is, the counter
party owes money on the swap transaction on a net present value basis.
48 Chapter 4. Differing exposure types
CreditMetrics™—Technical Document
The market and credit risk calculations for swaps are therefore intimately related. All
things remaining equal, the greater the market volatility, the greater the amount exposed
to loss during an unfavorable credit event.
To summarize, optionality (i.e., credit exposure to the swap counterparty only if the
counterparty is outofthemoney) is the feature that makes swaps distinct from bonds.
Although the exposure in the case of bonds is also marketdriven, there is no optionality
involved, since the issuer of debt is always “outofthemoney.” In other words, the net
present value for swaps can be either positive or negative for the counterparties. How
ever, the net present value for the issuer is always negative.
We next describe how the swap is reevaluated in each possible rating state at the risk
horizon. The purpose of this exercise is to ﬁll the “value” table analogous to
Table 2.4
for bonds. Essentially, we represent the value of the swap as a difference of two compo
nents:
1. The ﬁrst component is equal to the forward
riskfree
2
value of the swap cash
ﬂows. This hypothetical value is obtained by ﬁnding the forward value of the
swap cash ﬂows by using the government rates rather than the swap rates; there
fore the ﬁrst component is the same for all forward credit rating states.
2. The second component represents the loss expected on the swap due to a default
net of recoveries by the counterparty on the remaining cash ﬂows of the swap. By
“remaining” we mean all cash ﬂows that occur after the risk horizon (assumed to
be one year). Since the probability of this default varies by rating category, the
second component varies from one rating category to another.
Finally, the revaluation of the swap in any rating category is obtained by subtracting the
second (expected default loss) component from the ﬁrst (riskfree value) component.
Note that this valuation scheme essentially values the swap as if it were riskfree, and
then subtracts a penalty (the expected loss) to account for the risk due to the credit qual
ity of the counterparty. The intuition behind this procedure for calculating the swap value
is straightforward. First we calculate the value assuming that there is no risk whatsoever
of the counterparty’s default, using the government (i.e., creditriskfree) rates for this
calculation. We then subtract from this creditriskfree value the amount that we can
expect to lose due to a counterparty default. The probability of default is obviously an
important factor driving this latter component. A second factor is
optionality,
which is
implied in the fact that we have exposure to the counterparty only if the counterparty is
inthemoney. The value of this optionality is determined from the amount by which the
swap is expected to be inthemoney, and also from the volatility of interest rates.
An enhancement of this procedure might be to account for not only the expected loss due
to credit, but also for the random nature of the swap exposure. This could be achieved by
redeﬁning the expected loss penalty to include a measure of how much the swap is likely
to ﬂuctuate in value. The result would be that two swaps with the same expected losses
in each rating state would be distinguished by the amount of uncertainty in their losses.
2
Here we mean riskfree from a credit risk perspective.
Sec. 4.5 Marketdriven instruments 49
Part I: Risk Measurement Framework
We next detail the calculation of the two components of the swaps value at the risk hori
zon in each possible credit rating state. First, however, let us restate mathematically the
revaluation of the swap:
[4.1]
Value of swap in 1 year
Rating R
= Riskfree value in 1 year
– Expected loss in years 1 through maturity
Rating R
where “R” in the above expression can be any possible credit rating category including
default.
As mentioned before, the calculation of the riskfree value of the swap in one year is
straightforward. All we need do is discount the future swap cash ﬂows occurring
between Year 1 through maturity by the forward government zero curve. The calcula
tion of the expected loss, on the other hand, is complicated. This is because of the
optionality component in swap exposure explained earlier.
For each forward nondefault credit rating, the expected loss can be written as:
[4.2]
Expected loss
Rating R
= Average exposure
Year 1 through maturity
·
Probability of default in years 1 through maturity
Rating R
·
(1–Recovery Fraction)
The average exposure represents the average
3
of several expected exposure values calcu
lated at different forward points over the life of the swap starting from the end of ﬁrst
year. We use average exposure in the expected loss expression above to account for the
possibility of swap counterparty defaulting at any point in time between the end of the
ﬁrst year and maturity. Each of the expected exposure values that enter the average
exposure calculation requires a modiﬁed BlackScholes computation to account for the
inherent optionality feature. As a result, the average exposure calculation for swaps is
quite complicated and timeconsuming. We refer the interested reader to other sources
for a more thorough treatment of the expected and average exposure calculations.
4
The second term that enters the expected loss calculation is the probability of default for
each rating category between Year 1 and the maturity of the swap. For example, if the
maturity of the swap is ﬁve years, then the fouryear probability of default is required for
each of the rating categories AAA through CCC. These probabilities can be obtained by
multiplying the oneyear transition matrix four times to generate the fouryear transition
matrix. The fouryear default probabilities can then be simply read off from the last (i.e.,
default) column of this transition matrix.
Two assumptions are implicit in this method of generating the longterm default proba
bilities. First, we assume that the transition process is stationary in that the same transi
tion matrix is valid from one year to another. Second, we assume that there is no
autocorrelation in rating movements from one year to another.
3
The expected exposures are weighted by the appropriate discount factors for this average calculation.
4
One such source is “On measuring credit exposure,”
RiskMetrics
™
Monitor,
J.P. Morgan, March 1997. Also,
J.P. Morgan plans to provide a software tool in the near future that enables the user to calculate average and
expected exposures. This tool will be based on the RiskMetrics market risk methodology, a software implementa
tion of which is currently being marketed by J.P. Morgan under the name FourFifteen.
50 Chapter 4. Differing exposure types
CreditMetrics™—Technical Document
The method provided so far enables us to calculate the value of the swap in each of the
nondefault credit rating categories. To calculate the value in case of a default during the
risk interval, we must modify this procedure somewhat. This is mainly due to the fact
that the average exposure calculation over the life of the swap does not make any sense
here, since we know for sure that the swap counterparty has defaulted during the risk
interval. Therefore we write the expected loss in the defaulted state as:
[4.3]
Expected loss
Default
= Expected exposure
Year 1
·
(
1
– Recovery Fraction).
The implicit assumption in the above expression is that the risk interval is relatively
short, say one year, as compared to the maturity of the swap. If the risk interval is much
longer than this, say several years, it will be more accurate to replace the expected expo
sure with the average exposure value calculated over the risk interval. This is because
the swap counterparty can default at any point over the longer risk interval, and the
expected exposures at these points can be very different from the expected exposure at
the risk horizon. In this case, therefore, the average exposure is a more suitable measure.
This concludes our explanation of the value calculation for swaps at the risk horizon.
Let us next consider an example.
Assume a threeyear ﬁxed for ﬂoating swap on $10 mm notional beginning
January 24, 1997. Let the risk horizon be one year and the recovery rate in case of
default be 0.50.
On January 24, 1997, the average exposure at the end of one year is calculated to be
equal to be $61,627. This represents an average of the expected exposures between the
end of one year and the end of three years (a twoyear time period). Now, given a two
year default likelihood of 0.02% for the AA rating category, the value of the swap at the
end of risk horizon in the AA rating category is equal to:
[4.4]
FV in 1 year – p
AA
·
AE
·
(
1
– R) = FV in 1 year –
0.0002
·
61,627
·
(1 – 0.5)
= FV –
$6
where
FV
refers to the forward value, and
AE
refers to the average exposure in one year.
Similarly, given a 33.24% default likelihood for the CCC rating category, the corre
sponding value of the swap in the CCC rating category is equal to:
[4.5]
FV in 1 year – p
CCC
·
AE
·
(
1
– R) = FV in 1 year –
0.3344
·
61,627
·
(1 – 0.5)
= FV
–
$10,304
.
Next, let us consider what happens in default. The expected exposure at the end of the
year is calculated on January 24, 1997 to be equal to $101,721. Given a recovery rate of
50%, the value in the defaulted state is equal to:
[4.6]
FV in 1 year – EE
·
(
1–
R) = FV in 1 year –
101,721
·
(1–0.5)
= FV –
$50,860
where
EE
refers to the expected exposure in one year.
In
Table 4.5
we summarize the value of the swap in each possible credit rating states at
the risk horizon. We do not specify the riskfree component (FV), for two reasons:
Sec. 4.5 Marketdriven instruments 51
Part I: Risk Measurement Framework
1. This calculation is relatively straightforward and involves valuing the future cash
ﬂows with the riskfree yield.
2. More importantly, it is conceivable that, at least for the lower credit ratings and
default, the expected loss value far exceeds the riskfree forward value of the swap
itself. This is especially true when the swap value is near par, the risk interval is
quite small, and the interest rates are not changing too rapidly. In this circum
stance, it sufﬁces to set the riskfree value term to zero and just use the expected
loss term in the credit risk calculation.
Table 4.5
Value of swap at the risk horizon in each rating state
“FV” represents the riskfree forward value of the swap
cash ﬂows in one year.
We next discuss a reﬁnement to the calculation of the expected loss value in rating states
AAA through CCC that produces more accurate expected loss numbers. For the sake of
clarity we did not address it earlier; we now explain it below.
Recall that we calculate the expected loss value for rating states AAA through CCC by
multiplying the average exposure by the probability of default for the desired rating cate
gory. Both the average exposure and the default probability are valid from Year 1
through the maturity of the swap. Also, the average exposure represents the average of
the expected exposures calculated at several points between Year 1 and maturity.
Given this, we can more accurately calculate the expected loss component as follows:
1. Calculate the expected exposure in, say, oneyear increments between Year 1 and
maturity.
2. Weigh each of the expected exposures by a probability factor. This factor repre
sents the probability that the counterparty defaults in the year in which the
expected exposure is calculated, given that it does not default before then.
3. Add these weighted expected exposures after adjusting for the time value of
money effect.
The result is a expectedloss calculation which reﬂects reality more accurately than if we
were simply to multiply the average exposure by a single default probability. This is
because by thus breaking the expected loss calculation into smaller pieces at different
time horizons, we properly account for the timing of default.
Yearend
rating
Twoyear default
likelihood (%) Value ($)
AAA 0.00 FV – 1
AA 0.02 FV – 6
A 0.15 FV – 46
BBB 0.48 FV – 148
BB 2.59 FV – 797
B 10.41 FV – 3,209
CCC 33.24 FV – 10,304
Default — FV – 50,860
52 Chapter 4. Differing exposure types
CreditMetrics™—Technical Document
4.5.2 Extension for forwards and multiple transactions
The methodology that we have presented above for swaps can be used in exactly the
same manner for forwards. Furthermore, it can be easily extended to the case in which
there are several transactions with the same counterparty and netting is enforceable.
These transactions do not all have to be swaps, but can represent a variety of market
driven instruments, including forwards.
The methodology outlined for swaps can be extended to a portfolio of different instru
ment types with the same counterparty as follows. First, all the cash ﬂows from the dif
ferent transactions conducted by the same counterparty are netted to yield the resulting
net cash ﬂows. (Of course, this netting is done according to the particular netting
arrangements that are in place with the counterparty.) Next, the swaps methodology is
used to revalue these net cash ﬂows in different rating categories at the risk horizon.
Once again, this value comprises riskfree forward value and a expected loss value, both
of which are calculated in exactly the same manner as for swaps.
53
Part II
Model Parameters
54
CreditMetrics™—Technical Document
55
Part II: Model Parameters
Overview of Part II
We have seen in the previous section the general overview, scope and type of results of
CreditMetrics. Now we will give more detail to the main modeling parameters used in
the CreditMetrics calculation: our sources of data, how we use the data to estimate
parameters and why we have made some of the modeling choices we did. There is no
single step in the methodology that is particularly difﬁcult; there are simply a lot of
steps. We devote a chapter to each major parameter and have tried to present each chap
ter as a topic which can be read on its own. Although we encourage the reader to study
all chapters, reading only a particular chapter of interest is also possible.
Part II is organized into four chapters providing a detailed description of the major
parameters within the CreditMetrics framework for quantifying credit risks. Our intent
has been to make this description sufﬁciently detailed so that a practitioner can indepen
dently implement this model. This section is organized as follows:
•
Chapter 5: Overview of credit risk literature.
To better place our efforts within
the context of prior research in the credit risk quantiﬁcation ﬁeld, we give a brief
overview of some of the relevant literature.
•
Chapter 6: Default and credit quality migration.
We present an underlying
model of the ﬁrm
within which we integrate the process of ﬁrm default and, more
generally, credit quality migrations. We argue that default is just a special case of a
more general process of credit quality migration.
•
Chapter 7: Recovery rates.
Since changes in value are – naturally – greatest in the
state of default, our overall measure of credit risk is sensitive to the estimation of
recovery rates. We also model the uncertainty of recovery rates.
•
Chapter 8: Credit quality correlations.
The portfolio view of any risk requires an
estimation of – most generally – joint movement. In practice, this often means esti
mating correlation parameters. CreditMetrics requires the joint likelihood of credit
quality movements between obligors. Since the observation of credit events are
often rare or of poor quality, it is difﬁcult to further estimate their correlations of
credit quality moves. We show that the results of several different data sources cor
roborate each other and might be used to estimate credit quality correlations.
56
CreditMetrics™—Technical Document
57
Chapter 5. Overview of credit risk literature
One of our explicit goals is to stimulate broad discussion and further research towards a
better understanding of quantitative credit risk estimation within a full portfolio context.
We have sought to make CreditMetrics as competent as is possible within an objective
and workable framework. However, we are certain that it will improve with comments
from the broad community of researchers.
Extensive previous work has been done towards developing methodologies for estimat
ing different aspects of credit risk. In this chapter, we give a brief survey of the aca
demic literature so that our effort with CreditMetrics can be put in context and so that
researchers can more easily compare our approach to others. We group the previous aca
demic research on credit risk estimation within three broad categories:
• estimating particular individual parameters such as expected default frequencies or
expected recovery rate in the event of default;
• estimating volatility of value (often termed
unexpected
losses) with the assumption
of bond market level diversiﬁcation; and
• estimating volatility of value within the context of a speciﬁc portfolio that is not per
fectly diversiﬁed.
Also, there have been several papers on credit
pricing
, starting with Merton [74], which
discuss debt value as a result of ﬁrm risk estimation in an optiontheoretic framework.
There is more recent work in this area which has focused on incorporating corporate
bond yield spreads in valuation models, see Ginzburg, Maloney & Willner [93], Jarrow,
Lando & Turnbull [96] and Das & Tufano [96]. For CreditMetrics, we have chosen to
focus on the risk assessment side rather than focus on the pricing side.
5.1 Expected losses
Expected losses are driven by the expected probability of default and the expected recov
ery rate in default. We cover recovery rate expectations in much more detail in
Chapter
7
and so will devote this discussion to the expected default likelihood. The problem of
estimating the chance of counterparty default has been so difﬁcult that many systems
devote all their efforts to this alone. Certainly, if the underlying estimates of default
likelihood are poor, then a risk management system is unlikely to make up for this deﬁ
ciency in its other parts. We will discuss three approaches that are used in practice:
• the accounting analytic approach which is the method used by most rating agencies;
• statistical methods which encompass quite a few varieties; and
• the optiontheoretic approach which is a common academic paradigm for default.
We emphasize that CreditMetrics is not another rating service. We assume that expo
sures input into CreditMetrics will already have been labeled into discrete rating catego
ries as to their credit quality by some outside provider.
58 Chapter 5. Overview of credit risk literature
CreditMetrics™—Technical Document
As we discuss in
Chapter 6,
a transition matrix for use by CreditMetrics can be ﬁt to any
categorical rating system which has historical data. Indeed, we would argue that each
credit scoring system should be ﬁt with its own transition matrix. For some users with
their own internal rating systems, this will be a necessary ﬁrst step before applying
CreditMetrics to their portfolios. If these systems have limited historical data sets avail
able, then an estimation algorithm that expresses longterm behavior may be desirable
(see
Section 6.4
).
5.1.1 Accounting analytic approach
Perhaps the most widely applied approach for estimating ﬁrm speciﬁc credit quality is
fundamental analysis with the use of ﬁnancial ratios. Such
accounting analytic
methods
focus on leverage and coverage measures, coupled with an analysis of the quality and
stability of the ﬁrm’s earnings and cash ﬂows. A good statement of this approach is in
Standard and Poor’s
Debt Rating Criteria
.
1
These raw quantitative measures are then
tempered by the judgment and experience of an industry specialist. This broad descrip
tion is generally the approach of the major debt rating agencies. This approach yields
discrete ordinal groups (e.g., alphabetic ratings) which label ﬁrms by credit quality.
We are aware of at least 35 credit rating services worldwide. Also, it is common that
ﬁnancial institutions will maintain their own inhouse credit rating expertise. However,
letter (or numerical) rating categories by themselves only give an ordinal ranking of the
default likelihoods. A quantitative credit risk model such as CreditMetrics cannot utilize
ratings without additional information. Each credit rating label must have a statistical
meaning such as a speciﬁc default probability (e.g., 0.45% over a oneyear horizon).
The two major U.S. agencies, S&P and Moody’s, have published historical default likeli
hoods for their letter rating categories. An example from Moody’s is shown in
Table 5.1
.
There have been many studies of the historical default frequency of corporate publicly
rated bonds. These include Altman [92], [88], [87], Altman & Bencivenga [95], Altman
& Haldeman [92], Altman & Nammacher [85], Asquith, Mullins & Wolff [89], Carty &
Lieberman [96a] and S&P CreditWeek [96]. These studies are indispensable, and it is
important to highlight some important points from them:
• the evolution and change in the original issue high yield bond market is unique in its
history and future high yield bond issuance will be different;
• most of the default history is tagged to U.S. domestic issuers who are large enough
to have at least an S&P or Moody’s rating; and
• the deﬁnition of “default” has itself evolved (e.g., it now typically includes “dis
tressed exchanges”).
Thus, use of these data must be accompanied by a working knowledge of how they were
generated and what they represent.
1
See: http://www.ratings.standardpoor.com/criteria/index.htm
Table 5.1
Moody’s corporate bond
average cumulative default
rates (%)
Source: Carty & Lieberman [96a]
— Moody’s Investors Service
Years 1 2 3 4 5
Aaa 0.00 0.00 0.00 0.07 0.23
Aa1 0.00 0.00 0.00 0.31 0.31
Aa2 0.00 0.00 0.09 0.29 0.65
Aa3 0.09 0.15 0.27 0.42 0.60
A1 0.00 0.04 0.49 0.79 1.01
A2 0.00 0.04 0.21 0.57 0.88
A3 0.00 0.20 0.37 0.52 0.61
Baa1 0.06 0.39 0.79 1.17 1.53
Baa2 0.06 0.26 0.35 1.07 1.70
Baa3 0.45 1.06 1.80 2.87 3.69
Ba1 0.85 2.68 4.46 7.03 9.52
Ba2 0.73 3.37 6.47 9.43 12.28
Ba3 3.12 8.09 13.49 18.55 23.15
B1 4.50 10.90 17.33 23.44 29.05
B2 8.75 15.18 22.10 27.95 31.86
B3 13.49 21.86 27.84 32.08 36.10
Sec. 5.1 Expected losses 59
Part II: Model Parameters
On a more macroeconomic level, researchers have found that aggregate default likeli
hood is correlated with measures of the business and credit cycle. For example,
Fons [91] correlates aggregate defaults to GDP, while Jónsson & Fridson [96] examine
also corporate proﬁts, manufacturing hours, money supply, etc.
5.1.2 Statistical prediction of default likelihood
There is a large body of more statistically focused work devoted to building credit qual
ity estimation models, which seek to predict future default. One can identify three basic
approaches to estimating default likelihood: qualitative dependent variable models, dis
criminant analysis, and neural networks. All of these approaches are strictly quantitative
and will at least yield a ranking of anticipated default likelihoods and often can be tuned
to yield an estimate of default likelihood.
Linear
discriminant analysis
applies a classiﬁcation model to categorize which ﬁrms
have defaulted versus which ﬁrms survived. In this approach, a historical sample is com
piled of ﬁrms which defaulted with a matched sample of similar ﬁrms that did not
default. Then, the statistical estimation approach is applied to identify which variables
(and in which combination) can best classify ﬁrms into either group. The best example
of this approach is Edward Altman’s Zscores; ﬁrst developed in 1968 and now offered
commercially as Zeta Services Inc. This approach yields a continuous numerical score
based on a linear function of the relevant ﬁrm variables, which – with additional process
ing – can be mapped to default likelihoods.
The academic literature is full of alternative techniques ranging from principal compo
nents analysis, selforganizing feature maps, logistic regression, probit/logit analysis and
hierarchical classiﬁcation models. All of these methods can be shown to have some abil
ity to distinguish high from low default likelihoods ﬁrms. Authors who compare the
predictive strength of these diverse techniques include Alici [95], Altman, Marco &
Varetto [93], and Episcopos, Pericli & Hu [95].
The application of
neural network
techniques to credit scoring include Dutta &
Shekhar [88], Kerling [95], and Tyree & Long [94]. The popular press reports commer
cial applications of neural networks to large volume credit decisions such as credit card
authorizations, but there do not appear to be commercial application yet of these neural
network techniques for large corporate credits.
5.1.3 Optiontheoretic approach
The
optiontheoretic
approach was proposed by Fisher Black and Myron Scholes in the
context of option pricing, and subsequently developed by Black, Cox, Ingersoll, and
most notably, Robert Merton. In this view, a ﬁrm has a market value which evolves ran
domly through time as new information about future prospects of the ﬁrm become
known. Default occurs when the value of the ﬁrm falls so low that the ﬁrm’s assets are
worth less than its obligations. This approach has served as an academic paradigm for
default risk, but it is also used as a basis for default risk estimation. The leading com
mercial exemplar of this approach is KMV. In general, this method yields a continuous
numeric value such as the number of standard deviations to the threshold of default,
which – with additional processing – can be mapped to default likelihoods
60 Chapter 5. Overview of credit risk literature
CreditMetrics™—Technical Document
5.1.4 Migration analysis
Understanding the potential range of outcomes that are possible is fundamental to risk
assessment. As illustrated in
Chart 5.1
, knowing today’s credit rating allows us to esti
mate from history the possible pattern of behaviors in the coming period. More speciﬁ
cally, if an obligor is BBB today, then chances are the obligor will be BBB in one year’s
time; but it may be up(down)graded.
Table 5.2
shows that, for instance, 86.93% of the
time a BBBrated obligor will remain a BBB, but there is a 5.30% chance that a BBB
will downgrade to a BB in one year.
Table 5.2
Credit quality migration likelihoods for a BBB in one year
One of our fundamental techniques is
migration analysis
. Morgan developed transition
matrices for our own use as early as 1987. We have since built upon a broad literature of
work which applies migration analysis to credit risk evaluation. The ﬁrst publication of
transition matrices was in 1991 by both Professor Edward Altman of New York Univer
sity and separately by Lucas & Lonski of Moody’s Investors Service. They have since
been published regularly (see Moody’s Carty & Lieberman [96a] and Standard & Poor’s
Creditweek
[15Apr96]) and can be calculated by ﬁrms such as KMV.
There have been studies of their predictive power and stationarity (Altman & Kao [91]
and [92]). More recently, several practitioners (see Austin [92], Meyer [95], and Smith
& Lawrence [95]) have used migration analysis to better estimate an accountingbased
allowance for loan and lease losses
(what we would term
expected default losses
). Also,
these tools have been used to both estimate (Crabbe [95]) and even potentially improve
Lucas [95b]) holding period returns. Finally, academics have constructed arbitrage free
credit pricing models (see Ginzburg, Maloney and Willner [93], Jarrow, Lando & Turn
bull [96] and Das & Tufano [96]). In CreditMetrics, we extend this literature by showing
how to calculate the volatility of value due to credit quality changes (i.e., the potential
magnitude of
unexpected
losses) rather than just expected losses.
5.2 Unexpected losses
The volatility of losses, commonly termed
unexpected
losses, has proven to be generally
much more difﬁcult to estimate than expected losses. Since it is so difﬁcult to explicitly
address correlations there have been a number of examples where practitioners take one
of two approaches. First, they have applied methods which are statistically easy by
addressing either the special case of correlations all equaling zero (perfectly uncorre
lated) or correlations all equaling one (perfectly positively correlated). Neither of these
is realistic.
Second, they have taken a middle road and assumed that their speciﬁc portfolio will have
the same correlation effects as some index portfolio. The index portfolio can either be
the total credit market (“full” diversiﬁcation) or a sector index. Thus, the hope would be
that statistics drawn from observing the index of debt might be applied through analogy
to the speciﬁc portfolio. The institution’s portfolio would be assumed to have the same
AAA AA A BBB BB B CCC Default
BBB 0.02% 0.33% 5.95% 86.93% 5.30% 1.17% 0.12% 0.18%
Chart 5.1
Credit migration
BBB
AAA
AA
A
BB
BBB
B
CCC
Sec. 5.2 Unexpected losses 61
Part II: Model Parameters
correlations and proﬁle of composition as the overall credit markets. These approaches
can be grouped into two categories which we discuss in turn:
• historical default volatility; and
• volatility of holding period returns.
Although these may yield some estimate of general portfolio risk, they both suffer from
an inability to do meaningful marginal analysis. These techniques would not allow the
examination of marginal risk brought by adding some speciﬁc proposed transaction.
There would also be no guide to know which speciﬁc names contribute disproportionate
risk to the portfolio.
5.2.1 Historical default volatility
Historical default volatility is available from public studies: see for example
Table 5.3,
which is taken from Carty & Lieberman [96a]. There are several hypotheses to explain
why default rates would be volatile:
• defaults are simply random events and the number of ﬁrms in the credit markets is
not large enough to smooth random variation;
• the volume of high yield bond issuance across years is uneven; and
• the business cycle sees more ﬁrms default during downturns versus growth phases.
All three hypotheses are likely to have some truth for the corporate credit markets.
Table 5.3
Volatility of historical default rates by rating category
Source: Carty &Lieberman [96a
] —
Moody’s Investors Service
The problem with trying to understand the volatility of individual exposures in this fash
ion is that it must be viewed within a portfolio.
5.2.2 Volatility of holding period returns
The volatility of default events is only one component of credit risk. Thus, it may also be
useful to examine the volatilities of total holding period returns. A number of academic
studies have performed this exercise. For corporate bonds, there are two studies by Ben
Default rate standard deviations (%)
Credit rating Oneyear Tenyear
Aaa 0.0 0.0
Aa 0.1 0.9
A 0.1 0.7
Baa 0.3 1.8
Ba 1.4 3.4
B 4.8 5.6
62 Chapter 5. Overview of credit risk literature
CreditMetrics™—Technical Document
nett, Esser & Roth [93] and Wagner [96]. For commercial loans, there are studies by
Asarnow [96] and Asarnow & Marker [95].
Once the historical return volatility is estimated – perhaps grouped by credit rating,
maturity bucket, and industry/sector – some practitioners have applied them to analogous
exposures in the credit portfolio. In this approach, portfolio diversiﬁcation is addressed
only to the extent that the portfolio under analysis is assuming to be analogous to the
credit market universe. Again, there is the obvious problem of diversiﬁcation differ
ences. But there are also three practical concerns with this approach:
• historical returns are likely to poorly sample returns given credit quality migrations
(including defaults) which are lowfrequency but important
2
;
• the data as it has been collected would require a standard deviation estimate over a
sample size of less than 30 and so the standard error of the estimate is large; and
• the studies listed have commingled all sources of volatility – including interest rate
ﬂuctuations – rather than just volatility in value due to credit quality changes.
This general approach is sometimes termed the RAROC approach. Implementations
vary, but the idea is to track a benchmark corporate bond (or index) which has liquidity
and observable pricing. The resulting estimate of volatility of value is then used to proxy
for the volatility of some exposure (or portfolio) under analysis.
Potential problems with this approach arise because of its relative inefﬁciency in esti
mating infrequent events such as up(down)grades and defaults. Observing some bench
mark bond in this fashion over, say, the last year, will yield one of two qualitative results.
First, the benchmark bond will neither upgraded nor downgraded and the resulting
observed volatility will be (relatively) small. Second, the benchmark bond will have
realized
some credit quality migration and the resulting observed volatility will be (rela
tively) large.
This process of observing volatility should be unbiased over many trials. However, the
estimation error is potentially high due to the infrequent but meaningful impact of credit
quality migrations on value. Our approach in CreditMetrics uses long term estimate of
migration likelihood rather than observation within some recent sample period and so
should avoid this problem.
Consider
Chart 5.2
below. Bonds within each credit rating category can be said also to
have volatility of value due to daytoday credit spread ﬂuctuations. The RAROC
approach seeks to measure these ﬂuctuations, but will also sometimes
realize
a poten
tially large move due to a credit rating migration. Our approach is probabilistic. Credit
Metrics assumes that all migrations might have been realized and each is weighted by the
likelihoods of migration which we argue is best estimated using long term data.
2
The credit quality migration and revaluation mechanism in CreditMetrics gives a weight to remote but possible
credit quality migrations according to their longterm historical frequency without regard to how a shortterm (per
haps one year) sampling of bond prices would – or would not – have observed theses.
Sec. 5.3 A portfolio view 63
Part II: Model Parameters
Chart 5.2
Construction of volatility across credit quality categories
5.3 A portfolio view
Any analysis of a group of exposures could be called a portfolio analysis. We use the
term here to mean a Markowitztype analysis where the total risk of a portfolio is mea
sured by explicit consideration of the relationships between individual risks and expo
sure amounts in a variancecovariance framework. This type of analysis was originated
by Harry Markowitz, and has subsequently gone through considerable development, pri
marily in application to equity portfolios.
A growing number of major institutions estimate the portfolio effects of credit risk in a
Markowitztype framework. However, most institutions still rely on an intuitive assess
ment as to what level of over concentration to any one area may lead to problems. Thus,
bank lenders, for instance, typically set exposure limits against several types of portfolio
concentrations, such as industrial sector, geographical location, product type, etc. Lack
ing the guidance of a model, these groupings tend to be subjective rather than statistical.
For example, industrial sectors are generally deﬁned by aggregating fourdigit Standard
Industrial Classiﬁcation (SIC) codes into 60 or fewer groupings. This implies that the
banker is assuming that credit quality correlations are higher within an industry or sector
and lower between industries or sectors. It is not clear from the data that this is necessar
ily true. Although this is likely true for
commodity process
industries like oil reﬁning
and wood/paper manufacture, we believe it would be less true for
proprietary technology
industries like pharmaceuticals and computer software.
Modern portfolio theory is commonly applied to market risk. The volatilities and corre
lations necessary to calculate portfolio market volatility are generally readily
measurable. In contrast, there has been relatively little academic literature on the prob
lem of measuring diversiﬁcation or overconcentration within a credit portfolio. To do
this requires an understanding of credit quality correlations between obligors.
So, if we were interested in modeling the coincidence of
just
defaults, we might follow
Stevenson & Fadil [95]. They constructed 33 industry indices of default experience as
90 92.5 95 97.5 100 102.5 105 107.5 110
Relative frequency
Bond value
All potential migrations
BB
BBB AA
AAA
A (no sample period migration)
B, CCC
Default
64 Chapter 5. Overview of credit risk literature
CreditMetrics™—Technical Document
listed in Dun & Bradstreet’s
Business Failure Record
. The correlation between these
indices was their industry level estimate default correlation. While this approach is ﬁne
in concept, it suffers from the infrequency of defaults over which to correlate.
To get around this problem, another approach is to construct indices of, not just defaulted
ﬁrms, but default
likelihoods
of all ﬁrms. We know of two services which publish quan
titatively estimated default likelihood statistics across thousands of ﬁrms: KMV Corpo
ration and Zeta Services. Gollinger & Morgan [93] used time series of default
likelihoods (ZetaScores™ published by Zeta Services) to estimate default correlations
across 42 industry indices. Neither of these studies has been realized in a practicable
implementation.
In contrast to these academic suggestions, there is a practicable framework which is a
commercial offering by KMV Corporation. In brief, they estimate the value of a ﬁrm’s
debt within the option theoretic framework ﬁrst described in Merton [74]. Both expected
default frequencies (EDFs) and correlations of default expectation are addressed within a
consistent – and academically accepted – modeloftheﬁrm.
The approach practiced by KMV is to look to equity price series as a starting point to
understanding the volatility of a ﬁrm’s underlying (unlevered) asset value moves. Asset
value moves can be taken to be approximately normally distributed. These asset values
can in turn be mapped ordinally (onetoone) to credit quality measure, as illustrated in
Chart 3.3
. An assumption of bivariate normality between ﬁrms’ asset value moves then
allows credit quality correlations to be estimated from equity prices series. This is the
model on which we have constructed the equitybased correlation estimation in
Chapter 8
. J.P. Morgan has talked with KMV for at least four years on this approach to
correlation and we are grateful for their input.
65
Chapter 6. Default and credit quality migration
A fundamental source of risk is that the
credit quality
of an obligor may change over the
risk horizon. “Credit quality” is commonly used to refer to only the relative chance of
default. As we show here, however, CreditMetrics makes use of an extended deﬁnition
that includes also the volatility of up(down)grades. In this chapter we do the following:
• detail our modeloftheﬁrm which relates changes in underlying ﬁrm value to the
event of credit distress;
• generalize this model to incorporate up(down)grades in credit quality;
• discuss the historical tabulation of transition matrices by different providers;
• discuss anticipated longterm behavior of transition matrices; and
• detail an approach to estimate transition probabilities which is sensitive to both the
historical tabulation and anticipated longterm behavior.
6.1 Default
As discussed in the previous chapter, credit rating systems typically assign an alphabetic
or numeric label to rating categories. By itself, this only gives an ordinal ranking of the
default likelihoods across the categories. A quantitative framework, such as Credit
Metrics, must give meaning to each rating category by linking it with a default probabil
ity.
1
In the academic research, even the deﬁnition of the default event has evolved over time.
Up to 1989, it was common to look for only missed interest or principal payments (see
Altman [87]). Since then, starting with Asquith, Mullins & Wolff [89], researchers real
ized that distressed exchanges can play an important role in default statistics. Also
default rates can be materially different depending upon the population under study. If
rates are tabulated for the ﬁrst few years of newly issued, then the default rate will be
much lower than if the population broadly includes all extant debt.
6.1.1 Deﬁning credit distress
For our purposes in CreditMetrics, we look to the following characteristics when we
speak of the likelihood of credit distress:
• default rates which have been tabulated weighted by obligors rather than weighted
by number of issues or dollars of issuance;
• default rates which have been tabulated broadly upon all obligors rather than just
those with recent debt issuance; and
1
Rating agencies commonly also include a judgment for differing recovery rates in their subordinated and struc
tured debt rating. For instance, although senior and subordinated debt to a ﬁrm will encounter what we term
“credit distress” at the exact same time, the anticipated recovery rate for subordinated is lower and thus it is given
a lower rating. It is the senior rating that we look to as the most indicative of credit distress likelihood.
66 Chapter 6. Default and credit quality migration
CreditMetrics™—Technical Document
• default rates which are tabulated by senior rating categories (subordinated ratings
include recovery rate differences, which are separate from the
likelihood
of default).
This last point is worth elaborating. We utilize credit ratings as an indication of the
chance of default and credit rating migration likelihood. However, there are clearly dif
ferences in rating – to different debt of the same ﬁrm – between senior and subordinated
classes. The rating agencies assign lower ratings to subordinated debt in recognition of
differences in anticipated recovery rate in default. It is certainly true that senior debt
obligations may be satisﬁed in full during bankruptcy procedures while subordinated
debt is paid off only partially. In this circumstance we would say that the ﬁrm – and so
all
its debts – encountered
credit distress
even though only the subordinated class real
ized a
default
. Thus we take the senior credit rating as most indicative of the chance of a
ﬁrm encountering
credit distress
.
6.1.2 Fitting probabilities of default with a transition matrix
Based on historical default studies from both Moody’s and S&P credit rating systems,
we have transition matrices which include historically estimated oneyear default rates.
These are included as part of the dataset for CreditMetrics. Of course, there are many
rating agencies beyond S&P and Moody’s. There are two ways of using alternative
credit rating systems depending upon what historical information is available.
• If individual rating histories are available, then tabulating a transition matrix would
give ﬁrst direct estimate of the transition likelihoods including default.
• If all that is available are cumulative default histories by rating category,
2
then the
transition matrix which “best replicates” this history can be estimated.
In the absence of historical information, perhaps a onetoone correspondence could be
made to established rating systems based on each credit category’s rating criteria.
6.2 Credit quality migration
Credit rating migrations can be thought of as an extension of our model of ﬁrm defaults
discussed in
Section 3
and illustrated again in
Chart 6.1
. We say that a ﬁrm has some
underlying value – the value of its assets – and changes in this value suggest changes in
credit quality. Certainly it is the case that equity prices drop precipitously as a ﬁrm
moves towards bankruptcy. If we take the default likelihood as given by the credit rating
of the ﬁrm, then we can work backwards to the “threshold” in asset value that delimits
default. This is treated more formally in
Section 8.4
.
Likewise, just as our ﬁrm default model uses the default likelihood to place a threshold
below which a ﬁrm is deemed to be in default, so also do the rating migration probabili
ties deﬁne thresholds beyond which the ﬁrm would be deemed to up(down)grade from
its current credit rating. The data which drive this model are the default likelihood and
credit rating migration likelihoods for each credit rating. We can compactly represent
these rating migration probabilities using a transition matrix model (e.g.,
Table 6.2
).
2
Moody's terms these aggregated groupings “cohorts” and S&P terms them “static pools.”
Sec. 6.3 Historical tabulation 67
Part II: Model Parameters
In essence, a transition matrix is nothing more than a square table of probabilities. These
probabilities give the likelihood of migrating to any possible rating category (or perhaps
default) one period from now given the obligor’s credit rating today.
Chart 6.1
Model of ﬁrm value and migration
Many practical events (e.g., calls, enforced collateral provisions, spread resets) can be
triggered by a rating change. These actions can directly affect the realized value within
each credit rating category. For instance, a
pricing grid
– which predetermines a credit
spread schedule given changes in credit rating – can reduce the volatility of value across
up(down)grades.
3
Thus, we ﬁnd it very convenient to explicitly incorporate awareness
of rating migrations into our risk models.
6.3 Historical tabulation
We can tabulate historical credit rating migration probabilities by looking at time series
of credit ratings over many ﬁrms. This technique is both powerful and limited. It is
powerful in that we can freely model different volatilities of credit quality migration con
ditioned on the current credit standing. Said another way, each row in the transition
matrix describes a volatility of credit rating changes that is unique to that row’s initial
credit rating. This is clearly an advantage since migration volatilities can vary widely
between initial credit rating categories. There are, however, two assumption that we
make about transition matrices. They are:
1. We assume that all ﬁrms tagged with the “correct” rating label. By this we mean
that the rating agencies’ are diligent in consistently applying credit rating stan
dards across industries and countries (i.e., a “Baa” means the same for a U.S. elec
tric utility as it does for a French bank). Of course, there is no reason that
transition matrices could not be tabulated more speciﬁcally to reﬂect potential dif
ferences in the historical migration likelihoods of industries or countries. One
caveat to this reﬁnement might be the greater “noise” introduced by the smaller
sample sizes.
3
The securitised form of this structure is called a CreditSensitive Note (CSN) and is discussed in more detail in
Das & Tufano [96].
Default
CCC
B
BB
Firm remains
A
AA
AAA
Value of BBB firm at horizon date
BBB
Higher Lower
BBB
68 Chapter 6. Default and credit quality migration
CreditMetrics™—Technical Document
2. We assume that all ﬁrms tagged with a given rating label will act alike. By this we
mean that the full spectrum of credit migration likelihoods – not just the default
likelihood – is similar for each ﬁrm assigned to a particular credit rating.
There are several sources of transition matrices, each speciﬁc to a particular credit rating
service.
4
We advocate maintaining this correspondence even though it is common for
practitioners to use shorthand assumptions, e.g., Moody’s Baa is “just like” S&P’s BBB,
etc. Here we list three of these sources: Moody’s, S&P, and KMV. Each is shown for
the major credit rating categories – transition matrices which cover the minor (+/) credit
rating are also available, but are not shown here.
6.3.1 Moody’s Investors Service transition matrix
Moody’s utilizes a data set of 26 years’ worth of credit rating migrations over the issuers
that they cover. These issuers are predominantly U.S.based ﬁrms, but are including
more and more international ﬁrms. The transition matrix is tabulated upon issuers condi
tioned on those issuers continuing to be rated at the end of the year. Thus there is no
concern with having to adjust for a
nolongerrated
“rating.”
Table 6.1
Moody’s Investors Service: Oneyear transition matrix
Source: Lea Carty of Moody’s Investors Service
6.3.2 Standard & Poor’s transition matrix
It happens that the transition matrix published by Standard & Poor’s includes a
no
longerrated
“rating,” and so we pause to discuss this issue. The majority of these with
drawals of a rating occur when a ﬁrm’s only outstanding issue is paid off or its debt issu
ance program matures. Yet our assumption is that CreditMetrics will be applied to
obligations with a known maturity. So there should be no N.R. category in application.
Thus, it makes sense to eliminate the N.R. category and grossup the remaining percent
ages in some appropriate fashion. We do this as follows. Since S&P describes that they
track bankruptcies even after a rating is withdrawn, the default probabilities are already
fully tabulated. We believe that there is no systematic reason correlated with credit rat
4
KMV is not a credit rating service. They quantitatively estimate Expected Default Frequencies (EDF) which are
continuous values rather than categorical labels using an option theoretic approach.
Initial
Rating
Rating at yearend (%)
Aaa Aa A Baa Ba B Caa Default
Aaa 93.40 5.94 0.64 0 0.02 0 0 0
Aa 1.61 90.55 7.46 0.26 0.09 0.01 0 0.02
A 0.07 2.28 92.44 4.63 0.45 0.12 0.01 0
Baa 0.05 0.26 5.51 88.48 4.76 0.71 0.08 0.15
Ba 0.02 0.05 0.42 5.16 86.91 5.91 0.24 1.29
B 0 0.04 0.13 0.54 6.35 84.22 1.91 6.81
Caa 0 0 0 0.62 2.05 4.08 69.20 24.06
Sec. 6.3 Historical tabulation 69
Part II: Model Parameters
ing stating which would explain rating removals. We thus adjust all remaining migration
probabilities on a
pro rata
basis as shown in
Table 6.2
below:
Table 6.2
Standard & Poor’s oneyear transition matrix – adjusted for removal of N.R.
Source: Standard & Poor’s CreditWeek April 15, 1996
Both of these tables are included in the CreditMetrics data set.
6.3.3 KMV Corporation transition matrix
Both of the above transition matrices were tabulated by credit rating agencies. In con
trast, the sample transition matrix shown in Table 6.3 was constructed from KMV EDFs
(expected default frequency) for nonﬁnancial companies in the US using data from Jan
uary 1990 through September 1995. Each month, the rating group based on the EDF of
each company for that month was compared against the rating group it was in 12 months
hence, based on its EDF at that date. This gave a single migration. There are an average
of 4,780 companies in the sample each month, resulting in a total of 329,803 migration
observations. Firms that disappeared from the sample were allocated into the rating cat
egories proportionately to the population. Rating group #8 signiﬁes default, which is
treated as a a terminal event for the ﬁrm.
The purpose of this sample is to show how an alternative approach such as EDFs can be
utilized to generate a transition matrix. EDFs are default probabilities measured on a
continuous scale of 0.02% to 20.0%, but grouped into discrete “rating” ranges for appli
cation in CreditMetrics.
Initial
Rating
Rating at yearend (%)
AAA AA A BBB BB B CCC Default
AAA 90.81 8.33 0.68 0.06 0.12 0 0 0
AA 0.70 90.65 7.79 0.64 0.06 0.14 0.02 0
A 0.09 2.27 91.05 5.52 0.74 0.26 0.01 0.06
BBB 0.02 0.33 5.95 86.93 5.30 1.17 0.12 0.18
BB 0.03 0.14 0.67 7.73 80.53 8.84 1.00 1.06
B 0 0.11 0.24 0.43 6.48 83.46 4.07 5.20
CCC 0.22 0 0.22 1.30 2.38 11.24 64.86 19.79
70 Chapter 6. Default and credit quality migration
CreditMetrics™—Technical Document
Table 6.3
KMV oneyear transition matrices as tabulated from expected default frequencies (EDFs)
Source: KMV Corporation
Table 6.3
is presented as an example and will not be included in the CreditMetrics data
set. Subscribers to KMV’s Expected Default Frequencies utilize a measure of default
probability that is on a continues scale rather than discrete groupings offered by a credit
rating agency.
Both KMV and we ourselves advocate that each credit rating (or expected default fre
quency) be addressed by a transition matrix tailored to that system. For this reason, the
example KMV transition matrix shown here will not be part of the CreditMetrics data
set. Only subscribers to KMV’s expected default frequency (EDF) data would be users
of such a transition matrix and so KMV will be offering it as part of that subscription.
Although it would be ﬁne to have some issuers within a portfolio evaluated with one
service (i.e., ﬁnancials evaluated by IBCA) and other issuers evaluated by another ser
vice (i.e., corporates and industrials by Moody’s, say), it would be inappropriate to mix
systems (i.e., S&P ratings applied to Moody’s transition matrix).
6.4 Longterm behavior
In estimating transition matrices, there are a number of desirable properties that one
wants a transition to matrix to have, but which does not always follow from straightfor
ward compilation of the historical data. In general, it is good practice to impose at least
some of the desirable properties on the historical data in the form of estimation con
straints.
The nature and extent of the problems encountered will be a function of the particular
rating system, the number of grades considered, and the amount of historical data avail
able. The following discussion uses S&P ratings as the basis for explaining these issues
and how they can be addressed.
Historical tabulation is worthwhile in its own right. However, as with almost any type of
sampling, it represents a limited amount of observation with sampling error. In addition
to what we have historically observed, we also have strong expectations about credit rat
ing migrations. For instance, over sufﬁcient time we expect that any inconsistencies in
rank order across credit ratings will disappear. By
rank order
, we mean a consistent pro
gression in one direction such as default likelihoods always increasing – never then
Initial
Rating
Rating at Yearend (%)
1 (AAA) 2 (AA) 3 (A) 4 (BBB) 5 (BB) 6 (B) 7 (CCC) 8 (Default)
1 (AAA) 66.26 22.22 7.37 2.45 0.86 0.67 0.14 0.02
2 (AA) 21.66 43.04 25.83 6.56 1.99 0.68 0.20 0.04
3 (A) 2.76 20.34 44.19 22.94 7.42 1.97 0.28 0.10
4 (BBB) 0.30 2.80 22.63 42.54 23.52 6.95 1.00 0.26
5 (BB) 0.08 0.24 3.69 22.93 44.41 24.53 3.41 0.71
6 (B) 0.01 0.05 0.39 3.48 20.47 53.00 20.58 2.01
7 (CCC) 0.00 0.01 0.09 0.26 1.79 17.77 69.94 10.13
Sec. 6.4 Longterm behavior 71
Part II: Model Parameters
decreasing – as we move from high quality ratings to lower quality ratings. We list three
potential shortterm sampling error concerns here:
• Output cumulative default likelihoods should not violate proper rank order. For
instance,
Table 6.4
below shows that AAAs have defaulted more often at the 10year
horizon than have AAs.
• Limited historical observation yields “granularity” in estimates. For instance, the
AAA row in
Table 6.2
above is supported by 1,658 ﬁrmyears worth of observation.
This is enough to yield a “resolution” of 0.06% (i.e., only probabilities in increments
of 0.06% – or 1/1658 – are possible).
• This lack of resolution may erroneously suggest that some probabilities are identi
cally zero. For instance, if there were truly a 0.01% chance of AAA default, then we
would have to watch for another 80 years before there would be a 50% chance of
tabulating a nonzero AAA default probability.
There are other potential problems with historical sampling such as the business cycle
and regime shifts (e.g., the restructuring of the highyield market in the 1980’s). But
these will not be addressed here.
Table 6.4
Average cumulative default rates (%)
Source: S&P CreditWeek, Apr. 15, 1996
6.4.1 Replicate historical cumulative default rates
The major rating agencies have published tables of cumulative default likelihood over
holding periods as long as 20 years – reported in annual increments. If we ignore for the
moment the issue of autocorrelation, then it is generally true that
“there exists some
annual transition matrix which best replicates (in a least squares sense) this default his
tory.”
Said another way, we can always work backwards from a cumulative default table
to an implied transition matrix.
Table 6.4
illustrates part of a cumulative default proba
bility table published by Moody’s.
Term 1 2 3 4 5 ... 7 ... 10 ... 15
AAA 0.00 0.00 0.07 0.15 0.24 ... 0.66 ... 1.40 ... 1.40
AA 0.00 0.02 0.12 0.25 0.43 ... 0.89 ... 1.29 ... 1.48
A 0.06 0.16 0.27 0.44 0.67 ... 1.12 ... 2.17 ... 3.00
BBB 0.18 0.44 0.72 1.27 1.78 ... 2.99 ... 4.34 ... 4.70
BB 1.06 3.48 6.12 8.68 10.97 ... 14.46 ... 17.73 ... 19.91
B 5.20 11.00 15.95 19.40 21.88 ... 25.14 ... 29.02 ... 30.65
CCC 19.79 26.92 31.63 35.97 40.15 ... 42.64 ... 45.10 ... 45.10
72 Chapter 6. Default and credit quality migration
CreditMetrics™—Technical Document
Cumulative default rate tables like this can be ﬁt fairly closely by a single transition
matrix.
5
Thus, it is apparently true that defaults over time are closely approximated by a
transition matrix model.
6
This is an important result. It demonstrates that the statistical
behavior of credit rating migrations can be captured through a transition matrix model.
CreditMetrics uses a transition matrix to model credit rating migrations not only because
it is intuitive but also because it is an extremely powerful statistical tool.
Below we show a transition matrix that has been created using
nothing but
a least squares
ﬁt to the cumulative default rates in
Table 6.4
. At this point, we are most interested in
showing that: (i) such a matrix can be derived and (ii) that the process of defaults is
closely replicated by a Markov process. (We make no claim that
Table 6.5
is a faithful
replication of the historically tabulated
Table 6.2
.)
Table 6.5
Imputed transition matrix which best replicates default rates
For comparison to
Table 6.4
, we show below in
Table 6.6
the cumulative default rates
which result from this transition matrix. Again, the most important point is that
Table 6.4
and
Table 6.6
are quite close; thus the Markov process is a reasonable modeling tool.
The median difference between them is 0.16% with a maximum error of 2.13%.
This “best ﬁt” Markov process has yielded the side beneﬁt of resolving nonintuitive
rank order violations in its resulting cumulative default rates. For instance, our problem
of AAA’s having a 10 year default rate that was
greater
than AA’s is now gone. This
behavior – of noncrossing default likelihoods – is a feature that we would expect given
very long sampling histories.
5
Empirically, a transition matrix ﬁt is not as good for cumulative default rates of
newly issued
debt (as opposed to
the total debt population) due to a “seasoning” effect where subinvestment grades have an unusually low default
likelihood in the ﬁrst few years. This “seasoning” problem has not been apparent for bank facilities.
6
A transition matrix model is an example of a
Markov Process
. A Markov Process is a statespace model which
allows the next progression to be determined only by the current state and not information of previous states.
Initial
Rating
Rating at year end (%)
AAA AA A BBB BB B CCC Default
AAA 43.78 53.42 1.65 0.71 0.29 0.11 0.02 0.01
AA 0.60 90.60 6.20 1.45 0.93 0.16 0.04 0.01
A 0.22 2.84 92.97 3.12 0.56 0.14 0.07 0.07
BBB 2.67 3.29 12.77 75.30 5.07 0.60 0.14 0.17
BB 0.19 3.58 8.28 9.97 55.20 17.17 4.53 1.08
B 0.12 0.50 20.69 1.05 0.25 55.40 17.05 4.95
CCC 0.04 0.11 6.28 0.30 0.12 41.53 32.46 19.15
Sec. 6.4 Longterm behavior 73
Part II: Model Parameters
Table 6.6
Resulting cumulative default rates from imputed transition matrix (%)
6.4.2 Monotonicity (noncrossing) barrier likelihoods
Cumulative default rates are just a special case of what we term “barrier” likelihoods. In
general, we can ask, “what is the cumulative rate of crossing any given level of credit
quality?” For instance, if we managed a portfolio which was not allowed to invest in
subinvestment grade bonds, then we might be interested in the likelihood of any credit
quality migrations which were to or across the BB rating barrier. The cumulative proba
bilities for crossing the “BB barrier” using the transition matrix in
Table 6.5
are as shown
in
Table 6.7
. Notice that monotinicity (rank order) is violated for singleAs.
Table 6.7
“BB barrier” probabilities calculated from
Table 6.6
matrix (%)
Just as we would expect very longterm historical observation to resolve violations of
nonintuitive cumulative default rank order, we should expect resolution of barrier rank
ordering. This table above shows that our imputed transition matrix violates this antici
pated longterm behavior.
We can now replay the least squares ﬁt we performed when we produced
Table 6.4
with
the added constraint that all possible barrier probabilities must also be in rank order.
Table 6.8
shows these same BB barrier probabilities with our new ﬁt. (In fact, there are
six nondefault “barriers” for seven rating categories and our ﬁtting algorithm addressed
them all.)
Term 1 2 3 4 5 ... 7 ... 10 ... 15
AAA 0.01 0.04 0.09 0.18 0.31 ... 0.66 ... 1.37 ... 2.81
AA 0.01 0.06 0.15 0.27 0.44 ... 0.85 ... 1.63 ... 3.12
A 0.07 0.17 0.30 0.46 0.65 ... 1.11 ... 1.94 ... 3.50
BBB 0.17 0.41 0.78 1.25 1.79 ... 2.95 ... 4.60 ... 6.83
BB 1.08 3.41 6.14 8.76 11.05 ... 14.53 ... 17.71 ... 20.39
B 4.95 10.97 15.75 19.33 21.98 ... 25.46 ... 28.19 ... 30.35
CCC 19.15 27.43 32.63 36.32 39.01 ... 42.49 ... 45.14 ... 47.05
Term 1 2 3 4 5 ... 7 ... 10 ... 15
AAA
0.46 1.40 2.54 3.80 5.09 ... 7.74 ... 11.71 ... 18.13
AA
1.25 2.54 3.85 5.17 6.51 ... 9.17 ... 13.12 ... 19.47
A
0.91 2.00 3.20 4.49 5.82 ... 8.57 ... 12.69 ... 19.29
BBB
6.57 11.66 15.69 18.93 21.60 ... 25.78 ... 30.40 ... 36.25
74 Chapter 6. Default and credit quality migration
CreditMetrics™—Technical Document
Table 6.8
“BB barrier” probabilities calculated from
Table 6.6
matrix (%)
This reﬁnement was achieved with minimal change in the transition matrix’s ﬁt to the
cumulative default rates. The differences in predicted cumulative default rates averages
only 0.06% (median is 0.02%) between the two ﬁtted transition matrices. For compari
son with
Table 6.5
, we show this new ﬁt of our imputed transition matrix.
Table 6.9
Imputed transition matrix with default rate rank order constraint
Perhaps the difference between
Table 6.5
and
Table 6.9
is that the weight of probabilities
are generally moved towards the upperleft to lowerright diagonal. Also, without
directly trying, we are moving towards a better approximation of the historical transition
matrix shown in
Table 6.2
.
6.4.3 Steady state proﬁle matches debt market proﬁle
Another desirable dimension of “ﬁt” for a transition matrix is for it to exhibit a longterm
steady state that approximates the observed proﬁle of the overall credit markets. By this
we mean that – among those ﬁrms which do not default – there will be some distribution
of their credit quality across the available credit rating categories. To represent the rat
ing proﬁle across the bond market, we have taken the following data (
Table 6.10
) from
Standard & Poor’s
CreditWeek
April 15, 1996.
Term 1 2 3 4 5 ... 7 ... 10 ... 15
AAA
0.39 1.09 1.98 3.01 4.12 ... 6.52 ... 10.37 ... 16.97
AA
1.07 2.19 3.36 4.57 5.82 ... 8.39 ... 12.36 ... 19.01
A
1.13 2.42 3.82 5.29 6.80 ... 9.88 ... 14.48 ... 21.73
BBB
5.88 10.72 14.77 18.18 21.11 ... 25.89 ... 31.34 ... 38.13
Initial Rating
Rating at year end (%)
AAA AA A BBB BB B CCC Default
AAA
58.57 39.02 1.42 0.63 0.18 0.14 0.03 0.01
AA
0.71 89.45 7.47 1.39 0.72 0.18 0.05 0.02
A
0.25 3.83 91.15 3.73 0.77 0.14 0.07 0.06
BBB
2.07 2.26 10.03 80.29 4.53 0.50 0.15 0.18
BB
0.15 3.57 7.84 10.38 55.91 16.18 4.91 1.06
B
0.14 0.62 19.21 2.44 0.55 54.87 17.24 4.94
CCC
0.04 0.14 5.85 0.77 0.33 41.10 32.65 19.14
Sec. 6.4 Longterm behavior 75
Part II: Model Parameters
Table 6.10
Estimate of debt market proﬁle across credit rating categories
Mathematically, our transition matrix Markov process will have two longterm proper
ties (i.e., more than 100 periods). First, since default is an absorbing state, eventually all
ﬁrms will default. Second, since the initial state has geometrically less inﬂuence on
future states, the proﬁle of nondefaulted ﬁrms will converge to some steady state
regardless of the ﬁrm’s initial rating.
As the chart below shows, our ﬁtting algorithm can achieve a closer approximation of
the anticipated longterm steady state. The transition matrix in Table 6.9 shows too
strong a tendency to migrate towards singleA. Once we add an incentive to ﬁt the antic
ipated steady state, we see that a more balanced proﬁle is achieved.
Chart 6.2
Achieving a closer ﬁt to the longterm steady state proﬁle
This additional soft constraint was accomplished with a negligible effect on the matrix’s
ability to replicate cumulative default likelihoods – and monotonicity in the barrier was
still fully realized. Also, without directly trying, we are moving towards a better approx
imation of the historical transition matrix shown in Table 6.2.
6.4.4 Monotonicity (smoothly changing) transition likelihoods
Though it is certainly not a requirement of a transition matrix, our expectation is that
there is a certain rank ordering the likelihood of migrations as follows:
1. Better ratings should never have a higher chance of default;
2. The chance of migration should become less as the migration distance (in rating
notches) becomes greater; and
S&P 1996 AAA AA A BBB BB B CCC
Count 85 200 487 275 231 87 13
Proportion 6.2% 14.5% 35.3% 20.0% 16.8% 6.3% 0.9%
AAA AA A BBB BB BB CCC
0%
10%
20%
30%
40%
50%
60%
Credit rating
Table 6.10
Frequency
Section 6.4.3 result
Section 6.4.2 result
76 Chapter 6. Default and credit quality migration
CreditMetrics™—Technical Document
3. The chance of migrating to a given rating should be greater for more closely adja
cent rating categories.
As an example, we will refer to Table 6.10. Since the default likelihoods ascend
smoothly there is no violation of #1. However, since the chance that a singleB would
migrate to a singleA is greater than either a migration to BBB or BB, there is a “viola
tion” of #2. Also, since singleB has a greater chance of migrating to singleA than does
an initial BB or BBB, there is a “violation” of rule #3. The reader can ﬁnd other proba
bilities in this table which are not monotonic in our deﬁnition.
As before, we could add the soft constraint that our ﬁtting algorithm should endeavor to
mitigate these nonrank orderings of probabilities as it seeks to replicate the cumulative
default likelihoods. However, as we discuss next, there is one last source of data that we
should use in best estimating our transition matrix – an historically tabulated transition
matrix. Any ﬁtting algorithm that addresses smooth transition likelihoods would have to
revisit these same probabilities when it includes knowledge of the historically tabulated
transition matrix. So we address them both together below.
6.4.5 Match historically tabulated transition matrix
Standard & Poor’s historically tabulated transition matrix was shown above in Table 6.2.
Up to now we have discussed some of the characteristics of transition matrices and meth
ods of addressing these. Now we will bring all this together in Table 6.11 to give an esti
mate of a oneyear transition matrix which is rooted in the historical data and is also
sensitive to our expectation of longterm behavior.
Table 6.11
Achieving a closer ﬁt to the longterm steady state proﬁle
This transition matrix is meant to be close to the historically tabulated probabilities while
being adjusted somewhat to better approximate the longterm behaviors we have dis
cussed in this section. From a risk estimation standpoint we see that there are now small
but nonzero probabilities of default imputed for AAAs and AAs.
Initial
Rating
Rating at year end (%)
AAA AA A BBB BB B CCC Default
AAA 87.74 10.93 0.45 0.63 0.12 0.10 0.02 0.02
AA 0.84 88.23 7.47 2.16 1.11 0.13 0.05 0.02
A 0.27 1.59 89.05 7.40 1.48 0.13 0.06 0.03
BBB 1.84 1.89 5.00 84.21 6.51 0.32 0.16 0.07
BB 0.08 2.91 3.29 5.53 74.68 8.05 4.14 1.32
B 0.21 0.36 9.25 8.29 2.31 63.89 10.13 5.58
CCC 0.06 0.25 1.85 2.06 12.34 24.86 39.97 18.60
77
Chapter 7. Recovery rates
Residual value estimation in the event of default is inherently difﬁcult. At the time when
a banker makes a loan or an investor buys a bond, it is in the belief not that the obligor
will go bankrupt but that the instrument will outperform. So it can be especially difﬁcult
to imagine what the obligor’s position will be in the unlikely event of default. Will it be
a catastrophe which leaves no value to recover, or will it be a regrettable but well
behaved wrapping up which affects only shareholders but leaves debt holders whole?
It is in the remote chance of an outright default that a credit instrument will realize its
greatest potential loss. Across a typical portfolio, most of the credit risk will be attribut
able to default events. Investment grade credits will have relatively more of their volatil
ity attributable to credit spread moves versus subinvestment grade credits, which will be
primarily driven by potential default events. However, a typical portfolio will have a
mixture of each, with most of the portfolio risk coming from the subinvestment grades.
So the magnitude of any recovery rate in default is important to model diligently.
The academic literature in our bibliography focuses primarily on U.S. defaults post
October 1, 1979 – the effective date of the 1978 Bankruptcy Reform Act. However, the
general ﬁnding that recovery rates are highly uncertain with a distribution that can be
modeled is applicable internationally.
In this chapter, we will discuss not only the estimation of mean expected recovery rate in
default, but the important problem of addressing the wide uncertainty of recovery rate
experience. This chapter is organized as follows:
• estimating recovery rate distributions, their mean and standard deviation, by senior
ity level and exposure type; and
• ﬁtting a full distribution to recovery rate statistics while preserving the required 0%
to 100% bounds.
We have seen much effort devoted to estimating recovery rates based on: (i) seniority
ranking of debt, (ii) instrument type or use, (iii) credit rating
X
years before default, and
(iv) size and/or industry of the obligor. But the most striking feature of any historical
recovery data is its wide uncertainty. Any worthwhile credit risk model must be able to
incorporate recovery rate uncertainty in order to fully capture the volatility of value
attributable to credit. However, once we contemplate volatility of recoveries, we must
also address any potential correlation of recoveries across a portfolio. In this section, we
estimate any potential correlation of recoveries across the book.
7.1 Estimating recovery rates
There are many practical problems in estimating recovery rates of debt in the event of
default. Often there is no market from which to observe objective valuations, and if
there are market prices available they will necessarily be within a highly illiquid market.
Even if these issues are resolved there is the question of whether it is best to estimate val
ues: (i) immediately upon announcement of default, (ii) after some reasonable period for
information to become available – perhaps a month, or (iii) after a full settlement has
been reached – which can take years.
78 Chapter 7. Recovery rates
CreditMetrics™—Technical Document
Since there have been academic studies, see Eberhart & Sweeney [92], which conclude
that the bond market efﬁciently prices future realized liquidation values, we take comfort
in those studies which poll/estimate market valuations about one month after the
announcement of default. This certainly is the value which an active investor would face
whether or not he chose to hold his position after the default event.
1
We look to the following independent studies for use in CreditMetrics. These studies
reﬁne their estimates of recovery rate according to seniority type among bonds. Among
bank facilities (e.g., loans, commitments, letter of credit) the studies have viewed these
as a separate “seniority” class. It is clear from the data that the historical loan recovery
rates have been higher than recovery rates for senior bonds. It is not clear whether this is
attributable to differences in relationship, use of borrowing or security.
7.1.1 Recovery rates of bonds
For corporate bonds, we have two primary studies of recovery rate which arrive at simi
lar estimates (see Carty & Lieberman [96a] and Altman & Kishore [96]). For bond
recoveries we can look primarily to Moody’s 1996 study of recovery rates by seniority
class. This study has the largest sample of defaulted bond that we know of.
Table 7.1
is a
partial representation of Table 5 from Moody’s Investors Service, which shows statistics
for defaulted bond prices (1/1/70 through 12/31/1995).
Table 7.1
Recovery statistics by seniority class
Par (face value) is $100.00.
As this table shows, the subordinated classes are appreciably different from one another
in their recovery realizations. In contrast, the difference between secured versus unse
cured senior debt is not statistically signiﬁcant. It is likely that there is a selfselection
effect here. There is a greater chance for security to be requested in the cases where an
underlying ﬁrm has questionable hard assets from which to salvage value in the event of
default.
There is no public study we are aware of that seeks to isolate the effects of different lev
els of security controlling for the asset quality of the obligor ﬁrm. It becomes then a
1
There are two studies, see
Swank & Root [95] and Ward & Griepentrog [93], that report high average holding
period returns for debt held between the default announcement and the ultimate bankruptcy resolution. These
studies also note the high average uncertainty of returns and thus the market’s risk pricing efﬁciency.
Carty & Lieberman [96a] Altman & Kishore [96]
Seniority Class Number Average Std. Dev. Number Average Std. Dev.
Senior Secured 115 $53.80 $26.86 85 $57.89 $22.99
Senior Unsecured 278 $51.13 $25.45 221 $47.65 $26.71
Senior Subordinated 196 $38.52 $23.81 177 $34.38 $25.08
Subordinated 226 $32.74 $20.18 214 $31.34 $22.42
Junior Subordinated 9 $17.09 $10.90 — — —
Sec. 7.1 Estimating recovery rates 79
Part II: Model Parameters
practical problem for the risk manager to judge on a bondbybond basis what adjustment
is best made to recovery rate estimates for different levels of security.
2
7.1.2 Recovery rate of bank facilities
For bank facilities, we again have two primary studies of recovery rate which arrive at
similar estimates see (Asarnow & Edwards [95] and Carty & Lieberman [96b]). A&E
track 831 commercial and industrial loan defaults plus 89 structured loans while C&L
track 58 defaults of loans with Moody’s credit ratings. Both studies treat bank facilities
as essentially a seniority class of their own – with this being senior to all public bond
seniority classes.
Moody’s reports a 71% mean and 77% median recovery rate which is within sampling
error of Asarnow & Edwards 65.21% mean and 78.79% median recovery rates. So these
two studies are different by no more than 5%.
Chart 7.1
below is reproduced from A&E, and we have used it to estimate the standard
deviation of recovery rates, of 32.7%, which is beyond the information reported by A&E.
Chart 7.1
Distribution of bank facility recoveries
Source: Asarnow & Edwards [95]
A legitimate concern is that all of the studies referenced here are either exclusively based
on, or primarily driven by, U.S. bankruptcy experience. Since bankruptcy law and prac
tice differs from jurisdiction to jurisdiction (and even across time within a jurisdiction),
it is not clear that these historical estimates of recovery rate will be directly applicable
internationally.
2
For this reason, our software implementation of CreditMetrics, CreditManager,
will allow recovery rate estimates
to be overwritten on the individual exposure level
0

2
.
5
7
.
5

1
2
.
5
1
7
.
5

2
2
.
5
2
7
.
5

3
2
.
5
3
7
.
5

4
2
.
5
4
7
.
5

5
2
.
5
5
7
.
5

6
2
.
5
6
7
.
5

7
2
.
5
7
7
.
5

8
2
.
5
8
7
.
5

9
2
.
5
9
7
.
5

1
0
0
0
20
40
60
80
100
120
Frequency
Mean: 65.21%
Median: 78.79%
StDev: 32.70%
Recovery rate %
80 Chapter 7. Recovery rates
CreditMetrics™—Technical Document
7.2 Distribution of recovery rate
Recovery rates are best characterized, not by the predictability of their mean, but by their
consistently wide uncertainty. Loss rates are bounded between 0% and 100% of the
amount exposed. If we did not know anything about recovery rate, that is, if we thought
that all possible recovery rates were equally likely, then we would model them as a ﬂat
(i.e., uniform) distribution between the interval 0 to 1. Uniform distributions have a
mean of 0.5 and a standard deviation of 0.29 ( ). The standard deviations of
25.45% for senior unsecured bonds and 32.7% for bank facilities are on either side of
this and so represent relatively high uncertainties.
We can capture this wide uncertainty and the general shape of the recovery rate distribu
tion – while staying within the bounds of 0% to 100% – by utilizing a
beta distribution
.
Beta distributions are ﬂexible as to their shape and can be fully speciﬁed by stating the
desired mean and standard deviation.
Chart 7.2
illustrates beta distributions for different
seniority classes using some of statistics reported in
Table 7.1
.
Chart 7.2
Example beta distributions for seniority classes
.
This full representation of the distribution is unnecessary for the analytic engine of
CreditMetrics. It is used later in our simulation framework, where the shape of the full
distribution is required.
σ 1 12 ⁄ =
0% 25% 50% 75% 100%
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
Junior subordinated
Senior unsecured
Senior subordinated
Subordinated
Residual value
81
Chapter 8. Credit quality correlations
Central to our view of credit risk estimation is a diligent treatment of the portfolio effect
of credit. Whereas market risks can be diversiﬁed with a relatively small portfolio or
hedged using liquid instruments, credit risks are more problematic. For credit portfolios,
simply having many obligors’ names represented within a portfolio does not assure good
diversiﬁcation (i.e., they may all be large banks within one country). When diversiﬁca
tion is possible, it typically achieved by much larger numbers of exposures than for mar
ket portfolios.
The problem of constructing a Markowitztype portfolio aggregation of credit risk has
only recently been widely examined. We know of two academic papers which address
the problems of estimating correlations within a credit portfolio: Gollinger & Morgan
[93] use time series of default likelihoods (ZetaScores™ published by Zeta Services,
Inc.) to correlate across 42 constructed indices of industry default likelihoods, and
Stevenson & Fadil [95] correlate the default experience, as listed in Dun & Bradstreet’s
Business Failure Record
, across 33 industry groups. Both of these studies note the prac
tical difﬁculties of estimating default correlations.
Our portfolio treatment of credit risk was greatly inﬂuenced by various engagements
with KMV, which has studied models of credit correlations for a number of years.
The structure of this chapter is as follows:
• First, we discuss evidence from default histories which supports our assertion that
credit correlations actually exist.
• Second, we investigate the possibility of modeling joint rating changes directly using
historical rating change data.
• Third, we discuss the estimation of credit correlations through the observation of
bond spread histories.
• Fourth, we present a model which connects rating changes and defaults to move
ments in an obligor’s asset value. This allows us to model joint rating changes
across multiple obligors without relying on historical rating change or bond spread
data.
• Last, we discuss methods to estimate the parameters of the asset value model, and
present a dataset for this purpose.
8.1 Finding evidence of default correlation
In this section, before moving on to modeling correlations in credit rating changes, we
examine several histories of rating changes and defaults in order to establish that such
correlations in fact exist. One might claim that each ﬁrm is in many ways unique and its
changes in credit quality often are driven by events and circumstances speciﬁc to that
ﬁrm; this would argue for little correlation between different ﬁrms’ rating changes and
defaults. Thus, it would be desirable for us to ﬁrst ﬁnd evidence of defaults across a large
body of companies.
82 Chapter 8. Credit quality correlations
CreditMetrics™—Technical Document
We can do this by examining the default statistics reported by the major rating agencies
over many years. Since the studies we consider are based on a very large number of
observations, if defaults were uncorrelated, then we would expect to observe default
rates which are very stable from year to year. On the other hand, if defaults were per
fectly correlated, then we would observe some years where every ﬁrm in the study
defaults and others where no ﬁrms default. That our observations lie somewhere between
these two extremes (that is, we observe default rates which ﬂuctuate, but not to the extent
that they would under perfect correlation) is evidence that some correlation exists. We
make this observation more precise below.
We will use the formula below to compute average default correlation from the data;
for a full derivation, see
Appendix F
.
[8.1]
where the approximation is for large values of , the number of names covered by the
data. In the formula, denotes the average default rate over the years in the study and
is the standard deviation of the default rates observed from year to year.
Both Moody’s and S&P publish default rate statistics which could be used to make this
type of statistical inference of average default correlations. In
Table 8.1
, we use data
from Tables 3 and 6 from Moody’s most recent default study (see Carty & Lieberman
[96a]).
We can infer from these ﬁgures that the number of ﬁrmyears supporting the default rate,
, is in the thousands for all credit rating categories. Thus, our approximation formula
for is appropriate. However, there are only 25 yearly observations supporting the cal
culation of (and it is reported with signiﬁcant rounding), so the conﬁdence levels
around the resulting inferred correlation will be high.
Table 8.1
Inferred default correlations with conﬁdence levels
Source: Moody’s 19701995 1year default rates and volatilities (Carty & Lieberman [96a])
Credit
rating
category
Default
rate
Standard
deviation
defaults
Implied
default
correlation
Lower
conﬁdence
Upper
conﬁdence
Aa 0.03% 0.1% 0.33% 0.05% 1.45%
A 0.01% 0.1% 1.00% 0.15% 4.35%
baa 0.13% 0.3% 0.69% 0.29% 1.83%
ba 1.42% 1.4% 1.40% 0.79% 2.91%
B 7.62% 4.8% 3.27% 1.95% 6.47%
ρ
ρ
N
σ
2
µ µ
2
–

¸ ,
¸ _
1 –
N 1 –

σ
2
µ µ
2
–

≈ =
N
µ σ
µ
ρ
σ
µ σ ρ Pr ρ X < { } 2.5% = Pr ρ X > { } 2.5% =
Sec. 8.2 Direct estimation of joint credit moves 83
Part II: Model Parameters
There are at least four caveats to this approach:
• the standard deviations of default rates, , are calculated over a very limited number
of observations which lead to wide conﬁdence levels;
• the underlying periodic default rates for investment grade categories are not nor
mally distributed; thus the conﬁdence levels for the investment grades will be wider
than those calculated;
• the average default rate, , is assumed to be constant across all ﬁrms within the
credit rating category and constant across time; and
• the approach is sensitive to the proportion of recession versus growth years which –
in the 25year sample – may not be representative of the future.
The inferred default correlations are all positive and – using the conﬁdence interval tech
nique discussed above – are all statistically greater than zero to at least the 97.5% level.
This is a fairly objective indication that default events have statistically signiﬁcant corre
lations which cannot be ignored in a risk assessment model such as CreditMetrics.
In fact, our needs go beyond estimations of default correlations; we must estimate the
joint likelihood of any possible combination of credit quality outcomes. Thus, if the
credit rating system recognizes eight states (i.e.,
AAA, AA, …, CCC
plus
Default
), then
between two obligors there are 8•8 or 64 possible joint states whose likelihoods must to
be estimated.
8.2 Direct estimation of joint credit moves
Perhaps the most direct way to estimate joint rating change likelihoods is to examine
credit ratings time series across many ﬁrms which are synchronized in time with each
other. We have done this with a sample of 1,234 ﬁrms who have senior unsecured S&P
credit ratings reported quarterly for as much as the last 40 quarters. We note that this data
set does not include much of the default experience that S&P reports in their more com
prehensive studies and stress that we have assembled this data set only to illustrate the
principle that joint credit quality migration likelihoods can be estimated directly. With
this method, it is possible to avoid having to specify a correlation estimate and an accom
panying descriptive model.
Since we are interested in tabulating all possible pairwise combinations between ﬁrms,
there are over 1.13 million pairwise combinations within our particular sample. In gen
eral, if a rating series data set offers
N
observations in a tabulated transition matrix then it
will offer on the order of
N
2
observations of joint migration. For a rating system with
seven nondefault categories, there will be 28 unique joint likelihood tales. In
Table 8.2
we show one of these 28 tabulated results for the case where one ﬁrm starts the period as
a BBB and another ﬁrm starts the period with a singleA rating.
σ
µ
84 Chapter 8. Credit quality correlations
CreditMetrics™—Technical Document
Table 8.2
Historically tabulated joint credit quality comovements
This yields our nonparametric estimate of joint credit quality probabilities to be as in
Table 8.3
:
Table 8.3
Historically tabulated joint credit quality comovement (%)
We emphasize again that this illustration is only to demonstrate a technique for estimat
ing joint credit quality migration likelihoods directly. Unfortunately, our own access to
the rating agency’s data sets is inadequate to fully estimate a production quality study.
This method of estimation has the advantage that it does not make assumptions as to the
underlying process, the joint distribution shape, or rely on distilling the data down to a
single parameter – the correlation. However, it carries the limitation of treating all ﬁrms
with a given credit rating as identical. So two banks would be deemed to have the same
relationship as a bank and an oil reﬁner. In the following sections, we discuss a method
of estimating credit quality correlations which are sensitive to the characteristics of indi
vidual ﬁrms.
8.3 Estimating credit quality correlations through bond spreads
A second way to estimate credit quality correlations using historical data would be to
examine price histories of corporate bonds. Because it is intuitive that movements in
bond prices reﬂect changes in credit quality, it is reasonable to believe that correlations
Firm
starting
in BBB
Firm starting in A
AAA AA A BBB BB B CCC Default
AAA 0 0 0 0 0 0 0 0
AA 0 15 1,105 54 4 0 0 0
A 0 978 44,523 2,812 414 224 0 0
BBB 0 12,436 621,477 40,584 5,075 2,507 0 0
BB 0 839 41,760 2,921 321 193 0 0
B 0 175 7,081 532 76 48 0 0
CCC 0 55 2,230 127 18 15 0 0
Default 0 29 981 67 7 0 0 0
Firm
starting
in BBB
Firm starting in A
AAA AA A BBB BB B CCC Default
AAA        
AA  0.00 0.14 0.01 0.00   
A  0.12 5.64 0.36 0.05 0.03  
BBB  1.57 78.70 5.14 0.64 0.32  
BB  0.11 5.29 0.37 0.04 0.02  
B  0.02 0.90 0.07 0.01 0.01  
CCC  0.01 0.28 0.02 0.00 0.00  
Default  0.00 0.12 0.01 0.00   
Sec. 8.4 Asset value model 85
Part II: Model Parameters
of bond price moves might allow for estimations of correlations of credit quality moves.
Such an approach has two requirements: adequate data on bond price histories and a
model relating bond prices to credit events.
Where bond price histories are available, it is possible to estimate some type of credit
correlation by ﬁrst extracting credit spreads from the bond prices, and then estimating
the correlation in the movements of these spreads. It is important to note that such a cor
relation only describes how spreads tend to move together. To arrive at the parameters
we require for CreditMetrics (that is, likelihoods of joint credit quality movements), it is
necessary to adopt a model which links spread movements to credit events.
Models of risky bonds typically have three state variables: the ﬁrst is the risk free interest
rate, the second is the credit spread, and the third indicates whether the bond has
defaulted. A typical approach (see for example Duffee [95] or Nielsen and Ronn [94]) is
to assume that the risk free rate and credit spread evolve independently
1
and that defaults
are linked to the credit spread through some pricing model. This pricing model allows us
to infer the probability of the issuer defaulting from the observed bond spread
2
. An
extension of this type of model to two or more bonds would allow for the inference of
default correlations from the correlation in bond spread moves.
While an approach of this type is attractive because it is elegant and consistent with other
models of risky assets, its biggest drawback is practical. Bond spread data is notoriously
scarce, particularly for low credit quality issues, making the estimation of bond spread
correlations impossible in practice.
8.4 Asset value model
In this section, we present the approach which we introduced in
Chapter 3
and which we
will use in practice to model joint probabilities of upgrades, downgrades, and defaults
(all of which will be referred to generically as credit rating changes). We are motivated
to pursue such an approach by the fact that practical matters (such as the lack of data on
joint defaults) make it difﬁcult to estimate such probabilities directly. Our approach here
then will be indirect. It involves two steps:
1. Propose an underlying process which drives credit rating changes. This will estab
lish a connection between the events which we ultimately want to describe (rating
changes), but which are not readily observable, and a process which we understand
and can observe.
2. Estimate the parameters for the process above. If we have been successful in the
ﬁrst part, this should be easier than estimating the joint rating change probabilities
directly.
In this section, we propose that a ﬁrm’s asset value be the process which drives its credit
rating changes and defaults. This model is essentially the option theoretic model of Mer
ton [74], which is discussed further in Kealhofer [95]. We describe the model which
links changes in asset values to credit rating changes and explain how we parameterize
1
The evolution of these quantities is generally modeled by diffusion processes with some drift and volatility.
2
This is similar to the inference of implied volatilities from observed option premiums.
86 Chapter 8. Credit quality correlations
CreditMetrics™—Technical Document
the asset value model. We postpone the discussion of parameter estimation to
Section 8.5
.
It is evident that the value of a company’s assets determines its ability to pay its debt
holders. We may suppose then that there is a speciﬁc level such that if the company’s
assets fall below this level in the next year, it will be unable to meet its payment obliga
tions and will default. Were we only treating value changes due to default, this would be
a sufﬁcient model. However, since we wish to treat portfolio value changes resulting
from changes in credit rating as well, we need a slightly more complex framework.
Extending the intuition above, we assume there is a series of levels for asset value that
will determine a company’s credit rating at the end of the period in question. For exam
ple, consider a hypothetical company that is BB rated and whose assets are currently
worth $100 million. Then the assumption is that there are asset levels such that we can
construct a mapping from asset value in one year’s time to rating in one year’s time, as in
Chart 8.1
. Essentially, the assumption is that the asset value in one year determines the
credit rating (or default) of the company at that time. The asset values in the chart which
correspond to changes in rating will be referred to as asset value thresholds. We reiterate
that we are not yet claiming to know what these thresholds are, only that this relationship
exists.
Chart 8.1
Credit rating migration driven by underlying BB ﬁrm asset value
Assuming we know the asset thresholds for a company, we only need to model the com
pany’s change in asset value in order to describe its credit rating evolution. To do this,
we assert that the percent changes in asset value (that is, asset “returns,” which we will
denote by
R
) are normally distributed, and parameterized by a mean
µ
and standard devi
ation (or volatility)
σ
. Note that this volatility is not the volatility of value of a credit
• •
• •
• •
• •
• •
• •
• •
• •
0 20 40 60 80 100 120 140 160 180 200
Rating
Asset value in one year
Default
CCC
B
BB
BBB
A
AA
AAA
Sec. 8.4 Asset value model 87
Part II: Model Parameters
instrument (which is an output of CreditMetrics) but simply the volatility of asset returns
for a given name. For ease of exposition, we will assume
µ
=
0
3
.
Given this parameterization of the asset value process, we may now establish a connec
tion between the asset thresholds in the chart above and the transition probabilities for
our company. Continuing with our example of the BB rated obligor, we read from the
transition matrix that the obligor’s oneyear transition probabilities are as in the second
column of
Table 8.4
.
On the other hand, from the discussion of asset thresholds above, we know that there
exist asset return thresholds
Z
Def
, Z
CCC
, Z
BBB
,
etc., such that if
R< Z
Def
, then the obligor
goes into default; if
Z
Def
<R<Z
CCC
, then the obligor is downgraded to CCC; and so on.
So for example, if
Z
Def
were equal to 70%, this would mean that a 70% (or greater)
decrease in the asset value of the obligor would lead to the obligor’s default.
Since we have assumed that
R
is normally distributed, we can compute the probability
that each of these events occur:
[8.2]
and so on. (
Φ
denotes the cumulative distribution for the standard normal distribution.)
These probabilities are listed in the third column of
Table 8.4
.
Table 8.4.
One year transition probabilities for a BB rated obligor
The connection between asset returns and credit rating may be represented schematically
as in
Chart 8.2
, where we present the return thresholds superimposed on the distribution
of asset returns. The integral between adjacent thresholds corresponds to the probability
that the obligor assumes the credit rating corresponding to this region.
3
This likely will not be the case in practice, but for our purposes here, the value of will not inﬂuence the result.
It is in fact true that
σ
does not inﬂuence the ﬁnal result either – and the reader may choose to ignore
σ
in the
expressions to follow – but we retain it for illustrative purposes.
Rating
Probability from the
transition matrix (%)(
Probability according to the
asset value model
AAA 0.03
1−Φ
(Z
AA
/σ)
AA 0.14
Φ(
Z
AA
/σ)−Φ(
Z
A
/σ)
A 0.67
Φ(
Z
A
/σ)−Φ(
Z
BBB
/σ)
BBB 7.73
Φ(
Z
BBB
/σ)−Φ(
Z
BB
/σ)
BB 80.53
Φ(
Z
BB
/σ)−Φ(
Z
B
/σ)
B 8.84
Φ(
Z
B
/σ)−Φ(
Z
XXX
/σ)
CCC 1.00
Φ(
Z
XXX
/σ)
−
Φ
(
Z
Def
/
σ
)
Default 1.06
Φ
(
Z
De
f
/
σ
)
µ
Pr Default { } Pr R Z
Def
< { } Φ Z
Def
σ ⁄ ( ) , = =
Pr CCC { } Pr Z
De f < ( )
R Z
CCC
< { } Φ Z
CCC
σ ⁄ ( ) Φ Z
Def
σ ⁄ ( ) , – = =
88 Chapter 8. Credit quality correlations
CreditMetrics™—Technical Document
Chart 8.2
Distribution of asset returns with rating change thresholds
Now in order to complete the connection, we simply observe that the probabilities in the
two columns of the Table 1 must be equal. So considering the default probability, we see
that must equal 1.06%, which lets us solve for Z
Def
:
[8.3] ,
where gives the level below which a standard normal distributed random variable
falls with probability . Using this value, we may consider the CCC probability to solve
for Z
CCC
, then the B probability to solve for Z
B
, and so on, obtaining the values in
Table 8.5. Note there is no threshold Z
AAA
, since any return over 3.43σ implies an
upgrade to AAA.
4
Table 8.5
Threshold values for asset return
for a BBB rated obligor
Now consider a second obligor, A rated. Denote this obligor’s asset return by R′, the
standard deviation of asset returns for this obligor by σ′, and its asset return thresholds
4
We comment that to this point, we have not added anything to our model. For one obligor, we only need the tran
sition probabilities to describe the evolution of credit rating changes, and the asset value process is not necessary.
The beneﬁt of the asset value process is only in the consideration of multiple obligors.
Threshold Value
Z
AA 3.43σ
Z
A 2.93σ
Z
BBB 2.39σ
Z
BB 1.37σ
Z
B −1.23σ
Z
CCC −2.04σ
Z
Def −2.30σ
Asset return over one year
Downgrade to B
Firm defaults
Upgrade to BBB
Z
CCC
Z
B
Z
BBB
Z
A
Firm remains
BB rated
Z
AA
Z
AAA
Φ Z
Def
σ ⁄ ( )
Z
Def
Φ
1 –
1.06% ( ) σ ⋅ 2.30σ – = =
Φ
1 –
p ( )
p
Sec. 8.4 Asset value model 89
Part II: Model Parameters
by Z′
Def
, Z′
CCC
, and so on. The transition probabilities and asset return thresholds are
listed in Table 8.6.
Table 8.6
Transition probabilities and asset return thresholds for A rating
At this point, we have described the motion of each obligor individually according to its
asset value processes. To describe the evolution of the two credit ratings jointly, we
assume that the two asset returns are correlated and normally distributed,
5
and it only
remains to specify the correlation ρ between the two asset returns. We then have the
covariance matrix for the bivariate normal distribution:
[8.4]
This done, we know how the asset values of the two obligors move together, and can
then use the thresholds to see how the two credit ratings move together.
To be speciﬁc, say we wish to compute the probability that both obligors remain in their
current credit rating. This is the probability that the asset return for the BB rated obligor
falls between Z
B
and Z
BB
while at the same time the asset return for the A rated obligor
falls between Z′
BBB
and Z′
A
. If the two asset returns are independent (i.e., ρ=0), then this
joint probability is just the product of 80.53% (the probability that the BB rated obligor
remains BB rated) and 91.05% (the probability that the A rated obligor remains A rated).
If ρ is not zero, then we compute:
[8.5]
where f(r,r′;Σ) is the density function for the bivariate normal distribution with covari
ance matrix Σ
6
. We may use the same procedure to calculate the probabilities of each of
5
Technically, we assume that the two asset returns are bivariate normally distributed. We remark, however, that it is
not necessary to use the normal distribution. Any multivariate distribution (including those incorporating fat tails
or skewness effects) where the joint movements of asset values can be characterized fully by one correlation
parameter would be applicable.
6
The variables r and r′ in Eq. [8.5] represent the values that the two asset returns may take on within the speciﬁed
intervals.
Rating Probability Threshold Value
AAA 0.09%
AA 2.27% Z’
AA
3.12σ′
A 91.05% Z’
A
1.98σ′
BBB 5.52% Z’
BBB
−1.51σ′
BB 0.74% Z’
BB
−2.30σ′
B 0.26% Z’
B
−2.72σ′
CCC 0.01% Z’
CCC
−3.19σ′
Default 0.06% Z’
Def
−3.24σ′
Σ
σ
2
ρσσ'
ρσσ' σ'
2
¸ ,
¸ _
=
Pr Z
B
R Z
BB
Z'
BBB
R' Z'
A
< < , < < { } f r r′ Σ ; , ( ) r′ d ( ) r d
Z′
BBB
Z′
A
∫
Z
B
Z
BB
∫
=
90 Chapter 8. Credit quality correlations
CreditMetrics™—Technical Document
the 64 possible joint rating moves for the two obligors. As an example, suppose that
ρ=20%. We would then obtain the probabilities in Table 8.7.
Table 8.7
Joint rating change probabilities for BB and A rated obligors (%)
This table is sufﬁcient to compute the standard deviation of value change for a portfolio
containing only issues of these two obligors. Note that the totals for each obligor are just
that obligor’s transition probabilities. To compute the standard deviation for a larger
portfolio, it is only necessary to repeat this analysis for each pair of obligors in the port
folio.
7
The effect of the correlation merits further comment. Consider the worst case event for a
portfolio containing these two obligors – that both obligors default. If the asset returns
are independent, then the joint default probability is the product of the individual default
probabilities, or 0.0006%. On the other hand, if the asset returns are perfectly correlated
(ρ=1), then any time the A rated obligor defaults, so too does the BB rated obligor.
Thus, the probability that they both default is just the probability that the A rated obligor
defaults, or 0.06%, 100 times greater than in the uncorrelated case.
In Chart 8.3, we illustrate the effect of asset return correlation on the joint default proba
bility for our two obligors.
7
Note that if all pairs of obligors have the same correlation, then the maximum number of matrices like Table 8.7
which would be needed is 28, regardless of the size of the portfolio. Notice that Table 8.7 depends only on the
ratings of the two obligors and on the correlation between them, and not on the particular obligors themselves.
Thus, since there are only seven possible ratings for each obligor, there are only 28 possibilities for the ratings of
each pair of obligors, and 28 possible matrices.
Rating of
ﬁrst company
Rating of second company
AAA AA A BBB BB B CCC Def Total
AAA 0.00 0.00 0.03 0.00 0.00 0.00 0.00 0.00 0.03
AA 0.00 0.01 0.13 0.00 0.00 0.00 0.00 0.00 0.14
A 0.00 0.04 0.61 0.01 0.00 0.00 0.00 0.00 0.67
BBB 0.02 0.35 7.10 0.20 0.02 0.01 0.00 0.00 7.69
BB 0.07 1.79 73.65 4.24 0.56 0.18 0.01 0.04 80.53
B 0.00 0.08 7.80 0.79 0.13 0.05 0.00 0.01 8.87
CCC 0.00 0.01 0.85 0.11 0.02 0.01 0.00 0.00 1.00
Def 0.00 0.01 0.90 0.13 0.02 0.01 0.00 0.00 1.07
Total 0.09 2.29 91.06 5.48 0.75 0.26 0.01 0.06 100
Sec. 8.4 Asset value model 91
Part II: Model Parameters
Chart 8.3
Probability of joint defaults as a function of asset return correlation
We have pointed out before that for pairs of obligors, it is only necessary to specify joint
probabilities of rating changes and defaults, and that actual default correlations are not
used in any calculations. However, many people are accustomed to thinking in terms of
default correlations, and so we touch brieﬂy on them here. For an asset correlation ,
we have shown that it is possible to compute , the probability that obligors 1 and 2
both default. The default correlation between these two obligors can then be written as
[8.6] ,
where and are the probabilities that obligor 1 and obligor 2 default, respectively.
The translation from asset to default correlation lowers the correlation signiﬁcantly.
Asset correlations in the range of 40% to 60% will typically translate into default corre
lations of 2% to 4%. We see then that even the very small default correlation estimates
in Section 8.1 require that asset value moves exhibit relatively high correlations.
Chart 8.4 shows how the default correlation is a function of the two obligor's default
probabilities. An asset correlation of 30% was assumed and default probabilities range
from 1bp to above 10%. The high “mound” towards the back indicates that junk bond
defaults will be far more correlated with each other than will investment grade defaults.
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
0.00%
0.01%
0.02%
0.03%
0.04%
0.05%
0.06%
Correlation
Joint default probability
ρ
A
p
12
ρ
D
p
12
p
1
p
2
–
p
1
1 p
1
– ( ) p
2
1 p
2
– ( )

=
p
1
p
2
92 Chapter 8. Credit quality correlations
CreditMetrics™—Technical Document
Chart 8.4
Translation of equity correlation to default correlation
Before moving on to estimation of parameters, we make one important observation:
Equation [8.5] above does not depend on either of the volatilities σ or σ'. This may seem
counterintuitive, that in a risk model we are ignoring asset volatility, but essentially all of
the volatility we need to model is captured by the transition probabilities for each obli
gor. As an example, consider two obligors which have the same rating (and therefore the
same transition probabilities), but where the asset volatility for one obligor is ten times
greater than the other. We know that the credit risk is the same to either obligor. One
obligor does have a more volatile asset process, but this just means that its asset return
thresholds are greater than those of the other ﬁrm. In the end, the only parameters which
affect the risk of the portfolio are the transition probabilities for each obligor and the cor
relations between asset returns.
The consequence of this is that we may consider standardized asset returns, that is, asset
returns adjusted to have mean zero and standard deviation one. The only parameter to
estimate then is the correlation between asset returns, which is the focus of the next
section.
One last comment is that it is a simple matter to adjust for different time horizons. For
example, to perform this analysis for a sixmonth time horizon, the only change is that
we use the sixmonth transition probabilities to calibrate the asset return thresholds.
8.5 Estimating asset correlations
The user can pursue different alternatives to estimate ﬁrm asset correlations. The sim
plest is just to use some ﬁxed value across all obligor pairs in the portfolio. This pre
cludes the user having to estimate a large number (4,950 for a 100obligor portfolio) of
individual correlations, while still providing reasonable portfolio risk measures. How
ever, the ability to detail risk due to overconcentration in a particular industry, for exam
10.00%
1.00%
0.10%
0.01% 0.01%
0.10%
1.00%
10.00%
0.200
0.175
0.150
0.125
0.100
0.075
0.050
0.025
0.000
0.200
0.175
0.150
0.125
0.100
0.075
0.050
0.025
0
Default likelihood
Obligor #1
Default likelihood
Obligor #2
Default correlation
Default correlation
Sec. 8.5 Estimating asset correlations 93
Part II: Model Parameters
ple, is lost. A typical average asset correlation across a portfolio may be in the range of
20% to 35%.
8
For more speciﬁc correlations, there are independent data providers that can provide
models which are independent of – but can be consistently used in – CreditMetrics.
Below, we present our own interpretation of this type of underlying ﬁrm asset correlation
estimation.
One fundamental – and typically very observable – source of ﬁrmspeciﬁc correlation
information are equity returns. Here, we use the correlation between equity returns as a
proxy for the correlation of asset returns. While this method has the drawback of over
looking the differences between equity and asset correlations, it is more accurate than
using a ﬁxed correlation, and is based on much more readily available data than credit
spreads or actual joint rating changes.
In the best of all possible worlds, we could produce correlations for any pair of obligors
which a user might request. However, the scarcity of data for many obligors, as well as
the impossibility of storing a correlation matrix of the size that would be necessary, make
this approach untenable. Therefore, we resort to a methodology which relies on correla
tions within a set of indices and a mapping scheme to build the obligorbyobligor corre
lations from the index correlations.
Thus, to produce individual obligor correlations, there are two steps:
• First, we utilize industry indices in particular countries to construct a matrix of cor
relations between these industries. The result is that we obtain the correlation, for
example, of the German chemical industry with the United States insurance industry.
For reasons which will become clear below, we also report the volatility for each of
these indices.
9
• Next, we map individual obligors by industry participation. For example, a com
pany might be mapped as 80% Germany and 20% United States, and 70% chemicals
and 30% ﬁnance, resulting in 56% participation in the German chemicals industry,
24% in German ﬁnance, 14% in American chemicals, and 6% in American ﬁnance.
Using these weights and the countryindustry correlations from above, we obtain the
correlations between obligors.
In Section 8.5.1, we discuss the data we provide and the methodology which goes into its
construction. In the following subsection, we present an example to describe the meth
ods by which the user speciﬁes the weightings for individual obligors and arrives at indi
vidual obligor correlations. The last subsection is a generalization of this example.
8
Based on conversations with Patrick H. McAllister in 1994 when he was an Economist at the Board of Governors
of the Federal Reserve System. Part of his research inferred average asset correlations of corporate & industrial
loan portfolios within midsized US banks to be in the range 20%to25%. Our own research suggests that it is
easier to construct higher correlation portfolios versus lower correlation portfolios, hence a 20%to35% range.
9
Recall from Section 8.4 that volatilities do not ﬁgure into the model for joint rating changes. We will see that the
volatilities of the indices are necessary, however, for mapping individual obligors to the indices.
94 Chapter 8. Credit quality correlations
CreditMetrics™—Technical Document
8.5.1 Data
As mentioned above, we provide the user a matrix of correlations between industries in
various countries. In this section, we discuss the data and the methods by which we con
struct this matrix.
In Table 8.8, we list the countries for which we provide data, along with the family of
industry speciﬁc indices we use for each country. For each country, the broad country
index used is the MSCI index. For countries where no index family appears, insufﬁcient
industry index data was available and we utilize only the data for the broad country index.
Table 8.8
Countries and respective index families
In Table 8.9, we list the industries for which we provide indices in one or more of the
countries. We choose these industry groups by beginning with the major groups used by
Standard & Poor for the United States, and then eliminating groups which appear redun
dant. For instance, we ﬁnd that the correlation between the Health Care and
Pharmaceuticals indices is over 98%, and so consolidate these two groups into one, rea
soning that the two indices essentially explain the same movements in the market.
Country Index family Country Index family
Australia ASX Mexico Mexican SE
Austria New Zealand
Belgium Norway Oslo SE
Canada Toronto SE Philippines Philippine SE
Finland Helsinki SE Poland
France SBF Portugal
Germany CDAX Singapore AllSingapore
Greece Athens SE South Africa
Hong Kong Hang Seng Spain
Indonesia Sweden Stockholm SE
Italy Milan SE Switzerland SPI
Japan Topix Thailand SET
Korea Korea SE United Kingdom FTSEA
Malaysia KLSE United States S&P
Sec. 8.5 Estimating asset correlations 95
Part II: Model Parameters
Table 8.9
Industry groupings with codes
Because the industry coverage in each country is not uniform, we also provide data on
MSCI worldwide industry indices. In a case such French chemicals, where there is no
countryindustry index, the user may then choose to proxy the French chemical index
with a combination of the MSCI France index and the MSCI worldwide chemicals index.
Finally, realizing that it may at times be more feasible to describe a company by a
regional index rather than a set of country indices, we provide data on six MSCI regional
indices. In the end, we select the indices for which at least three years of data are avail
able, leaving us with 152 countryindustry indices, 28 country indices, 19 worldwide
industry indices, and 6 regional indices. The available countryindustry pairs are pre
sented in
Table 8.10
. For the speciﬁc index titles used in each case, refer to
Appendix I
.
Grouping Code Grouping Code
General country index GNRL Insurance INSU
Automobiles AUTO Machinery MACH
Banking & ﬁnance BFIN Manufacturing MANU
Broadcasting & media BMED Metals Mining MMIN
Chemicals CHEM Oil & gas – reﬁning & marketing OGAS
Construction & building materials CSTR Paper & forest products PAPR
Electronics ELCS Publishing PUBL
Energy ENRG Technology TECH
Entertainment ENMT Telecommunications TCOM
Food FOOD Textiles TXTL
Health care & pharmaceuticals HCAR Transportation TRAN
Hotels HOTE Utilities UTIL
96 Chapter 8. Credit quality correlations
CreditMetrics™—Technical Document
Table 8.10
Countryindustry index availability
Country
G
N
R
L
A
U
T
O
B
F
I
N
B
M
E
D
C
H
E
M
C
S
T
R
E
L
C
S
E
N
R
G
E
N
M
T
F
O
O
D
H
C
A
R
H
O
T
E
I
N
S
U
M
A
C
H
M
A
N
U
M
M
I
N
O
G
A
S
P
A
P
R
P
U
B
L
T
E
C
H
T
C
O
M
T
X
T
L
T
R
A
N
U
T
I
L
T
o
t
a
l
Australia X X X X X X X X X X 10
Austria X 1
Belgium X 1
Canada X X X X X X X X X X X X X X X 15
Finland X X X X X 5
France X X X X X X 6
Germany X X X X X X X X X X X 11
Greece X X X 3
Hong Kong X X X 3
Indonesia X 1
Italy X X X X X X 6
Japan X X X X X X X X X X X X X X X X 16
Korea X X X X X X X X X X X 11
Malaysia X X X 3
Mexico X X X X 4
New Zealand X 1
Norway X X X 3
Philippines X X X 3
Poland X 1
Portugal X 1
Singapore X X X 3
South Africa X X X 3
Spain X 1
Sweden X X X X X 6
Switzerland X X X X X 5
Thailand X X X X X X X X X X X X X X X X X 17
United Kingdom X X X X X X X X X X X X X X X X X 17
United States X X X X X X X X X X X X X X X X X X X X X X X X 24
MSCI worldwide X X X X X X X X X X X X X X X X X X X 19
Total 28 5 20 6 12 13 7 8 2 10 6 6 12 6 1 13 4 11 3 2 3 7 10 4 199
Sec. 8.5 Estimating asset correlations 97
Part II: Model Parameters
For each of the indices, we consider the last 190 weekly returns, and compute the mean
and standard deviation of each return series. Thus, if we denote the t
th
week’s return on
the k
th
index by , we compute the average weekly return on this index by
[8.7] ,
where T is 190 in our case, and the weekly standard deviation of return by
[8.8] .
As mentioned above, we provide the user with the standard deviations (volatilities), and
discuss their use in the next section. In addition, for all pairs of indices, we compute the
covariance of weekly returns by
[8.9] ,
and the correlation of weekly returns by
[8.10] .
We provide these correlations to the user.
Note that our computations of volatilities and correlations differ from the standard vola
tility computations in RiskMetrics in that we weight all of the returns in each time series
equally. The motivation for this is that we are interested in computing correlations
which are valid over the longer horizons for which CreditMetrics will be used. The sta
tistics here tend to be more stable over time, and reﬂect longer term trends, whereas the
statistics in RiskMetrics vary more from day to day, and capture shorter term behavior.
Note also that the correlations we compute are based on historical weekly returns. It is
therefore an assumption of the model that the weekly correlations which we provide are
accurate reﬂections of the quarterly or yearly asset moves which drive the CreditMetrics
model.
8.5.2 Obligor correlations – example
Now that we have described how to calculate correlations between countryindustry
pairs, it only remains to illustrate how to apply these to obtain correlations between indi
vidual obligors. The steps of this computation are as follows:
R
t
k ( )
R
k ( ) 1
T
 R
t
k ( )
t 1 =
T
∑
=
σ
k
1
T 1 –

R
t
k ( )
R
k ( )
– ( )
2
t 1 =
T
∑
=
COV k l , ( )
1
T 1 –

R
t
k ( )
R
k ( )
– ( ) R
t
l ( )
R
l ( )
– ( )
t 1 =
T
∑
=
ρ
k l ,
COV k l , ( )
σ
k
σ
l

=
98 Chapter 8. Credit quality correlations
CreditMetrics™—Technical Document
1. Assign weights to each obligor according to its participation in countries and
industries, and specify how much of the obligor’s equity movements are not
explained by the relevant indices.
2. Express the standardized returns for each obligor as a weighted sum of the returns
on the indices and a companyspeciﬁc component.
3. Use the weights along with the index correlations to compute the correlations
between obligors.
By specifying the amount of an obligor’s equity price movements are not explained by
the relevant indices, we are describing this obligor’s ﬁrmspeciﬁc, or idiosyncratic, risk.
Generally, prices for companies with large market capitalization will track the indices
closely, and the idiosyncratic portion of the risk to these companies is small; on the other
hand, prices for companies with less market capitalization will move more independently
of the indices, and the idiosyncratic risk will be greater.
We will explain each of the steps above through an example.
Suppose we wish to compute the correlation between two obligors, ABC and XYZ.
Assume that we decide that ABC participates only in the United States chemicals indus
try, and that its equity returns are explained 90% by returns on the United States chemi
cals index and 10% by companyspeciﬁc movements. We assume that these company
speciﬁc movements are independent of the movements of the indices, and also indepen
dent of the companyspeciﬁc movements for all other companies. Assume that XYZ
participates 75% in German insurance and 25% in German banking and ﬁnance and that
20% of the movements in XYZ’s equity are companyspeciﬁc.
To apply these weights and describe the standardized returns for the individual obligors,
we need the volatilities and correlations of the relevant indices. We present these in the
Table 8.11
. The volatilities listed are for weekly returns.
Table 8.11
Volatilities and correlations for countryindustry pairs
For the ﬁrm ABC, the volatility explained by the U.S. chemicals index is 90% of the
ﬁrm’s total volatility. The remainder is explained by ABC’s ﬁrm speciﬁc movements.
Thus, we consider two independent standard normal random variables, and
, which represent the standardized returns of the U.S. chemical index and ABC’s
ﬁrm speciﬁc standardized returns, respectively. We then write ABC’s standardized
returns as
[8.11] .
Index Volatility
Correlations
U.S.
Chemicals
Germany
Insurance
Germany
Banking
U.S.: Chemicals 2.03% 1.00 0.16 0.08
Germany: Insurance 2.09% 0.16 1.00 0.34
Germany: Banking 1.25% 0.08 0.34 1.00
r
USCm ( )
r
ˆ
ABC ( )
r
ABC ( )
w
1
r
USCm ( )
w
2
r
ˆ
ABC ( )
+ =
Sec. 8.5 Estimating asset correlations 99
Part II: Model Parameters
We know that 90% of ABC’s volatility is explained by the index, and thus we know that
. We also know that the total volatility must be one (since the returns are stan
dardized), and thus .
For XYZ, we proceed in a similar vein. We ﬁrst ﬁgure the volatility of the index move
ments for XYZ, that is, the volatility of an index formed by 75% German insurance and
25% German banking, by
[8.12]
We then scale the weights so that the total volatility of the index portion of XYZ’s stan
dardized returns is 80%. Thus, the weight on the German insurance index is
[8.13] ,
and the weight on the German banking index is
[8.14] .
Finally, in order that the total standardized return of XYZ have variance one, we know
that the weight on the idiosyncratic return must be .
At this point, we have what we will refer to as each ﬁrm’s
standard weights
, that is, the
weightings on standardized index returns which allow us to describe standardized ﬁrm
returns. Recall that for our example we describe the returns for ABC and XYZ by:
[8.15] ,
and
[8.16] ,
where and are the idiosyncratic returns for the two ﬁrms. Since the idio
syncratic returns are independent of all the other returns, we may compute the correlation
between ABC and XYZ by:
[8.17]
The above illustrates the method for computing correlations between pairs of obligors,
and suggests a more general framework. In the next subsection, we present the same
methods, but generalized to handle obligors with participations in more industries and
countries.
w
1
0.9 =
w
2
1 w
1
2
– 0.44 = =
σ
ˆ
0.75
2
σ
DeIn
2
0.25
2
σ
DeBa
2
2 0.75 0.25 ρ DeIn DeBa , ( ) σ
DeIn
σ
DeBa
⋅ ⋅ ⋅ ⋅ ⋅ + ⋅ + ⋅ 0.017. = =
0.8
0.75 σ
DeIn
⋅
σˆ

⋅ 0.74 =
0.8
0.25 σ
DeBa
⋅
σˆ

⋅ 0.15 =
1 0.8
2
– 0.6 =
r
ABC ( )
0.90r
USCm ( )
0.44r
ˆ
ABC ( )
+ =
r
XYZ ( )
0.74r
DeIn ( )
0.15r
DeBa ( )
0.6r
ˆ
XYZ ( )
+ + =
r
ˆ
ABC ( )
r
ˆ
XYZ ( )
ρ ABC XYZ , ( ) 0.90 0.74 ρ USCm DeIn , ( ) ⋅ ⋅ 0.90 0.15 ρ USCm DeBa , ( ) ⋅ ⋅ + 0.11 = =
100 Chapter 8. Credit quality correlations
CreditMetrics™—Technical Document
Note that the index volatilities do not actually enter into the correlation calculations, but
do play a role when we convert industry participations to standard weights. This allows
us to account for cases like our example, where industry participation is split 75% and
25%, or 3 to 1, but since the industry with 75% participation (insurance) is more volatile
than the other industry (banking), the standard weight on insurance is actually more than
three times greater than the standard weight on banking.
8.5.3 Obligor correlations – generalization
To complete our treatment of obligor correlations, we provide generalizations of the
methods above for computing standard weights and for calculating correlations from
these weights.
First, to compute standard weights, consider a ﬁrm with industry participations of
, , and , where the indices account for α of the movements of the ﬁrm’s equity.
We compute the ﬁrm’s standard weights in the following steps:
Compute the volatility of the weighted index for the ﬁrm, that is,
[8.18]
Scale the weights on each index such that the indices represent only α of the volatility of
the ﬁrm’s standardized returns. The scaling is as below:
[8.19] .
Compute the weight on the idiosyncratic returns by taking .
The generalization to the case of four or more indices should be clear.
Now suppose we have n different ﬁrms with standard weightings on m indices, and we
wish to compute the equity correlations between these ﬁrms. Let the correlation matrix
for the indices be denoted by C. Since the weightings are on both the indices and the
idiosyncratic components, we need to create a correlation matrix, , which covers both
of these. This matrix will be m+n by m+n, and constructed as below:
w
1
ˆ w
ˆ 2
w
ˆ 3
σˆ
w
ˆ 1
2
σ
1
2
w
ˆ 2
2
σ
2
2
w
ˆ 3
2
σ
3
2
2w
ˆ 1
w
ˆ 2
ρ
1 2 ,
σ
1
σ
2
2w
ˆ 2
w
ˆ 3
ρ
2 3 ,
σ
2
σ
3
2w
ˆ 1
w
ˆ 3
ρ
1 3 ,
σ
1
σ
3
+ + + + + =
w
1
α
w
ˆ 1
σ
1
σ
ˆ

w
2
, ⋅ α
w
ˆ 2
σ
2
σ
ˆ

and w
3
, ⋅ α
w
ˆ 3
σ
3
σ
ˆ

⋅ = = =
1 α
2
–
C
Sec. 8.5 Estimating asset correlations 101
Part II: Model Parameters
Thus, the upper left of is the m by m matrix C, representing the correlations between
indices; the lower right is the n by n identity matrix, reﬂecting that each ﬁrm’s idiosyn
cratic component has correlation one with itself and is independent of the other ﬁrms’
idiosyncratic components; and the remainder consists of only zeros, reﬂecting that there
is no correlation between the idiosyncratic components and the indices. For the example
in the previous subsection (where and ), we would have
[8.20] .
We then create a m+n by n weight matrix W, where each column represents a different
ﬁrm, and each row represents weights on indices and idiosyncratic components. Thus, in
the k
th
column of W, the ﬁrst m entries will give the ﬁrst ﬁrm’s weights on the indices,
the m+n+k entry will give the ﬁrm’s idiosyncratic weight, and the remaining entries will
be zero. For our example, the matrix would be given by
[8.21] .
The n by n matrix giving the correlations between all of the ﬁrms is then given by
.
C
0 … … 0
.
.
.
… … .
.
.
.
.
.
… … .
.
.
.
.
.
… … .
.
.
0 … … 0
0 …… 0
.
.
.
…… .
.
.
.
.
.
…… .
.
.
0 …… 0
1 0 … 1
0 .
.
.
… 0
.
.
.
.
.
.
.
.
.
0 … … 1
n
r
o
w
s
m
r
o
w
s
n columns
m columns
C
m 3 = n 2 =
C
1 0.16 0.08 0 0
0.16 1 0.34 0 0
0.08 0.34 1 0 0
0 0 0 1 0
0 0 0 0 1
=
W
W
0.90 0
0 0.74
0 0.15
0.44 0
0 0.60
=
W′ C W ⋅ ⋅
102 Chapter 8. Credit quality correlations
CreditMetrics™—Technical Document
103
Part III
Applications
104
CreditMetrics™—Technical Document
105
Part III: Applications
Overview of Part III
To this point, we have detailed an analytic approach to compute the mean and standard
deviation of portfolio value change, presented calculations for one and twoasset portfo
lios, and discussed the inputs to these calculations. In this section we discuss
approaches to computing risk measures other than standard deviation and apply the
CreditMetrics methodology to a larger portfolio.
Both issues – alternative measures of risk and computations for a larger portfolio – point
us to a central theme of this section: simulation. By this we mean the generation of
future portfolio scenarios according to the models already discussed.
Implementation of a simulation approach involves a tradeoff. On the one hand, we are
able to describe in much more detail the distribution of portfolio value changes; on the
other, we introduce noise into what has been an exact solution for the risk estimates. We
will continue to discuss this tradeoff as we go.
Part III is composed of four chapters which describe the methods and discuss the outputs
of the CreditMetrics methodology for larger portfolios. The chapters dealing with simu
lation focus on computing advanced (beyond the mean and standard deviation) risk esti
mates. This section is organized as follows:
•
Chapter 9: Analytic portfolio calculation.
We extend the methods discussed in
Chapter 3
for computing the standard deviation and marginal standard deviation to a
large (more than two instruments) portfolio.
•
Chapter 10: Simulation.
We address the assumptions necessary to specify the
portfolio distribution completely, describe the Monte Carlo approach to this distribu
tion, and discuss how to produce percentile levels as well as marginal statistics. We
focus on computing advanced (beyond the mean and standard deviation) risk esti
mates for larger portfolios.
•
Chapter 11: Portfolio example.
We choose a portfolio of 20 instruments of vary
ing maturities and rating and specify the asset correlations between their issuers. We
then utilize the simulation approach of the previous section to estimate certain risk
statistics and interpret these results in the context of the portfolio.
•
Chapter 12: Application of model outputs.
We consider how the analysis in
Chapter 11
might lead to risk management actions such as prioritizing risk reduc
tion, setting credit risk limits, and assessing economic capital.
106
CreditMetrics™—Technical Document
107
Chapter 9. Analytic portfolio calculation
In
Chapter 3
, we discussed the computation of the standard deviation of value change for
a portfolio of two instruments. We refrained from extending this computation to larger
portfolios, stating that the standard deviation of value for larger portfolios involves no
different calculations than the standard deviation for twoasset portfolios. In this chapter,
we illustrate this point for a threeasset portfolio, and discuss as well the calculation of
marginal standard deviations for this portfolio. The generalization of these calculations
to portfolios of arbitrary size is straightforward, and is detailed in
Appendix A
.
9.1 Threeasset portfolio
Our example is a portfolio consisting of three assets, all annual coupon bonds. We take
the ﬁrst two of these bonds to be issued by the BBB and A rated ﬁrms of
Chapter 3
and
the third to be a twoyear bond paying a 10% coupon and issued by a CCC rated ﬁrm.
We will refer to the ﬁrms respectively as Firms 1, 2, and 3. Suppose that the Firm 1 issue
has a notional amount of 4mm, the Firm 2 issue an amount of 2mm, and the Firm 3 issue
an amount of 1mm. Denote by , , and , the values at the end of the risk horizon
of the three respective issues.
We present transition probabilities for the three ﬁrms in
Table 9.1
below, and revalua
tions in
Table 9.2
.
Table 9.1
Transition probabilities (%)
Transition probability (%)
Rating
Firm 1 Firm 2 Firm 3
AAA 0.02 0.09 0.22
AA 0.33 2.27 0.00
A 5.95 91.05 0.22
BBB 86.93 5.52 1.30
BB 5.30 0.74 2.38
B 1.17 0.26 11.24
CCC 0.12 0.01 64.86
Default 0.18 0.06 19.79
V
1
V
2
V
3
108 Chapter 9. Analytic portfolio calculation
CreditMetrics™—Technical Document
Table 9.2
Instrument values in future ratings ($mm)
Utilizing the methods of
Chapter 2
and the information in the tables above, we may com
pute the mean value for each issue:
[9.1] , , and ,
giving a portfolio mean of . We may also compute the variance of value
for each of the three assets, obtaining
[9.2] , , and .
Note that since the standard deviations are in units of ($mm), the units for ,
, and are ($mm)
2
.
Now to compute , the standard deviation of value for the portfolio, we could use the
standard formula
[9.3] .
This would require the calculation of the various covariance terms. Alternatively, noting
that
[9.4] ,
we may express by
[9.5] .
Value of issue ($mm)
Future rating Firm 1 Firm 2 Firm 3
AAA 4.375 2.132 1.162
AA 4.368 2.130 1.161
A 4.346 2.126 1.161
BBB 4.302 2.113 1.157
BB 4.081 2.063 1.142
B 3.924 2.028 1.137
CCC 3.346 1.774 1.056
Default 2.125 1.023 0551
µ
1
$4.28mm = µ
2
$2.12mm = µ
3
$0.97mm =
µ
p
$7.38mm =
σ
2
V
1
( ) 0.014 = σ
2
V
2
( ) 0.001 = σ
2
V
3
( ) 0.044 =
σ
2
V
1
( )
σ
2
V
2
( ) σ
2
V
3
( )
σ
p
σ
p
2
σ
2
V
1
( ) σ
2
V
2
( ) σ
2
V
3
( ) 2 COV V
1
V
2
, ( )
2 + COV V
1
V
3
, ( ) 2 COV V
2
V
3
, ( ) ⋅ + ⋅
⋅ + + + =
σ
2
V
1
V
2
+ ( ) σ
2
V
1
( ) 2 COV V
1
V
2
, ( ) σ
2
V
2
( ) + ⋅ + =
σ
p
σ
p
2
σ
2
V
1
V
2
+ ( ) σ
2
V
1
V
3
+ ( ) σ
2
V
2
V
3
+ ( )
σ
2
V
1
( ) – σ
2
V
2
( ) – σ
2
V
3
( ) –
+ + =
Sec. 9.1 Threeasset portfolio 109
Part III: Applications
The above formula has the attractive feature of expressing the portfolio standard devia
tion in terms of the standard deviations of single assets (e.g. ) and the standard
deviations of twoasset subportfolios (e.g. ). Thus, to complete our computa
tion of , it only remains to identify each twoasset subportfolio, compute the standard
deviations of each, and apply Eq. [9.5].
The standard deviation for twoasset portfolios was covered in
Chapter 3
, and so in prin
ciple, we have described all of the portfolio calculations. We present the twoasset case
again as a review. Consider the ﬁrst pair of assets, the BBB and A rated bonds. In order
to compute the variance for the portfolio containing only these assets, we utilize the joint
transition probabilities in
Table 3.2
, which are an output of the asset value model of the
previous chapter, with an assumed asset correlation of 30%. Along with these probabili
ties we need the values of this twoasset portfolio in each of the 64 joint rating states; we
present these values in
Table 9.3
. Note that the values in
Table 9.3
differ from those in
Table 3.2
since the notional amounts of the issues in these two cases are different.
Table 9.3
Values of a twoasset portfolio in future ratings ($mm)
Applying Eq. [3.1] to the probabilities in
Table 3.2
and the values in
Table 9.3
, we then
compute . In a similar fashion, we specify that the asset correlations
between the ﬁrst and third and between the second and third obligors are also 30%, and
then create analogs to
Table 3.2
and
Table 9.3
. This allows us to compute
and . Finally, we apply Eq. [9.6] to obtain
, and thus .
The calculation of portfolio variance in terms of the variance of twoasset subportfolios
may seem unusual to those accustomed to the standard covariance approach. We remark
that we have all of the information necessary to compute the covariances and correlations
between our three assets. Thus, since
[9.6] ,
we have . Similarly, we obtain and
. This allows us to then compute correlations between the asset
values using
New rating for
Firm 1
(currently BBB)
New rating for Firm 2 (currently A)
AAA AA A BBB BB B CCC Default
AAA 6.51 6.51 6.50 6.49 6.44 6.40 6.15 5.40
AA 6.50 6.50 6.49 6.48 6.43 6.40 6.14 5.39
A 6.48 6.48 6.47 6.46 6.41 6.37 6.12 5.37
BBB 6.43 6.43 6.43 6.42 6.37 6.33 6.08 5.33
BB 6.21 6.21 6.21 6.19 6.14 6.11 5.86 5.10
B 6.06 6.05 6.05 6.04 5.99 5.95 5.70 4.95
CCC 5.48 5.48 5.47 5.46 5.41 5.37 5.12 4.37
Default 4.26 4.26 4.25 4.24 4.19 4.15 3.90 3.15
σ V
1
( )
σ V
1
V
2
+ ( )
σ
p
σ
2
V
1
V
2
+ ( ) 0.018 =
σ
2
V
1
V
3
+ ( ) 0.083 = σ
2
V
2
V
3
+ ( ) 0.051 =
σ
p
2
0.093 = σ
p
$0.305mm =
COV V
1
V
2
, ( )
σ
2
V
1
V
2
+ ( ) σ
2
V
1
( ) – σ
2
V
2
( ) –
2

=
COV V
1
V
2
, ( ) 0.0015 = COV V
1
V
3
, ( ) 0.0125 =
COV V
2
V
3
, ( ) 0.0030 =
110 Chapter 9. Analytic portfolio calculation
CreditMetrics™—Technical Document
[9.7] .
We then have , , and
. It is a simple matter then to check that the standard formula
Eq. [9.1] yields the same value for as we computed above.
We refer to σ
p
as the absolute measure of the portfolio standard deviation. Alternatively,
we may express this risk in percentage terms; we thus refer to σ
p
/µ
p
(which is equal to
4.1% in our example) as the percent portfolio standard deviation. These notions of abso
lute and percent measures will be used for other portfolio statistics, with the percent sta
tistic always representing the absolute statistic as a fraction of the mean portfolio value.
To extend this calculation to larger portfolios is straightforward. We present the details
of this in Appendix A.
9.2 Marginal standard deviation
As deﬁned in Section 3.3, the marginal standard deviation for a given instrument in a
portfolio is the difference between the standard deviation for the entire portfolio and the
standard deviation for the portfolio not including the instrument in question. Thus, since
we now are able to compute the standard deviation for a portfolio of arbitrary size, the
calculation of marginal standard deviations is clear.
Consider the Firm 1 issue in our portfolio above. We have seen that the standard devia
tion for the entire portfolio is $0.46mm. If we remove the Firm 1 issue, then the new
portfolio variance is given by , making the new portfolio stan
dard deviation . The marginal standard deviation of the Firm 1 issue is
then the difference between the absolute portfolio standard deviation and this ﬁgure, or
. Thus, we see that we can reduce the total portfolio standard devia
tion by $0.080mm if we liquidate the Firm 1 issue While this is a measure of the abso
lute risk contributed by the Firm 1 issue, we might also wish to characterize the riskiness
of this instrument independently of its size. To this end, we may express the marginal
standard deviation as a percentage of , the mean value of the Firm 1 issue. We refer to
this ﬁgure, 1.9% in this case, as the percent marginal standard deviation of this issue.
The difference between marginal and standalone statistics gives us an idea of the effect
of diversiﬁcation on the portfolio. Note that if we consider the Firm 1 issue alone, its
standard deviation of value is $0.117mm. If this asset were perfectly correlated with the
other assets in the portfolio, its marginal impact on the portfolio standard deviation
would be exactly this amount. However, we have seen that the marginal impact of the
Firm 1 issue is only $0.080mm, and thus that we beneﬁt from the fact that this issue is
not in fact perfectly correlated with the others.
The risk measures produced in this section may strike the reader as a bit small, particu
larly in light of the riskiness of the CCC rated issue in our example. This might be
explained by the fact that the size of this issue is quite small in comparison with the other
assets in the portfolio. However, since we have only considered the standard deviation to
this point, it may be that to adequately describe the riskiness of the portfolio, we need
CORR V
1
V
2
, ( )
COV V
1
V
2
, ( )
σ
2
V
1
( ) σ
2
V
2
( ) ×

=
CORR V
1
V
2
, ( ) 40.1% = CORR V
1
V
3
, ( ) 50.4% =
CORR V
2
V
3
, ( ) 45.2% =
σ
p
σˆ
p
2
σ
2
V
2
V
3
+ ( ) 0.051 = =
σ
ˆ
p
$0.225mm =
σ
p
σ
ˆ
p
– $0.080mm =
µ
1
Sec. 9.2 Marginal standard deviation 111
Part III: Applications
more detailed information about the portfolio distribution. In order to obtain this higher
order information, it will be necessary to perform a simulation based analysis, which is
the subject of the following two chapters.
112 Chapter 9. Analytic portfolio calculation
CreditMetrics™—Technical Document
113
Chapter 10. Simulation
Our methodology up to this point has focused on analytic estimates of risk, that is, esti
mates which are computed directly from formulas implied by the models we assume.
This analytical approach has two advantages:
1.
Speed
. Particularly for smaller portfolios, the direct calculations require fewer
operations, and thus can be computed more quickly.
2.
Precision.
No random noise is introduced in the calculations and, therefore, no
error in the risk estimates.
However, it has also two principal disadvantages. One is that for large portfolios,
number 1 above is no longer true. The other is that by restricting ourselves to analytical
approaches, we limit the available of statistics that can be estimated.
Throughout this document, we have discussed methods to compute the standard devia
tion of portfolio value; yet we have also stressed that this may not be a meaningful mea
sure of the credit risk of the portfolio. To provide a methodology that better describes the
distribution of portfolio values, we present in this chapter a simulation approach known
as “Monte Carlo.”
The three sections of this chapter treat the three steps to a Monte Carlo simulation:
1.
Generate scenarios.
Each scenario corresponds to a possible “state of the world”
at the end of our risk horizon. For our purposes, the “state of the world” is just the
credit rating of each of the obligors in our portfolio.
2.
Value portfolio.
For each scenario, we revalue the portfolio to reﬂect the new
credit ratings. This step gives us a large number of possible future portfolio values.
3.
Summarize results.
Given the value scenarios generated in the previous steps, we
have an estimate for the distribution of portfolio values. We may then choose to
report any number of descriptive statistics for this distribution.
We will continue to consider the example portfolio of the previous chapter: three two
year par bonds issued by BBB, A, and CCC rated ﬁrms. The notional values of these
bonds are $4mm, $2mm, and $1mm.
10.1 Scenario generation
In this section, we will discuss how to generate scenarios of future credit ratings for the
obligors in our portfolio. We will rely heavily on the asset value model discussed in
Sec
tion
8.4
. The steps to scenario generation are as follows:
1. Establish asset return thresholds for the obligors in the portfolio.
2. Generate scenarios of asset returns according to the normal distribution.
3. Map the asset return scenarios to credit rating scenarios.
114 Chapter 10. Simulation
CreditMetrics™—Technical Document
In
Table 10.1
below, we restate the
transition probabilities for the three issues.
We then present in
Table 10.2
1
the asset return thresholds for the three ﬁrms, which are
obtained using the methods of
Section 8.4.
Recall that the thresholds are labeled such that a return falling just below a given thresh
olds corresponds to the rating in the threshold’s subscript. That is, a return less than
Z
BB
(but greater than
Z
B
) corresponds to a rating of BB.
In order to describe how the asset values of the three ﬁrms move jointly, we state that the
asset returns in for each ﬁrm are normally distributed, and specify the correlations for
each pair of ﬁrms
2
. For our example, we assume the correlations in
Table 10.3
.
1
Recall the comment at the end of
Chapter 8
that asset return volatility does not affect the joint probabilities of rat
ing changes. For this reason, we may consider standardized asset returns, and report the thresholds for these.
2
Technically, the assumption is that the joint distribution of the asset returns of any collection of ﬁrms is multivari
ate normal.
Table 10.1
Transition probabilities (%)
Transition Probability (%)
Rating Firm 1 Firm 2 Firm 3
AAA 0.02 0.09 0.22
AA 0.33 2.27 0.00
A 5.95 91.05 0.22
BBB 86.93 5.52 1.30
BB 5.30 0.74 2.38
B 1.17 0.26 11.24
CCC 0.12 0.01 64.86
Default 0.18 0.06 19.79
Table 10.2
Asset return thresholds
Threshold Firm 1 Firm 2 Firm 3
Z
AA
3.54 3.12 2.86
Z
A
2.78 1.98 2.86
Z
BBB
1.53 1.51 2.63
Z
BB
1.49 2.30 2.11
Z
B
2.18 2.72 1.74
Z
CCC
2.75 3.19 1.02
Z
Def
2.91 3.24 0.85
Sec. 10.1 Scenario generation 115
PartI III: Applications
Table 10.3
Correlation matrix for example portfolio
Generating scenarios for the asset returns of our three obligors is a simple matter of gen
erating correlated, normally distributed variates. There are a number of methods for
doing this – Cholesky factorization, singular value decomposition, etc. – for discussions
of which see, for example, Strang [88]. In
Table 10.4
, we list ten scenarios which might
be produced by such a procedure. In each scenario, the three numbers represent the stan
dardized asset return for each of the three ﬁrms.
Table 10.4
Scenarios for standardized asset returns
To fully specify our scenarios, it is only necessary to assign ratings to the asset return
scenarios. For example, consider scenario 2 of
Table 10.4
. The standardized return for
Firm 1 is –2.1060, which falls between
Z
B
(–2.18 from
Table 10.2
) and
Z
BB
(–1.49 from
Table 10.2
) for this name. This corresponds to a new rating of BB. For Firm
2, the return is –2.0646, which falls between
Z
BB
and
Z
BBB
for this name, corresponding
to a new rating of BBB. Continuing this process, we may ﬁll in
Table 10.5
, which com
pletes the process of scenario generation
Firm 1 Firm 2 Firm 3
Firm 1 1.0 0.3 0.1
Firm 2 0.3 1.0 0.2
Firm 3 0.1 0.2 1.0
Scenario Firm 1 Firm 2 Firm 3
1 0.7769 0.8750 0.6874
2 2.1060 2.0646 0.2996
3 0.9276 0.0606 2.7068
4 0.6454 0.1532 1.1510
5 0.4690 0.5639 0.2832
6 0.1252 0.5570 1.9479
7 0.6994 1.5191 1.6503
8 1.1778 0.6342 1.7759
9 1.8480 2.1202 1.1631
10 0.0249 0.4642 0.3533
116 Chapter 10. Simulation
CreditMetrics™—Technical Document
Notice that for this small number of trials, the scenarios do not correspond precisely to
the transition probabilities in
Table 10.1
. (For example, in four of the ten scenarios, Firm
3 defaults, while the probability that this occurs is just 20%.) These random ﬂuctuations
are the source of the lack of precision in Monte Carlo estimation. As we generate more
scenarios, these ﬂuctuations become less prominent, but it is important to quantify how
large we can expect the ﬂuctuations to be. This is the topic of
Appendix B
.
10.2 Portfolio valuation
For nondefault scenarios, this step is no different here than in the previous chapters. For
each scenario and each issue, the new rating maps directly to a new value. To recall the
speciﬁcs of valuation, refer back to
Chapter 4
.
For default scenarios, the situation is slightly different. We discussed in
Chapter 7
that
recovery rates are not deterministic quantities but rather display a large amount of varia
tion. This variation of value in the case of default is a signiﬁcant contributor to risk. To
model this variation, we obtain the mean and standard deviation of recovery rate for each
issue in our portfolio according to the issue’s seniority. For example, in our BBB rated
senior unsecured issue, the recovery mean is 53% and the recovery standard deviation is
33%. For each default scenario, we generate a random recovery rate according to a beta
distribution
3
with these parameters
4
. These recovery rates then allow us to obtain the
value in each default scenario.
In the end, we obtain a portfolio value for each scenario. The results for the ﬁrst ten sce
narios for our example are presented in
Table 10.6
.
3
Recall that the beta distribution only produces numbers between zero and one, so that we are assured of obtaining
meaningful recovery rates.
4
Note that we assume here that the recovery rate for a given obligor is independent of the value of all other instru
ments in the portfolio.
Table 10.5
Mapping return scenarios to rating scenarios
Asset Return New Rating
Scenario Firm 1 Firm 2 Firm 3 Firm 1 Firm 2 Firm 3
1 0.7769 0.8750 0.6874 BBB A CCC
2 2.1060 2.0646 0.2996 BB BBB CCC
3 0.9276 0.0606 2.7068 BBB A A
4 0.6454 0.1532 1.1510 BBB A Default
5 0.4690 0.5639 0.2832 BBB A CCC
6 0.1252 0.5570 1.9479 BBB A Default
7 0.6994 1.5191 1.6503 BBB A Default
8 1.1778 0.6342 1.7759 BBB A Default
9 1.8480 2.1202 1.1631 A AA B
10 0.0249 0.4642 0.3533 BBB A CCC
Sec. 10.3 Summarizing the results 117
PartI III: Applications
Note that for a given issue, the value is the same in scenarios with the same (nondefault)
credit rating. For defaults, this is not the case – the values of the Firm 3 issue in the
default scenarios are different – since recovery rates are themselves uncertain. Thus,
each default scenario requires an independently generated recovery rate.
10.3 Summarizing the results
At this point, we have created a number of possible future portfolio values. The ﬁnal task
is then to synthesize this information into meaningful risk estimates.
In this section, we will examine a number of descriptive statistics for the scenarios we
have created. In the section to follow, we will examine the same statistics, but for an
example in which we consider a larger portfolio and a larger number of scenarios, so as
to obtain more signiﬁcant results.
In order to gain some intuition about the distribution of values, we ﬁrst examine a plot of
the ten scenarios for our example. This plot is presented in
Chart 10.1
. For a larger num
ber of scenarios, we would expect this plot to become more smooth, and approach some
thing like the histogram we will see in
Chart 11.1
.
Table 10.6
Valuation of portfolio scenarios ($mm)
Rating Value
Scenario Firm 1 Firm 2 Firm 3 Firm 1 Firm 2 Firm 3 Portfolio
1 BBB A CCC 4.302 2.126 1.056 7.484
2 BB BBB CCC 4.081 2.063 1.056 7.200
3 BBB A A 4.302 2.126 1.161 7.589
4 BBB A Default 4.302 2.126 0.657 7.085
5 BBB A CCC 4.302 2.126 1.056 7.484
6 BBB A Default 4.302 2.126 0.754 7.182
7 BBB A Default 4.302 2.126 0.269 6.697
8 BBB A Default 4.302 2.126 0.151 6.579
9 A AA B 4.346 2.130 1.137 7.613
10 BBB A CCC 4.302 2.126 1.056 7.484
118 Chapter 10. Simulation
CreditMetrics™—Technical Document
Chart 10.1
Frequency plot of portfolio scenarios
Even for small number of scenarios, we begin to see the heavy downside tail typical of
credit portfolio distributions.
The ﬁrst statistics we examine are those which we are able to compute analytically: the
mean and standard deviation of future portfolio value. Let
V
(1)
,
V
(2)
,
V
(3)
,...
indicate the
portfolio value in the respective scenarios. Then we may compute the sample mean (
µ
)
and standard deviation (
σ
) of the scenarios as follows:
[10.1]
where
N
is the number of scenarios (in our case,
N
=10).
As we have mentioned before, the mean and standard deviation may not be the best mea
sures of risk in that, since the distribution of values is not normal, we cannot infer per
centile levels from the standard deviation. We are thus motivated to perform simulations
in order to capture more information about the distribution of values. Estimates of per
centile levels are straightforward. For example, to compute the tenth percentile given our
scenarios, we choose a level (
x
) at which one of the ten scenarios is less than
x
and the
other nine scenarios are greater than
x
. For our scenarios, this level is between $6.58mm
and $6.70mm. This imprecision is due to simulation noise, but we will see in the next
chapter that as we consider more scenarios, our estimates of percentiles become more
precise.
To this point, we have considered only statistics which describe the portfolio distribu
tion. We would also like to consider individual assets and to ascertain how much risk
each asset contributes to the portfolio. To this end, we will describe marginal statistics.
We have discussed marginal standard deviations previously. This concept may be gener
alized, and we may compute a marginal analog of any of the statistics (standard devia
tion, percentile) discussed above. In general, the marginal statistic for a particular asset is
the difference between that statistic for the entire portfolio and that statistic for the port
folio not including the asset in question. Thus, if we wish to compute the marginal tenth
percentile of the third asset in our portfolio (the CCC rated bond), we take
6.4 6.6 6.8 7.0 7.2 7.4 7.5 7.7
0.00
0.05
0.10
0.15
0.20
0.25
0.30
Frequency
µ
p
1
N
 V
i ( )
$7.24mm and σ
p
=
i 1 =
N
∑
1
N 1 –
 V
i ( )
µ – ( )
2
i 1 =
N
∑
$0.37mm = = =
Sec. 10.3 Summarizing the results 119
PartI III: Applications
[10.2]
where
V
1
,
V
2
, and
V
3
represent the future values of the ﬁrst, second, and third assets,
respectively, and
θ
10
represents the tenth percentile of the values in question. For the sce
narios above, the tenth percentile for the entire portfolio is $6.64mm, while that for just
the ﬁrst two assets is $6.29mm; and thus the marginal standard deviation for the third
asset is $0.35mm. This marginal ﬁgure may be interpreted as the amount by which we
could decrease the risk on our portfolio by removing the CCC rated bond.
As we have mentioned a number of times, the statistics obtained through Monte Carlo
simulation are subject to ﬂuctuations; any set of scenarios may not produce a sample
mean or sample 5
th
percentile which is equal to the true mean or 5
th
percentile for the
portfolio. Thus, it is important to quantify, given the number of scenarios which are gen
erated, how close we expect our estimates of various portfolio statistics to be to their true
value. In fact, a reasonable way to choose the number of scenarios to be generated is to
specify some desired level of precision for a particular statistic, and generate enough sce
narios to achieve this. Quantifying the precision of simulation based statistics is the sub
ject of
Appendix B
.
θ
10
V
1
V
2
V
3
+ + ( ) θ
10
V
1
V
2
+ ( ) –
120 Chapter 10. Simulation
CreditMetrics™—Technical Document
121
Chapter 11. Portfolio example
In this chapter, we examine a more realistic example portfolio and discuss the results of a
simulationbased analysis of this portfolio. The risk estimates are no different than those
in the previous chapter, but should take on more meaning here in the context of a larger
portfolio.
11.1 The example portfolio
In this chapter, we consider a portfolio of 20 corporate bonds (each with a different
issuer) of varying rating and maturity. The bonds are listed in
Table 11.1
. The total mar
ket value of the portfolio is $68mm.
Table 11.1.
Example portfoli
o
Recall that for each asset, the credit rating determines the distribution of future credit rat
ing, and thus also the distribution of future value. For the portfolio, however, we must
also specify the asset correlations in order to describe the distribution of future ratings
and values. For this example, we assume the correlations in
Table 11.2
.
Asset
Credit
rating
Principal
amount
Maturity
(years)
Market
value
1 AAA 7,000,000 3 7,821,049
2 AA 1,000,000 4 1,177,268
3 A 1,000,000 3 1,120,831
4 BBB 1,000,000 4 1,189,432
5 BB 1,000,000 3 1,154,641
6 B 1,000,000 4 1,263,523
7 CCC 1,000,000 2 1,127,628
8 A 10,000,000 8 14,229,071
9 BB 5,000,000 2 5,386,603
10 A 3,000,000 2 3,181,246
11 A 1,000,000 4 1,181,246
12 A 2,000,000 5 2,483,322
13 B 600,000 3 705,409
14 B 1,000,000 2 1,087,841
15 B 3,000,000 2 3,263,523
16 B 2,000,000 4 2,527,046
17 BBB 1,000,000 6 1,315,720
18 BBB 8,000,000 5 10,020,611
19 BBB 1,000,000 3 1,118,178
20 AA 5,000,000 5 6,181,784
122 Chapter 11. Portfolio example
CreditMetrics™—Technical Document
Table 11.2
Asset correlations for example portfolio
Observe that there are ﬁve groups of issuers (those for assets 14, 610, 1115, 1618,
and 1920, in the shaded areas of the table) within which the asset correlations are rela
tively high, while the correlations between these groups are lower. This might be the
case for a portfolio containing issues from ﬁrms in ﬁve different industries; the correla
tions between ﬁrms in a given industry are high, while correlations across industries are
lower.
11.2 Simulation results
Using the methodology of the previous chapter, we generate 20,000 portfolio scenarios,
that is, 20,000 possible future occurrences in one year’s time of the credit ratings for
each of our issues. For each scenario, we then obtain a portfolio value for one year into
the future. In Charts
11.1
through
11.3
, we present histograms of the portfolio value sce
narios. Note the axes on each chart carefully. The ﬁrst chart illustrates the distribution
of the most common scenarios, the second moves a bit further into the left tail of the dis
tribution, and the third shows the distribution of the most extreme 5% of all cases. The
vertical axis, which represents relative frequency, is ten times smaller in the second chart
than in the ﬁrst, and twenty times smaller in the third chart than in the second.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
1 1 0.45 0.45 0.45 0.15 0.15 0.15 0.15 0.15 0.15 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
2 0.45 1 0.45 0.45 0.15 0.15 0.15 0.15 0.15 0.15 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
3 0.45 0.45 1 0.45 0.15 0.15 0.15 0.15 0.15 0.15 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
4 0.45 0.45 0.45 1 0.15 0.15 0.15 0.15 0.15 0.15 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
5 0.15 0.15 0.15 0.15 1 0.35 0.35 0.35 0.35 0.35 0.2 0.2 0.2 0.2 0.2 0.15 0.15 0.15 0.1 0.1
6 0.15 0.15 0.15 0.15 0.35 1 0.35 0.35 0.35 0.35 0.2 0.2 0.2 0.2 0.2 0.15 0.15 0.15 0.1 0.1
7 0.15 0.15 0.15 0.15 0.35 0.35 1 0.35 0.35 0.35 0.2 0.2 0.2 0.2 0.2 0.15 0.15 0.15 0.1 0.1
8 0.15 0.15 0.15 0.15 0.35 0.35 0.35 1 0.35 0.35 0.2 0.2 0.2 0.2 0.2 0.15 0.15 0.15 0.1 0.1
9 0.15 0.15 0.15 0.15 0.35 0.35 0.35 0.35 1 0.35 0.2 0.2 0.2 0.2 0.2 0.15 0.15 0.15 0.1 0.1
10 0.15 0.15 0.15 0.15 0.35 0.35 0.35 0.35 0.35 1 0.2 0.2 0.2 0.2 0.2 0.15 0.15 0.15 0.1 0.1
11 0.1 0.1 0.1 0.1 0.2 0.2 0.2 0.2 0.2 0.2 1 0.45 0.45 0.45 0.45 0.2 0.2 0.2 0.1 0.1
12 0.1 0.1 0.1 0.1 0.2 0.2 0.2 0.2 0.2 0.2 0.45 1 0.45 0.45 0.45 0.2 0.2 0.2 0.1 0.1
13 0.1 0.1 0.1 0.1 0.2 0.2 0.2 0.2 0.2 0.2 0.45 0.45 1 0.45 0.45 0.2 0.2 0.2 0.1 0.1
14 0.1 0.1 0.1 0.1 0.2 0.2 0.2 0.2 0.2 0.2 0.45 0.45 0.45 1 0.45 0.2 0.2 0.2 0.1 0.1
15 0.1 0.1 0.1 0.1 0.2 0.2 0.2 0.2 0.2 0.2 0.45 0.45 0.45 0.45 1 0.2 0.2 0.2 0.1 0.1
16 0.1 0.1 0.1 0.1 0.15 0.15 0.15 0.15 0.15 0.15 0.2 0.2 0.2 0.2 0.2 1 0.55 0.55 0.25 0.25
17 0.1 0.1 0.1 0.1 0.15 0.15 0.15 0.15 0.15 0.15 0.2 0.2 0.2 0.2 0.2 0.55 1 0.55 0.25 0.25
18 0.1 0.1 0.1 0.1 0.15 0.15 0.15 0.15 0.15 0.15 0.2 0.2 0.2 0.2 0.2 0.55 0.55 1 0.25 0.25
19 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.25 0.25 0.25 1 0.65
20 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.25 0.25 0.25 0.65 1
Sec. 11.2 Simulation results 123
Part III: Applications
Chart 11.1
Histogram of future portfolio values – upper 85% of scenarios
Chart 11.2
Histogram of future portfolio values – scenarios
between 95
th
and 65
th
percentiles
65 66.8 67.3 67.6 67.8 68.0
0
1
2
3
4
5
6
7
8
9
10
Portfolio value ($mm)
Relative frequency
See Cht. 11.2
59 65 66.4 66.9 67.2 67.5
0
0.2
0.4
0.6
0.8
1.0
1.2
Portfolio value ($mm)
Relative frequency
See Cht. 11.3
124 Chapter 11. Portfolio example
CreditMetrics™—Technical Document
Chart 11.3
Histogram of future portfolio values – lower 5% of scenarios
We may make several interesting observations of these charts. First, by far the most
common occurrence (almost 9% of all scenarios, exhibited by the spike near $67.8mm in
Table 11.1
) is that none of the issuers undergoes a rating change. Further, in well over
half of the scenarios, there are no signiﬁcant credit events, and the portfolio appreciates.
The second observation is the odd bimodal structure of the distribution. This is due to
the fact that default events produce much more signiﬁcant value changes than any other
rating migrations. Thus, the distribution of portfolio value is driven primarily by the
number of issues which default. The second hump in the distribution (the one between
$67mm and $67.2mm) represents scenarios in which one issue defaults.
The two other humps further to the left in the distribution represent scenarios with two
and three defaults, respectively. For larger portfolios, these humps become even more
smoothed out, while for smaller ones, the humps are generally more prominent.
Regardless of the particulars of the shape of the value distribution, one feature persists:
the heavy downward skew. Our example distribution is no different, displaying a large
probability of a marginal increase in value along with a small probability of a more sig
niﬁcant drop in value.
As in the previous chapter, the ﬁrst two statistics we present are the mean and standard
deviation of the portfolio value. For our case, we have:
• Mean portfolio value (
µ
) = $67,284,888.
• Standard deviation of portfolio value (
σ
) = $1,136,077.
As we have mentioned before, the mean and standard deviation may not be the best mea
sures of risk in that, since the loss distribution is not normal, we cannot infer conﬁdence
levels from these parameters. We can however estimate percentiles directly from our
scenarios.
Relative frequency
56 58 60 62 64 65
0
0.005
0.010
0.015
0.020
0.025
0.030
Portfolio value ($mm)
Sec. 11.3 Assessing precision 125
Part III: Applications
For example, if we wish to compute the 5
th
percentile (the level below which we esti
mate that 5% of portfolio values fall), we sort our 20,000 scenarios in ascending order
and take the 1000
th
of these sorted scenarios (that is, $64.98mm) as our estimate. (Our
assumption is then that since 5% of the simulated changes in value were less than
$5.69mm, there is a 5% chance that the actual portfolio value change will be less than
this level.) Here we see the advantage of the simulation approach, in that we can esti
mate arbitrary percentile levels, where in the analytic approach, because the portfolio
distribution is not normal, we are only able to compute two statistics.
In
Table 11.3
below, we present various percentiles of our scenarios of future portfolio
values. For comparison and in order to illustrate the nonnormality of the portfolio dis
tribution, we also give the percentiles which we would have estimated had we utilized
the sample mean and standard deviation, and assumed that the distribution was normal.
Table 11.3
Percentiles of future portfolio values ($mm)
Using the scenarios, we estimate that 2.5% of the time (or one year in forty), our portfo
lio in one year will drop in value to $63.97mm or less. If we had used a normal assump
tion, we would have estimated that this percentile would correspond to only a drop to
$65.06mm, a much more optimistic risk estimate.
On the other hand, if we examine the median value change (the 50% level), the normal
assumption leads to a more pessimistic forecast: there is a 50% chance that the portfolio
is less valuable than the mean value of $67.28mm. By contrast, the scenarios point to a
higher mean, and thus to a greater than 50% chance that the portfolio value will exceed
its mean.
Another interesting observation is that the 5
th
and 1
st
percentiles of the scenarios are 2
and 2.9 standard deviations, respectively, below the mean. This is further evidence that
it is best not to use the standard deviation to infer percentile levels for a credit portfolio.
11.3 Assessing precision
In this section, we utilize the methods of
Appendix B
to give conﬁdence bands around
our estimated statistics, and examine how these conﬁdence bands evolve as we increase
the number of scenarios which we consider.
For the 20,000 scenarios in our example, we have the results shown in
Table 11.4
.
Actual scenarios Normal distribution
Percentile
Portfolio value
($mm) Formula
Portfolio value
($mm)
95% 67.93
µ+1.65σ
69.15
50% 67.80
µ
67.28
5% 64.98
µ−1.65σ
65.42
2.5% 63.97
µ−1.96σ
65.06
1% 62.85
µ−2.33σ
64.64
0.5% 61.84
µ−2.58σ
64.36
0.1% 57.97
µ−3.09σ
63.77
126 Chapter 11. Portfolio example
CreditMetrics™—Technical Document
Table 11.4
Portfolio value statistics with 90% conﬁdence levels ($mm)
For both the mean and standard deviation, and for the 5
th
and 1
st
percentiles, the conﬁ
dence bands are reasonably tight, and we feel assured of making decisions based on our
estimates of these quantities. For the more extreme percentiles, we see that the true loss
level could well be at least 10% greater than our estimate. If we desire estimates for
these levels, we would be best off generating more scenarios.
With regard to the question of how many scenarios we need to obtain precise estimates,
we may examine the evolution of our conﬁdence bands for each estimate as we consider
more and more scenarios. We present this information for the six statistics above in the
following charts.
Chart 11.4
Evolution of conﬁdence bands for portfolio mean ($mm)
Statistic Lower bound Estimate Upper bound
Mean portfolio value 67.27 67.28 67.30
Standard deviation 1.10 1.14 1.17
5th percentile 64.94 64.98 65.02
1st percentile 62.66 62.85 62.97
0.5 percentile
1
61.26 61.84 62.08
0.1 percentile
2
56.11 57.97 58.73
1
1 in 200 chance of shortfall
2
1 in 1,000 chance of shortfall
Number of scenarios
0 5,000 10,000 15,000 20,000
67.0
67.1
67.2
67.3
67.4
67.5
Confidence bands
Estimate
Mean portfolio value ($mm)
Sec. 11.3 Assessing precision 127
Part III: Applications
Chart 11.5
Evolution of conﬁdence bands for standard deviation ($mm)
Chart 11.6
Evolution of conﬁdence bands for 5
th
percentile ($mm)
0 5,000 10,000 15,000 20,000
0.75
1.00
1.25
1.50
1.75
Confidence bands
Estimate
Number of scenarios
Standard deviation
0 5,000 10,000 15,000 20,000
58
60
62
64
66
Number of scenarios
Confidence bands
Estimate
Portfolio value ($mm)
128 Chapter 11. Portfolio example
CreditMetrics™—Technical Document
Chart 11.7
Evolution of conﬁdence bands for 1
st
percentile ($mm)
Chart 11.8
Evolution of conﬁdence bands for 0.5 percentile ($mm)
0 5,000 10,000 15,000 20,000
56
58
60
62
64
Number of scenarios
Confidence bands
Estimate
Portfolio value ($mm)
0 5,000 10,000 15,000 20,000
56
58
60
62
64
Number of scenarios
Confidence bands
Estimate
Portfolio value ($mm)
Sec. 11.4 Marginal risk measures 129
Part III: Applications
Chart 11.9
Evolution of conﬁdence bands for 0.1 percentile ($mm)
It is interesting to note here that few of the plots change beyond about 10,000 scenarios;
we could have obtained similar estimates and similar conﬁdence bands with only half the
effort. In fact, if we had been most concerned with the 5
th
percentile, we might have
been satisﬁed with the precision of our estimate after only 5000 trials, and could have
stopped our calculations then. For the most extreme percentile level, note that the esti
mates and conﬁdence bands do not change frequently. This is due to the fact that on
average only one in one thousand scenarios produces a value which truly inﬂuences our
estimate. This suggests that to meaningfully improve our estimate will require a large
number of additional scenarios.
11.4 Marginal risk measures
To examine the contribution of each individual asset to the risk of the portfolio, we com
pute marginal statistics. Recall that for any risk measure, the marginal risk of a given
asset is the difference between the risk for the entire portfolio and the risk of the portfolio
without the given asset.
As an example, let us consider the standard deviation. For each asset in the portfolio, we
will compute four numbers. First, we compute each asset’s
standalone standard devia
tion
of value, that is the standard deviation of value for the asset computed without
regard for the other instruments in the portfolio. Second, we compute the
standalone
percent standard deviation
, which is just the standalone standard deviation expressed as
a percentage of the mean value for the given asset. Third, we compute each asset’s
mar
ginal standard deviation
, the impact of the given asset on the total portfolio standard
deviation. Last, we express this ﬁgure in percent terms, giving the
percent marginal
standard deviation
. These four statistics are presented for each of the 20 assets in
Table 11.5
.
0 5,000 10,000 15,000 20,000
50
52
54
56
58
60
62
Number of scenarios
Confidence bands
Estimate
Portfolio value ($mm)
130 Chapter 11. Portfolio example
CreditMetrics™—Technical Document
Table 11.5
Standard deviation of value change
The difference between the standalone and marginal risk for a given asset is an indica
tion of the effect of diversiﬁcation. We see in general that for the higher rated assets,
there is a greater reduction from the standalone to marginal risk than for the lower rated
assets. This is in line with our intuition that a much larger portfolio is required to diver
sify the effects of riskier credit instruments.
An interesting way to visualize these outputs is to plot the percent marginal standard
deviations against the market value of each asset, as in
Chart 11.10
. Points in the upper
left of the chart represent assets which are risky in percent terms, but whose exposure
sizes are small, while points in the lower right represent large exposures which have rel
atively small chances of undergoing credit losses. Note that the product of the two coor
dinates (that is, the percent risk multiplied by the market value) gives the absolute
marginal risk. The curve in
Chart 11.10
represents points with the same absolute risk;
points which fall above the curve have greater absolute risk, while points which fall
below have less.
Standalone Marginal
Asset Credit rating Absolute ($) Percent Absolute ($) Percent
1 AAA 4,905 0.06 239 0.00
2 AA 2,007 0.17 114 0.01
3 A 17,523 1.56 693 0.06
4 BBB 40,043 3.37 2,934 0.25
5 BB 99,607 8.63 16,046 1.39
6 B 162,251 12.84 37,664 2.98
7 CCC 255,680 22.67 73,079 6.48
8 A 197,152 1.39 35,104 0.25
9 BB 380,141 7.06 105,949 1.97
10 A 63,207 1.99 5,068 0.16
11 A 15,360 1.30 1,232 0.10
12 A 43,085 1.73 4,531 0.18
13 B 107,314 15.21 25,684 3.64
14 B 167,511 15.40 44,827 4.12
15 B 610,900 18.72 270,000 8.27
16 B 322,720 12.77 89,190 3.53
17 BBB 28,051 2.13 2,775 0.21
18 BBB 306,892 3.06 69,624 0.69
19 BBB 1,837 0.16 120 0.01
20 AA 9,916 0.16 389 0.01
Sec. 11.4 Marginal risk measures 131
Part III: Applications
Chart 11.10
Marginal risk versus current value for example portfolio
Based on the discussion above, we may identify with the aid of the curve the ﬁve greatest
contributors to portfolio risk. Some of these “culprits” are obvious: Asset 7 is the CCC
rated issue, and has a much larger likelihood of default, whereas Asset 18 is BBB rated,
but is a rather large exposure.
On the other hand, the other “culprits” seem to owe their riskiness as much to their corre
lation with other instruments as to their individual characteristics. For instance, Asset 9
has a reasonably secure BB rating, but has a correlation of 35% with Asset 7, the CCC
rated issue, while Asset 16 is rated B, but has a 55% correlation with Asset 18. Finally,
the appearance of Asset 15 as the riskiest in absolute terms seems to be due as much to
its 45% correlation with two other B issues as to its own B rating.
With this, we conclude the chapter. The reader should now an understanding of the vari
ous descriptors of the future portfolio distribution which can be used to assess risk. In the
following chapter, we step away from the technical, and discuss what policy implications
the assessment of credit risk might have, as well as how the use of a risk measure should
inﬂuence the decision on precisely which measure to use.
•
•
•
••
•
•
••
•
•
• •
•
•
•
•
• •
•
•
0.0%
2.5%
5.0%
7.5%
10.0%
0 5 10 15
Marginal standard deviation
Asset 7
Asset 16
Asset 9
Asset 18
Credit exposure ($mm)
Asset 15
132 Chapter 11. Portfolio example
CreditMetrics™—Technical Document
133
Chapter 12. Application of model outputs
The measures of credit risk outlined in the preceding sections can have a variety of appli
cations; we will highlight just a few:
• to set priorities for actions to reduce the portfolio risk;
• to measure and compare credit risks so that an institution can best apportion scarce
risktaking resources by limiting overconcentrations; and
• to estimate
economic capital
required to support risktaking.
The objective of all of the above is to utilize risktaking capacity more efﬁciently.
Whether this is achieved by setting limits and insisting on being adequately compensated
for risk, or by allocating capital to functions which have proven to take risk most effec
tively, is a policy issue. The bottom line is that in order to optimize the return we receive
for the risk we take, it is necessary to measure the risk we take; and this is the contribu
tion of CreditMetrics.
Note that we do not address the issue of credit pricing. Although credit risk can be an
important input into a credit pricing decision, we believe that there are signiﬁcant other
determinants for pricing which are beyond the scope of CreditMetrics. These additional
factors are nontrivial and so we have chosen to focus this current version on the already
challenging task of risk estimation.
1
12.1 Prioritizing risk reduction actions
The primary purpose of any risk management system is to direct
actions
. But there are
many actions that may be taken towards addressing risk – so they must be prioritized.
For this discussion, we will make reference to
Chart 12.1
, which is exactly like
Chart 11.10
, but for a hypothetical portfolio with a very large number of exposures.
There are at least two features of risk which are worth reducing, but the tradeoff
between them is judgmental: (i) absolute exposure size, and (ii) statistical risk level.
Thus approaches include:
• reevaluate obligors having the largest
absolute size
(the lower right corner of the
chart) arguing that a single default among these would have the greatest impact.
• reevaluate obligors having the highest
percentage level of risk
(the upper left corner
of the chart) arguing that these are the most likely to contribute to portfolio losses.
1
Researchers interested in valuation and pricing models may refer to the following: Das & Tufano [96], Foss [95],
Jarrow & Turnbull [95], Merton [74], Shimko, Tejima & Van Deventer [93], Skinner [94], and Sorensen & Bollier
[94]. Other research on historical credit price levels and relationships includes: Altman & Haldeman [92], Eber
hart, Moore & Roenfeldt [90], Fridson & Gao [96], Hurley & Johnson [96], Madan & Unal [96], Neilsen & Ronn
[96], and Sarig & Warga [89].
134 Chapter 12. Application of model outputs
CreditMetrics™—Technical Document
• reevaluate obligors contributing the largest
absolute amount of risk
(points towards
the upper right corner of the chart) arguing that these are the single largest contribu
tors to portfolio risk.
Although all three approaches are perfectly valid, we advocate the last one, setting as the
highest priority to address those obligors which are both relatively high percentage risk
and relatively large exposure. These are the parties which contribute the greatest volatil
ity to the portfolio. In practice, these are often “fallen angels,” whose large exposures
were created when their credit ratings were better, but who now have much higher per
centage risk due to recent downgrades.
Chart 12.1
Risk versus size of exposures within a typical credit portfolio
Like
Chart 11.10
, this chart illustrates a risk versus size proﬁle for a credit portfolio.
Obligors with high percentage risk – and presumably high anticipated return – can be tol
erated if they are small in size. Large exposures are typically allowed only if they have
relatively small percentage risk levels. Unfortunately, the quality of a credit can change
over time and a large exposure may have its credit rating downgraded (i.e., its point will
move straight up in this chart). The portfolio will then have a large exposure with also a
relatively large absolute level of risk. It is this type of obligor which we advocate
addressing ﬁrst.
Chart 12.1
does not completely describe the portfolio in question, however, as it does
not address the issue of returns. Thus, there is another issue to consider when consider
ing which exposures should be addressed: whether the returns on the exposures in ques
tion adequately compensate their risk. This is where the power of a portfolio analysis
becomes evident. In general, it can be assumed that assets will be priced according to
their risk on a standalone basis, or otherwise, in a CAPM (capital asset pricing model)
framework, according to their correlation with a broad universe of assets. What this
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
0
1
2
3
4
5
6
7
8
9
10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Marginal standard deviation % (by obligor)
Absolute exposure size (by obligor)
Highest standard deviation %  Low impact
High risk and large size
High absolute standard deviation  low percentage risk
Sec. 12.2 Credit risk limits 135
Part III: Applications
means is that a given asset may contribute differently to the risk of distinct portfolios,
and yet yield the same returns in either case.
Consequently, we can imagine the following situation. Two managers identify a risky
asset in their portfolios. It turns out that the two assets are of the same maturity, credit
rating, and price, and are expected to yield equivalent returns. However, because of the
structure of the two portfolios, if the managers swap these assets, the risk of both portfo
lios will be reduced without the expected return on either being affected. This might be
the case if two managers are heavily concentrated in two different industries. By swap
ping similar risky assets, the managers reduce their concentration, and thus their risk,
without reducing their expected proﬁts.
We see then not only the importance of evaluating the contribution of each asset to the
risk of the portfolio, but also the identiﬁcation of how each asset makes its contribution.
When the risk of an asset is due largely to concentrations particular to the portfolio, as in
the example above, an opportunity could well exist to restructure the portfolio in such a
way as to reduce its risk without altering its proﬁtability.
12.2 Credit risk limits
In terms of policy rigor, the next step beyond using risk statistics for prioritization is to
use them for limit setting. Of course, what type of risk measure to use for limits, as well
as what type of policy to take with regard to the limits, are management decisions. In
this section, we discuss three aspects a user might consider with regard to using
CreditMetrics for limit purposes: what type of limit to set, which risk measure to use for
the limits, and what policy to employ with regard to the limits.
12.2.1 Types of credit risk limits
This section’s discussion will make reference to
Chart 12.2
, which the reader might rec
ognize as exactly the same as
Chart 11.10
, but with two additional barriers included.
Chart 12.2
Possible risk limits for an example portfolio
•
•
•
••
•
•
•
•
•
•
• •
•
•
•
•
• •
•
•
0%
1%
2%
3%
4%
5%
6%
7%
8%
9%
10%
$0 $5,000,000 $10,000,000 $15,000,000
Marginal standard deviation
Market value
Asset 15
Asset 7
Asset 16
Asset 9
Asset 18
136 Chapter 12. Application of model outputs
CreditMetrics™—Technical Document
We might consider each of the three possibilities mentioned in the previous section as
candidates for credit risk limits. We treat each in turn:
•
Set limits based on percentage risk.
This would correspond to a limit like the hori
zontal line in
Chart 12.2.
If we measured risk in absolute terms, this would corre
spond exactly to a limit on credit quality, that is, a limit restricting the portfolio to
contain only exposures rated, say, B or higher. Since we measure risk in marginal
terms, this limit would be slightly different in that it would also restrict exposures
that are more correlated to the portfolio, since these contribute more to portfolio risk.
•
Set limits based on exposure size.
This would correspond to a limit like the vertical
bar in
Chart 12.2.
Such a limit would restrict the portfolio to have no exposures,
regardless of credit quality, above a given size.
•
Set limits based on absolute risk.
This would correspond to a limit like the curve in
Chart 12.2.
Such a limit would prevent the addition to the portfolio of any exposure
which increased the portfolio risk by more than a given amount. In effect, this
would cap the total risk of the portfolio at a certain amount above the current risk.
In the previous section, we argued that it is best to address exposures with the highest
level of absolute risk ﬁrst, since these have the greatest impact on the total portfolio risk.
By the same token, it is most sensible to set limits in terms of absolute (rather than per
cent) risk. Moreover, limiting absolute risk is consistent with the natural tendencies of
portfolio managers; in other words, it is perfectly intuitive to require that exposures
which pose a greater chance of decreases in value due to credit be smaller, while allow
ing those with less chance of depreciating to be greater.
We see the natural tendency to structure portfolios in this way in both
Charts
12.1
and
12.2;
in both cases, the risk proﬁles tend to align themselves with the curve rather than
with either the vertical or horizontal line. Thus, setting limits based on absolute risk would
take the qualitative intuition that currently drives decisions and make it quantitative.
It is worth mentioning here that the risk limits we have discussed are not meant to replace
existing limits to individual names. Limits based on the notion that there is a maximum
amount of exposure we desire to a given counterparty, regardless of this counterparty’s
credit standing, are certainly appropriate. Such limits may be thought of as conditional, in
that they reﬂect the amount we are willing to lose conditioned on a given counterparty’s
defaulting, and do not depend on the probability that the counterparty actually defaults.
The limits proposed in this section are meant to supplement, but not replace, these condi
tional limits.
12.2.2 Choice of risk measure
Given a choice of what type of limit to implement, it is necessary next to choose the spe
ciﬁc risk measure to be used. Essentially, there are two choices to make: ﬁrst, whether
to use a marginal or standalone statistic, and second, whether to use standard deviation,
percentile level, or another statistic.
The arguments for using marginal statistics have been made before. These statistics
allow the user to examine an exposure with regard to its effect on the actual portfolio, tak
Sec. 12.2 Credit risk limits 137
Part III: Applications
ing into accounts the effects of correlation and diversiﬁcation. Thus, marginal statistics
provide a better picture of the true concentration risk with respect to a given counterparty.
On the other hand, certain circumstances suggest the use of absolute risk measures for
limits. For instance, suppose a portfolio contains a large percentage of a bond issue of a
given name. Even if the name has a very low correlation with the remainder of the port
folio (meaning that the bond has low marginal risk), the position should be considered
risky because of the liquidity implications of holding a large portion of the issue. Thus,
it is important in this case to know the standalone riskiness of the position.
As to what statistic to use, we describe four statistics below, and discuss the applicability
of each to credit risk limits.
As always, the easiest statistic to compute is the
standard deviation
. However, as a mea
sure of credit risk, it has a number of deﬁciencies. First, the standard deviation is a “two
sided” measure, measuring the portfolio value’s likely ﬂuctuations to the upside or
downside of the mean. Since we are essentially concerned with only the downside, this
makes the standard deviation somewhat misleading. In addition, since distributions of
credit portfolios are mostly nonnormal, there is no way to infer concrete information
about the distribution from just the standard deviation.
We have also discussed the use of
percentile levels
at some length. The advantages of
this statistic are that it is easy to deﬁne and has a very concrete meaning. When we state
the ﬁrst percentile level of a portfolio, we know that this is precisely the level below
which we can expect losses only one percent of the time. There is a price for this preci
sion, however, as we cannot derive such a measure analytically, and must resort to simu
lations. Thus, our measure is subject to the random errors inherent in Monte Carlo
approaches.
Another statistic which is often mentioned for characterizing risk is
average shortfall
.
This statistic is deﬁned as the expected loss given that a loss occurs, or as the expected
loss given that losses exceed a given level. While this does give some intuition about a
portfolio’s riskiness, it does not have quite as concrete an interpretation as a percentile
level. For instance, if we say that given a loss of over $3mm occurs, we expect that loss
to be $6mm, we still do not have any notion of how likely a $6mm loss is. Along the
same lines, we might consider using the
expected excession of a percentile level
. For the
1
st
percentile level, this statistic is deﬁned as the expected loss given that the loss is more
extreme than the 1
st
percentile level. If this statistic were $12mm, then the interpretation
would be that in the worst 1 percent of all possible cases, we would expect our losses to
be $12mm. This is a very reasonable characterization of risk, but like percentile level
and average shortfall, requires a simulation approach.
When choosing a risk statistic, it is important to keep in mind its application. For limits,
and particularly for prioritization, it is not absolutely necessary that we be able to infer
great amounts of information about the portfolio distribution from the risk statistics that
we use. What is most important is that the risk estimates give us an idea of the
relative
riskiness of the various exposures in our portfolio. It is reasonable to claim that the stan
dard deviation does this. Thus, for the purpose of prioritization or limit setting, it would
be sensible to sacriﬁce the intuition we obtain from percentile levels or expected exces
sions if using the standard deviation provides us with signiﬁcant improvements in com
putational speed.
138 Chapter 12. Application of model outputs
CreditMetrics™—Technical Document
12.2.3 Policy issues
The fundamental point of a limit is that it triggers action. There can be many levels of
limits which we classify according to the severity of action taking in the case the limit is
exceeded.
For informational limits, an excession of the limit might require more indepth reporting,
additional authorization to increase exposure size, or even supplemental covenant pro
tection or collateral. The common thread is that exposures which exceed the limits are
permitted, but trigger other actions which are not normally necessary.
Alternatively, one might set hard limits, which would preclude any further exposure to an
individual name, industry, geographical region, or instrument type. In practice, one might
implement both types of limits – an informational limit at some low level of risk or expo
sure and a hard limit at a higher level. And these limits might even be based on two differ
ent risk measures – a marginal measure at one level and an absolute measure at the other.
The assumption for both types of limits above is that the limits are in place before the
exposures, and each exposure we add to the portfolio satisﬁes the limits. However, for
the aforementioned fallen angels, this will not be the case. These exposures satisfy the
risk limits when they are added to the portfolio, but subsequently exceed the limits due to
a change in market rates or to a credit rating downgrade. Excessions of this type are
essentially uncontrollable, although a portfolio manager might seek to reduce the risk in
these cases by curtailing additional exposure, reducing existing exposure, or hedging
with a credit derivative.
It is not uncommon to set limits at different levels of aggregation since different levels of
oversight may occur at higher and higher levels. For instance, there might be limits on
individual names, plus industry limits, plus sector limits, plus even an overall credit port
folio limit.
It should always be the case that a limit will be less than or equal to the sum of limits one
level lower in the hierarchy. Thus, the ﬁnancial sector limit should not be greater than
the sum of limits to industries underneath it such as banks, insurers, brokers, etc. This
will be true whether limits are set according to exposures (which can be aggregated by
simply summing them) or according to risk (which can be aggregated only after account
ing for diversiﬁcation).
12.3 Economic capital assessment
For the purposes of prioritization and limit setting, the subjects of the ﬁrst two sections,
we examined risk measures in order to evaluate and manage individual exposures. The
total risk of the portfolio might guide the limitsetting process, but it was the relative
riskiness of individual exposures which most concerned us.
In this section, we examine a different application of credit risk measures, that of assess
ing the capital which a ﬁrm puts at risk by holding a credit portfolio. We are no longer
trying to compare different exposures and decide which contribute most to the riskiness
of the portfolio, but rather are seeking to understand the risk of the entire portfolio with
regard to what this risk implies about the stability of our organization.
Sec. 12.3 Economic capital assessment 139
Part III: Applications
To consider risk in this way, we look at risk in terms of capital; but rather than consider
ing the standard regulator or accounting view of capital, we examine capital from a risk
management informational view. The general idea is that if a ﬁrm’s liabilities are con
stant, then it is taking risk by holding assets that are volatile, to the extent that the asset
volatility could result in such a drop in asset value that the ﬁrm is unable to meet its lia
bility obligations.
This risktaking capability is not unlimited, as there is a level beyond which no manager
would feel comfortable. For example, if a manager found that given his asset portfolio,
there was a ten percent chance for such a depreciation to occur in the next year as to cause
organizationwide insolvency, then he would likely seek to decrease the risk of the asset
portfolio. For a portfolio with a more reasonable level of risk, the manager cannot add
new exposures indiscriminately, since eventually the portfolio risk will surpass the “com
fort level.” Thus, each additional exposure utilizes some of a scarce resource, which
might be thought of as risktaking capability, or alternately, as economic capital.
To measure or assess the economic capital utilized by an asset portfolio, we may utilize
the distribution of future portfolio values which we describe elsewhere in this document.
This involves a choice, then, of what statistic to use to describe this distribution. The
choice is in some ways similar to the choice of risk statistic for limits which we dis
cussed in the previous section; however, the distinct use of risk measures here make the
decision different. For limits, we were concerned with individual exposures and relative
measures; for economic capital, we are interested in a portfolio measure and have more
need for a more concrete meaning for our risk estimate. These issues should become
clear as we consider the risk statistics below.
For limits we could argue that the standard deviation was an adequate statistic in that it
could capture the relative risks of various instruments. In this case, however, it is difﬁ
cult to argue that a standard deviation represents a good measure of capital since we are
unable to attach a concrete interpretation to this statistic. Yet this statistic is practical to
compute and for this reason alone may be the logical choice.
As an indicator of economic capital, a percentile level seems quite appropriate. Using
for example the 1
st
percentile level, we could deﬁne economic capital as the level of
losses on our portfolio which we are 99% certain (or in the words of Jacob Bernoulli,
“morally certain”
2
) that we will not experience in the next year. This ﬁts nicely with our
discussion of capital above. If it is our desire to be 99% certain of meeting our ﬁnancial
obligations in the next year, then we may think of the 1
st
percentile level as the risk we
are taking, or as the economic capital which we are allocating to our asset portfolio. If
this level ever reaches the point at which such a loss will prevent us from meeting obliga
tions, then we will have surpassed the maximum amount of economic capital we are will
ing to utilize.
As with limits, we may consider average shortfall as a potential statistic. Yet just as in
the case of limits, it is difﬁcult to consider an expected shortfall of $6mm as a usage of
capital since we do not know how likely such a loss actually is. On the other hand, the
expected excession of a percentile level does seem worth consideration. Recall that if
this statistic were $12mm at the ﬁrst percentile level, then the interpretation would be
2
As quoted in Bernstein [96].
140 Chapter 12. Application of model outputs
CreditMetrics™—Technical Document
that in the worst 1 percent of all possible cases, we would expect our losses to be
$12mm. So like the percentile level above, this seems to coincide with our notion of
economic capital, and thus seems a very appropriate measure.
All of the above measures of economic capital differ fundamentally from the capital
measures mandated for bank regulation by the Bank for International Settlements (BIS).
For a portfolio of positions not considered to be trading positions, the BIS riskbased
capital accord of 1988 requires capital that is a simple summation of the capital required
on each of the portfolio's individual transactions, where each transaction's capital
requirement depends on a broad categorization (rather than the credit quality) of the obli
gor; on the transaction's exposure type (e.g., drawn loans versus undrawn commitments);
and, for offbalancesheet exposures, on whether the transaction's maturity is under one
year or over one year. The weaknesses of this riskbased structure – such as its onesize
ﬁtsall risk weight for all corporate loans and its inability to distinguish diversiﬁed and
undiversiﬁed portfolios – are increasingly apparent to regulators and market participants,
with particular concern paid to the uneconomic incentives created by the regulatory
regime and the inability of regulatory capital adequacy ratios to accurately portray actual
bank risk levels. In response to these concerns, bank regulators are increasingly looking
for insights in internal credit risk models that generate expected losses and a probability
distribution of unexpected losses.
3
12.4 Summary
In summary, the CreditMetrics methodology gives the user a variety of options to use for
measuring economic capital which may in turn lead to further uses of CreditMetrics. We
brieﬂy touch on three applications of an economic capital measure:
exposure reduction
,
limit setting
, and
performance evaluation
.
An assessment of economic capital may guide the user to actions which will alter the
characteristics of his portfolio. For example, if the use of economic capital is too high, it
will be necessary to take actions on one or more exposures, possibly by prohibiting addi
tional exposure, or else by reducing existing exposures by unwinding a position or hedg
ing with a credit derivative. How to choose which exposures to treat could then be
guided by the discussions in
Section 12.1
.
On the other hand, one might wish to use the measure of economic capital in order to aid
the limitsetting process, assuring that if individual or industry level exposures are within
the limits, then the level of capital utilization will be at an acceptable level.
A third use is performance evaluation. The traditional practice has been to evaluate port
folio managers based on return, leading to an incentive structure which encourages these
managers to take on lower rated exposures in order to boost performance. Adding a
measure of economic capital utilization allows for a more comprehensive measure of
performance; when managers’ returns are paired with such a risk measure, it can be seen
which managers make the most efﬁcient use of the ﬁrm’s economic capital. Examining
performance in this way retains the incentive to seek high returns, but penalizes for tak
ing undue risks to obtain these returns.
3
See Remarks by Alan Greenspan, Board of Governors of the Federal Reserve System, before the 32
nd
Annual
Conference on Bank Structure and Competition, FRB of Chicago, May 2, 1996.
Sec. 12.4 Summary 141
Part III: Applications
By examining rates of return on economic capital and setting targets for these returns, a
manager or ﬁrm goes a step beyond the traditional practice of requiring one rate of return
on its most creditworthy assets and a higher rate on more speculative ones; the new
approach is to consider a hurdle rate of return on risk, which is more clear and more uni
form than the traditional practice. Identifying portfolios or businesses that achieve higher
returns on economic capital essentially tells a manager which areas are providing the
most value to the ﬁrm. And just as it is possible to allocate any other type of capital,
areas where the return on risk is higher may be allocated more economic capital, or more
risktaking ability. By focusing capital on the most efﬁcient parts of a ﬁrm or portfolio,
proﬁts are maximized, but within transparent, responsible risk guidelines.
142 Chapter 12. Application of model outputs
CreditMetrics™—Technical Document
143
Appendices
144
CreditMetrics™—Technical Document
145
Appendices
Appendices
In CreditMetrics we use certain general statistical formulas, data, and indices in several
different capacities. We have chosen to address each of them in detail here in an appen
dix so that we may give them the depth they deserve without cluttering the main body of
this
Technical Document
. These appendices include:
Appendix A: Analytic standard deviation calculation.
A generalization of the methods presented in
Chapter 9
to compute the standard
deviation for a portfolio of arbitrary size.
Appendix B: Precision of simulationbased estimates.
Techniques to assess the precision of portfolio statistics obtained through simulation.
Appendix C: Derivation of the product of N random variables.
Used to: (i) combine the uncertainty of spread and exposure risk and (ii) for the deri
vation of risk across mutually exclusive outcomes.
Appendix D: Derivation of risk across mutually exclusive outcomes.
Used for both: the value variance of a position across
N
states and the covariance
between positions across
N
states.
Appendix E: Derivation of the correlation of two binomials.
Used to link correlation between ﬁrms’ value to their default correlations.
Appendix F: Inferring default correlations from default volatilities.
Used as alternative method to estimate default correlations which corroborates our
equity correlation approach.
Appendix G: International bankruptcy code summary.
Contains this information in tabular format.
Appendix H: Model inputs.
Describes the CreditMetrics data ﬁles and required inputs.
Appendix I: Indices used for asset correlations.
Contains this information in tabular format.
146
CreditMetrics™—Technical Document
147
CreditMetrics™—Technical Document
Appendix A. Analytic standard deviation calculation
In
Chapter 9
, we presented the calculation of the standard deviation for an example three
asset portfolio, and stated that the generalization of this calculation to a portfolio of arbi
trary size was straightforward. In this appendix, we present this generalization in detail.
Consider a portfolio of
n
assets. Denote the value of these assets at the end of the hori
zon by
V
1
,
V
2
, ...,
V
n
; let these values’ means be
µ
1
,
µ
2
,...,
µ
n
and their variances be
, ,..., . The calculation of these individual means and variances is
detailed in
Chapter 2
.
The value of the portfolio at the end of the forecast horizon is just
V
1
+V
2
+...+V
n
, and the
mean value is
µ
p
=
µ
1
+
µ
2
+...+
µ
n
. To compute the portfolio standard deviation (
σ
p
), we
may use the standard formula:
[A.1]
Alternatively, we may relate the covariance terms to the variances of pairs of assets,
[A.2] ,
and using this fact, express the portfolio standard deviation in terms of the standard devi
ations of subportfolios containing two assets:
[A.3]
As in
Chapter 9
, we see that the portfolio standard deviation depends only on the vari
ances for pairs of assets and the variances of individual assets. This makes the computa
tion of the portfolio standard deviation straightforward. We begin by computing the
variances of each individual asset; we then identify all pairs of assets among the
n
assets
in the portfolio
1
and compute the variances for each of these pairs using the methods in
Chapter 3
; ﬁnally, we apply Eq. [A.3].
1
There will be pairs.
σ
2
V
1
( ) σ
2
V
2
( ) σ
2
V
n
( )
σ
p
2
σ
2
V
i
( )
i 1 =
n
∑
2 COV V
i
V
j
, ( )
˙
.
j i 1 + =
n
∑
i 1 =
n 1 –
∑
⋅ + =
σ
2
V
i
V
j
+ ( ) σ
2
V
i
( ) 2 COV V
i
V
j
, ( ) σ
2
V
j
( ) + ⋅ + =
σ
p
2
σ
2
V
i
V
j
+ ( )
j i 1 + =
n
∑
i 1 =
n 1 –
∑
n 2 – ( ) – σ
2
V
i
( ).
i 1 =
n
∑
⋅ =
n n 1 – ( ) 2 ⁄ ⋅
148
CreditMetrics™—Technical Document
149
Appendices
Appendix B. Precision of simulationbased estimates
In
Chapter 10
, we presented a methodology to compute portfolio statistics using Monte
Carlo simulation and mentioned that statistics which are estimated in this way are subject
to random errors. In this appendix, we discuss how we may quantify the sizes of these
errors, and thus discover how conﬁdent we may be of the risk estimates we compute.
We devote one subsection each to the treatment of the sample mean, sample standard
deviation, and sample percentile levels.
Throughout this section, we will use
V
(1)
,
V
(2)
,
V
(3)
,...,V
(N)
to indicate the portfolio values
across scenarios and
V
[1]
,
V
[2]
,
V
[3]
,...,V
[N]
to indicate the same values sorted into ascend
ing order (so that, for example
V
[2]
is the second smallest value). Further, let
µ
n
denote
the sample mean and
σ
n
the sample standard deviation of the ﬁrst
n
scenarios.
B.1 Sample mean
Quantifying the error about our estimate of the mean portfolio value is straightforward.
For large
n
,
µ
n
will be approximately normally distributed with standard deviation
. Thus, after generating
n
scenarios, we may say that we are 68%
2
conﬁdent that
the true mean portfolio value lies between and and 90% conﬁ
dent the true mean lies between and . Note that
these bands will tighten as
n
increases.
B.2 Sample standard deviation
Our conﬁdence in the estimate
σ
n
is more difﬁcult to quantify since the distribution of
the estimate is less well approximated by a normal distribution, and the standard devia
tion of the estimate is much harder to estimate.
The simplest approach here is to break the full set of scenarios into several subsets, com
pute the sample standard deviation for each subset, and examine how much ﬂuctuation
there is in these estimates. For example, if we have generated 20,000 portfolio scenarios,
then we might divide these scenarios into ﬁfty separate groups of 400. We could then
compute the sample standard deviation within each group, obtaining ﬁfty different esti
mates
σ
(1)
,
σ
(2)
,...,
σ
(50)
. The sample standard deviation of these estimates, which we
denote by
s
, is then an estimate for the standard error of
σ
400
. In order to extrapolate to
an estimate for the standard error of σ
20000
, we assume that the same scaling holds as
with the sample mean, and take . Then we can say that we are approximately 90%
conﬁdent that the true value of our portfolio standard deviation lies between
and
3
. This procedure is commonly referred
to as “jackkniﬁng.”
For the sample mean and standard deviation, our approach to assessing precision was the
same. Motivated by the fact that the estimates we compute are sums over a large number
2
Since the probability that a normally distributed random variable falls within one standard deviation of its mean is
68%.
3
This methodology is somewhat sensitive to the choice of how many separate groups to divide the sample into. We
choose 50 here, but in practice suggest that the user experiment with various numbers in order to get a feel for the
sensitivity of the conﬁdence estimates to this choice.
σ
n
n ⁄
µ
n
σ
n
n ⁄ – µ
n
σ
n
n ⁄ +
µ
n
1.65 σ
n
n ⁄ ⋅ ( ) – µ
n
1.65 σ ⋅
n
n ⁄ +
s 50 ⁄
µ
2000
1.65 s 50 ⁄ ⋅ ( ) – µ
2000
1.65 s 50 ⁄ ⋅ ( ) +
150 Appendix B. Precision of simulationbased estimates
CreditMetrics™—Technical Document
of independent trials, we approximated the distributions of the estimates as normal. The
rest of the analysis then focused on computing the standard errors for the estimates.
Moreover, in some sense, the assessment of precision for estimates of these two statistics
is somewhat redundant, as it is possible to obtain exact values in both cases.
In the next section, we treat estimates of percentile levels, for which neither of these
points applies. Estimates are not just sums over the scenarios, and thus we cannot expect
the distributions of the estimates to be normal; further, we have no way of computing
percentile levels directly, and thus are much more concerned with the precision of our
estimates.
B.3 Sample percentile levels
As an example, say we are trying to estimate the 5
th
percentile level, and let θ
5
be the
true value of this level. Each scenario which we generate then (by deﬁnition) has a 5%
chance of producing a portfolio value less than θ
5
. Now consider 1000 independent sce
narios, and let N
5
be the number of these scenarios which fall below θ
5
. Note that N
5
fol
lows the binomial distribution. Clearly, the expected value of N
5
is ,
while the standard deviation is . For this many trials,
it is reasonable to approximate the distribution of N
5
by the normal. Thus, we estimate
that there is a 68% chance that N
5
will be between 506.9=43.1 and 50+6.9=56.9, and a
slightly higher chance that N
5
will be between 43 and 57. Further, there is a 90% chance
that N
5
falls between and .
At this point we have characterized N
5
. This may not seem particularly useful, however,
since N
5
is not actually observable. In other words, since we do not actually know the
level θ
5
(this is what we are trying to estimate), we have no way of knowing how many
of our scenarios fell below θ
5
. We assert that it is not necessary to know N
5
exactly,
since we can gain a large amount of information from its distribution.
Observe that if N
5
is greater than or equal to 43, then at least 43 of our scenarios are less
than θ
5
. This implies that θ
5
is at least as large as the 43rd smallest of our portfolio val
ues. (Recall that in our notation, this scenario is denoted by V
[43]
.) On the other hand, if
N
5
is less than or equal to 57, then it must be true that θ
5
is no larger than the 57th small
est of the portfolio values (that is, V
[57]
). Thus, we have argued that the event
[B.1]
is exactly the same as the event
[B.2]
Now since these two events are the same, they must have the same probability, and thus
[B.3]
and so we have a conﬁdence bound for our estimate of θ
5
. To recap, using 1000 scenar
ios, we estimate the 5
th
percentile portfolio value by the 50
th
smallest scenario, and state
1000 5% = 50 ⋅
1000 5% 100% 5% – ( ) ⋅ ⋅ 6.9 =
50 1.65 6.9 ⋅ – 38.6 = 50 1.65 6.9 ⋅ + 61.2 =
43 N
5
57 ≤ ≤
V
43 [ ]
θ
5
V
57 [ ]
. < <
Pr V
45 [ ]
θ
5
V
57 [ ]
< < { } Pr 43 N
5
57 ≤ ≤ { } = 68% =
Appendix B. Precision of simulationbased estimates 151
that we are 68% conﬁdent that the true percentile lies somewhere between the 43
rd
and
57
th
smallest scenarios.
In general, if we wish to estimate the p
th
percentile using N scenarios, we ﬁrst consider
the number of scenarios that fall below the true value of this percentile. We characterize
this number via the following:
[B.4]
where α depends on the level of conﬁdence which we desire. (That is, if we desire 68%
conﬁdence, then α=1, if we desire 90%, then α=1.65, etc.) If either l or m are not whole
numbers, we round them downwards, while if u is not a whole number, we round
upwards. We then estimate our percentile by V
[m]
and state with our desired level of
conﬁdence that the true percentile lies between V
[l]
and V
[u]
.
For further discussion of these methods, see DeGroot [86], p. 563. Note that the only
assumption we make in this analysis is that the binomial distribution is well approxi
mated by the normal. In general, this will be the case as long as the expected number of
scenarios falling below the desired percentile (that is, N·p) is at least 20 or so. In cases
where this approximation is not accurate, we may take the same approach as in this sec
tion, but characterize the distribution precisely rather than using the approximation. The
result will be similar, in that we will obtain conﬁdence bands on the number of scenarios
falling below the threshold, and then proceed to infer conﬁdence intervals on the esti
mated percentile.
lower bound: l N p α N p 1 p – ( ) ⋅ ⋅ ⋅ – ⋅ =
mean: m N p s N p 1 p – ( ) ⋅ ⋅ = , ⋅ =
and
upper bound: u N p α N p 1 p – ( ) ⋅ ⋅ ⋅ + ⋅ =
152
CreditMetrics™—Technical Document
153
Appendices
Appendix C. Derivation of the product of N random variables
First we examine in detail the volatility of the product of two random variables. Let X
and Y be any independent and uncorrelated distributions deﬁned as follows:
[C.1]
where all distributions, Z, are independent and standardized but can otherwise have any
desired shape: normal, highly skewed, binomial, etc.
[C.2]
First, we will multiply out x · y.
[C.3]
Since the expected value of Z is zero, the E( )’s simplify greatly.
[C.4]
Now σ
X·Y
is only a matter of algebra.
[C.5]
By induction, we can we can extend the volatility estimation for the product of arbitrarily
many independent events. First, the expectation of this product is simply the product of
its expectations:
[C.6]
The variance of the product of N distributions will in general have, , terms. For
the case of the product of three distributions, the result is:
X µ
x
∼ σ
x
Z
x
⋅ + Y µ
y
∼ σ
y
Z
y
where denotes distributed as ∼ ( ) ⋅ +
σ
X Y ⋅
2
E X
2
Y
2
⋅ ( ) E X Y ⋅ ( )
2
(Textbook formula) – =
X Y ⋅ µ
x
µ
y
µ
x
σ
y
Z
Y
µ
y
σ
x
Z
X
σ
x
Z
X
σ
y
Z
Y
+ + + =
E X Y ⋅ ( ) µ
x
µ
y
=
E X Y ⋅ ( )
2
µ
x
2
µ
y
2
(Since: E(Z) = 0) =
E X
2
Y
2
⋅ ( ) µ
x
2
µ
y
2
µ
x
2
σ
y
2
µ
y
2
σ
x
2
σ
x
2
σ
y
2
(Since: E(Z)
2
+ + + 1) = =
σ
X Y ⋅
2
µ
x
2
µ
y
2
µ
x
2
σ
y
2
µ
y
2
σ
x
2
σ
x
2
σ
y
2
+ + + ( ) = µ
x
2
µ
y
2
( ) –
µ
x
2
σ
y
2
µ
y
2
σ
x
2
σ
x
2
σ
y
2
+ + =
σ
X Y ⋅
2
µ
x
2
σ
y
2
µ
y
2
σ
x
2
σ
x
2
σ
y
2
+ + =
E Φ
i
i
N
∏
¸ ,
¸ _
µ
i
where all Φ
i
µ
i
σ
i
Z
i
⋅ + ∼
i
N
∏
=
and all Z
i
are standardized (0,1)
2
N
1 –
154 Appendix C. Derivation of the product of N random variables
CreditMetrics™—Technical Document
[C.7]
In general, the pattern continues and can be denoted as follows for N distributions. In
this notation, j and m denote sets whose elements comprise the product sums:
[C.8]
VAR Φ
X
Φ
Y
Φ
Z
⋅ ⋅ ( )
µ
x
2
µ
y
2
σ
z
2
µ
x
2
σ
y
2
σ
z
2
+ +
µ
x
2
σ
y
2
µ
z
2
σ
x
2
µ
y
2
σ
z
2
σ
x
2
σ
y
2
σ
z
2
+ + +
σ
x
2
µ
y
2
µ
z
2
σ
x
2
σ
y
2
µ
z
2
+ +
¸ ,
¸ _
=
VAR Φ
i
i
N
∏
¸ ,
¸ _
σ
j
2
µ
m
2
m S N ( ) j – =
∏
⋅
¸ ,
¸ _
J s N k , ( ) ∈
∏
µ
i
2
i
N
∏
–
k 1 =
N
∑
=
where the sets S N ( ) 1 2 3 … N , , , , { } =
and s N k , ( ) j
i
…j
k
1 j
1
… j
k
N k N ≤ , ≤ < < ≤ , { }
˙
. =
155
Appendices
Appendix D. Derivation of risk across mutually exclusive outcomes
Imagine that there were two alternative outcomes (subscripts
1
and
2
) that might occur in
the event of default with probabilities of
p
1
and
p
2
which sum to the total probability of
default. For completeness, subscript
ω
is the case of no default. Each of these three
cases has some distribution of losses denoted,
Φ
i
(
x
), with statistics,
µ
i
and
σ
i
.
Deﬁnitions:
and .
[D.1] Expected Total Loss
[D.2] Variance of Total Loss
The above derivation requires a substitution for an integral that merits further discussion.
The problem of multiplying a random variable by itself was addressed in the prior appen
dix note (see
Appendix C)
. If the two are the same distribution, then the correlation is
simply 1.0.
[D.3] Mean of Product of Two Random Variables
1 p =
1
p
2
p
ω
+ + Φ
T
x ( ) p
1
Φ
1
x ( ) p
2
Φ
2
x ( ) p
ω
Φ
ω
x ( ) + + =
µ
T
xΦ
T
x ( ) x d
∫
=
x p
1
Φ
1
x ( ) p
2
Φ
2
x ( ) p
ω
Φ
ω
x ( ) + + ( ) x d
∫
=
p
1
µ
1
p
2
µ
2
p
ω
µ
ω
+ + =
σ
T
2
= x µ
T
– ( )
2
Φ
T
x ( ) x d
∫
x
2
2xµ
T
– µ
T
2
+ ( ) p
1
Φ
1
x ( ) p
2
Φ
2
x ( ) p
ω
Φ
ω
x ( ) + + ( ) x d
∫
=
p
1
x
2
Φ
1
x ( )
∫
p
2
x
2
Φ
2
x ( )
∫
p
ω
x
2
Φ
ω
x ( )
∫
+ +
These simplify
2µ
T
p
1
µ
1
p
2
µ
2
p
ω
µ
ϖ
+ + ( ) –
Note that this equals µ
T
see above
µ
T
2
p
1
p
2
p
ω
+ + ( ) +
Note that this sums to 1
¸ ,
¸ _
=
p
1
µ
1
2
σ
1
2
+ ( ) p
2
µ
2
2
σ
2
2
+ ( ) p
ω
µ
ω
2
σ
ω
2
+ ( ) + +
2µ
T
2
–
µ
T
2
+
¸ ,
¸ _
=
p
1
µ
1
2
σ
1
2
+ ( ) p
2
µ
2
2
σ
2
2
+ ( ) p
ω
µ
ω
2
σ
ω
2
+ ( ) µ
T
2
– + + =
¹ ¹ ¹ ' ¹ ¹ ¹ ¹ ¹ ¹ ' ¹ ¹ ¹ ¹ ¹ ¹ ' ¹ ¹ ¹
¹ ¹ ¹ ¹ ¹ ' ¹ ¹ ¹ ¹ ¹
¹ ¹ ¹ ¹ ' ¹ ¹ ¹ ¹
µ
i j ⋅ ( )
µ
i
µ
j
ρσ
i
σ
j
See prior appendix note. + =
µ
i
2
σ
i
2
Since i = j and ρ + 1.0 = =
x
2
Φ
i
x ( ) x Substitution made above. d
∫
=
156 Appendix D. Derivation of risk across mutually exclusive outcomes
CreditMetrics™—Technical Document
For completeness, we have included terms describing the losses in the case of no default:
µ
ω
and σ
ω
. But these are both zero since there will be no losses in the case of no default.
Thus the overall total mean and standard deviation of losses in this process simpliﬁes as
follows:
[D.4] σ
T
p
i
µ
i
2
σ
i
2
+ ( )
i 1 =
S
∑
µ
T
2
– where µ
T
= p
i
µ
i
i 1 =
S
∑
=
157
Appendices
Appendix E. Derivation of the correlation of two binomials
The traditional textbook formula for covariance is shown below.
The expected probabilities, p’s, of the two binomials, x and y, are termed µ
x
and µ
y
respectively. Normally all the n observations would be equally weighted (
1
/
n
), but here
the probability weights W
i
will equal the likelihood of each possible outcome. For the
joint occurrence of two binomials, there will be exactly four possible outcomes. We can
simply list them explicitly. The probability weights W
i
are easily calculated for the case
of independence, but we will leave them as variables to allow for any degree of possible
correlation. As shown below, defaults will have value 1 and nondefaults will have
value 0.
[E.1]
The difﬁcult problem in deﬁning the probability weights W’s is knowing the correlated
joint probability of default (cell #1 above). We will label this joint probability as α.
Multiplying and simplifying the resulting formula, see below, yields an intuitive result
for our covariance. If the joint default probability, α, is greater than the independent
probability, (that is µ
x
times µ
y
), then the covariance is positive; otherwise it is negative.
Obligor Y Obligor X
Default No Default Default No Default
1: X& Y default 3: Only X defaults 1: X& Y default 2: Only Y defaults
2: Only Y defaults 4: Neither defaults 3: Only X defaults 4: Neither defaults
cov
x y ,
W
i
x
i
µ
x
– ( ) y
i
µ
y
– ( )
i 1 =
n
∑
=
cov
x y ,
W
1
1 µ
x
– ( ) 1 µ
y
– ( )
W
2
0 µ
x
– ( ) 1 µ
y
– ( ) +
W
3
1 µ
x
– ( ) 0 µ
y
– ( ) +
W
4
0 µ
x
– ( ) 0 µ
y
– ( ) +
=
158 Appendix E. Derivation of the correlation of two binomials
CreditMetrics™—Technical Document
[E.2]
Now that we have derived the covariance as a function of the joint default probability, α,
we can redeﬁne α in terms of the correlation of our two binomials. Again, we can start
with a textbook formula for the covariance:
[E.3] thus
Interestingly, the above deﬁnition of α and ρ is identical the formula for the mean of the
product of two correlated random variables as shown above (see Appendix A). Impor
tantly, this correlation ρ
xy
is the resulting correlation of the joint binomials
4
. It does not
represent some underlying ﬁrmasset correlation that (via a bivariate normal assumption)
might lead to correlated binomials. The σ’s here are the usual binomial standard devia
tions, . This formula for ρ
xy
implies that there are bounds on ρ
xy
since α is at
least max(0, µ
x
+µ
y
1) and at most min (µ
x
, µ
y
). Thus:
[E.4]
4
Other researchers have used this same binomial correlation, see Lucas [95a].
cov
x y ,
W
1
1 µ
x
– ( ) 1 µ
y
– ( )
+W
2
0 µ
x
– ( ) 1 µ
y
– ( )
+W
3
1 µ
x
– ( ) 0 µ
y
– ( )
+W
4
0 µ
x
– ( ) 0 µ
y
– ( )
¸ ,
¸ _
α [ ] 1 µ
x
– ( ) 1 µ
y
– ( )
+ µ
y
α – [ ] 0 µ
x
– ( ) 1 µ
y
– ( )
+ µ
x
α – [ ] 1 µ
x
– ( ) 0 µ
y
– ( )
+ 1 µ
x
– µ
y
– α + [ ] 0 µ
x
– ( ) 0 µ
y
– ( )
¸ ,
¸ _
α αµ
y
– αµ
x
– αµ
x
µ
y
+
µ –
x
µ
y
µ
x
µ
y
2
αµ
x
αµ
x
µ
y
– + +
µ –
x
µ
y
µ
x
2
µ
y
αµ
y
αµ
x
µ
y
– + +
+µ
x
µ
y
µ
x
2
µ
y
– µ
x
µ
y
2
– αµ
x
µ
y
+
¸ ,
¸ _
α µ
x
µ
y
– =
=
=
=
cov
x y ,
ρ
x y ,
σ
x
σ
y
=
so
α µ
x
µ
y
– ρ
x y ,
σ
x
σ
y
=
α µ
x
µ
u
= ρ
x y ,
σ
x
σ
y
+
and
ρ
x y ,
α µ
x
µ
y
– ( ) σ
x
σ
y
⁄ =
µ 1 µ – ( )
max 0 µ
x
µ
y
1 – + , ( ) µ
x
µ
y
– ( )
σ
x
σ
y

ρ
x y ,
min µ
x
µ
y
, ( ) µ
x
µ
y
– ( )
σ
x
σ
y

≤ ≤
159
Appendices
Appendix F. Inferring default correlations from default volatilities
For N ﬁrms in a grouping with identical default rate (i.e., within a single credit rating cat
egory), let be a random variable which is either 1 or 0 according to each ﬁrm’s
default event realization with mean default rate, , and binomial default standard
deviation, , deﬁned as follows:
[F.1]
Let D represent the number of defaults, . So the variance of D is as follows:
[F.2]
Rather than each , we are interested in the average correlation, , and deﬁne this
as follows
[F.3]
and so we can now deﬁne
[F.4]
Across many ﬁrms we can observe the volatility of defaults, , thus:
X
1
µ X
1
( )
σ X
1
( )
X
i
1 if company i defaults
0 otherwise
¹
'
¹
=
¸ ,
¸ _
µ
CrRt
µ X
i
( )
1
N
 X
i
i
N
∑
= =
σ X
i
( ) µ
CrRt
1 µ
CrRt
– ( ) =
D X
i
i
N
∑
=
VAR D ( ) ρ
ij
σ X
i
( )σ X
j
( )
j
N
∑
i
N
∑
=
ρ
ij
σ X
i
( )
2
j
N
∑
i
N
∑
=
ρ
ij
µ
CrRt
µ
CrRt
2
– ( )
j
N
∑
i
N
∑
=
µ
CrRt
µ
CrRt
2
– ( ) N ρ
ij
j i ≤
N
∑
i
N
∑
+ =
Since all i and j have the same default rate.
ρ
ij
ρ
CrRt
ρ
CrRt
ρ
ij
j i ≤
N
∑
i
N
∑
N
2
N – ( ) ⁄ =
VAR D ( ) µ
CrRt
µ
CrRt
2
– ( ) N N
2
N – ( )ρ
CrRt
+ [ ] = .
σ
CrRt
2
VAR D N ⁄ ( ) =
160 Appendix F. Inferring default correlations from default volatilities
CreditMetrics™—Technical Document
[F.5]
This can be applied with good result in a simpliﬁed form if N is “large”:
[F.6]
The estimate of 8,5000 ﬁrmyears above stems from Moody’s reporting of 120 ﬁrms
being rated Ba one calendar year prior to default (8,500 ≅ 120/1.42%), see Carty & Lie
berman [96a].
σ
CrRt
2
VAR
D
N

¸ ,
¸ _
VAR D ( )
N
2

µ
CrRt
µ
CrRt
2
– ( )
1 N 1 – ( )ρ
CrRt
+
N

ρ
CrRt
N
σ
CrRt
2
µ
CrRt
µ
CrRt
2
–

¸ ,
¸ _
1 –
N 1 –

= ∴ ⋅
= =
=
ρ
CrRt
N
σ
CrRt
2
µ
CrRt
µ
CrRt
2
–

¸ ,
¸ _
1 –
N 1 –

8 500
1.4%
Ba
2
1.42%
Ba
1.42%
Ba
2
–

¸ ,
¸ _
1 – ,
8 500 1 – ,

1.3886% = =
σ
CrRt
2
µ
CrRt
µ
CrRt
2
–

1.4%
Ba
2
1.42%
Ba
1.42%
Ba
2
–
 1.4002% = = =
=
161
Appendices
Appendix G. International bankruptcy code summary
The practical result of the seniority standing of debt will vary across countries according
to local bankruptcy law. Of course, this will affect the likely recovery rate distributions.
Major differences will apply to secured versus unsecured debt. The following summary
table is reproduced from Rajan & Zingales [95] – who in turn reference Keiser [94], Lo
Pucki & Triantis [94], and White [93].
Table G.1
Summary of international bankruptcy codes
Country
Forms
of Liquidation
Forms of
Reorganization
Management Control
in Bankruptcy Automatic Stay
Rights of
Secured Creditors
United
States
Chapter 7: Can be voluntary
(management ﬁles) or invol
untary (creditors ﬁle).
Chapter 11: Can be voluntary
(management ﬁles) or invol
untary (creditors ﬁle).
Trustee appointed in Chapter
7. Management stays in con
trol in Chapter 11.
Automatic stay on any
attempts to collect debt
once ﬁling takes place.
Secured creditors get highest
priority in any attempts to col
lect payment are also stayed
unless court or trustee approves
Japan Court Supervised Liquida
tion (Hasan) and Special
Liquidation (Tokubetsu Sei
san). The latter is less costly
and a broader set of ﬁrms are
eligible to ﬁle.
Composition (Wagiho), Cor
porate Arrangement (Kaisha
Seiri) and Reorganization
(Kaisha Koseiho). The list
in order of increasing eligibil
ity. Only debtors ﬁle.
Third party is appointed ex
cept in composition and cor
porate arrangement.
All creditors are stayed
except in court super
vised liquidation and
composition where
only unsecured credi
tors are stayed.
Secured Creditors have highest
priority and greater voting
rights in renegotiation. Howev
er, can be subject on the petition
that is ﬁled.
Germany Liquidation (Konkursord
nung) can be requested by
creditors or debtor. Manage
ment required to ﬁle as soon
as it learns it is insolvent.
Composition (Vergleich or
Zwangvergleich) can be ﬁled
for only by debtor.
Receiver appointed to man
age ﬁrm.
Only unsecured credi
tors are stayed.
Secured creditors can recover
their claims even after a bank
ruptcy ﬁling. No stay for
secured creditors.
France Liquidation (Liquidation Ju
dicaire)
Negotiated Settlement
(Reglement Amiable) where a
court appointed conciliator
attempts a settlement with
creditors and Judicial Ar
rangement (Redressement
Judiciare).
Debtor loses control in liqui
dation. Debtor remains in
control otherwise but submits
to court appointed adminis
trator’s decisions in a judicial
arrangement.
Stay on all creditors in
judicial arrangement.
Secured creditors may lose sta
tus if court determines the
security is necessary for contin
uation of the business, or if the
securing asset is sold as part of
settlement.
Italy Bankruptcy (Fallimento) Preventive Composition
(Concordato Preventino)
Debtor is removed from con
trol over the ﬁrm.
Stay on all creditors. Secured creditors stayed in
bankruptcy, through composi
tion allowed only if enough
value exists to pay secured cred
itors in full and 40% of
unsecured creditor claims.
Secured creditors follow ad
ministrative claims in priority.
United
Kingdom
Members’ voluntary wind
ing up, Creditors’ voluntary
winding up, Compulsory
winding up.
Administration, Administra
tive Receivership (usually
ends in sale of business), and
Voluntary Arrangement.
Debtor is removed from con
trol except in members’ vol
untary winding up.
Stay on all creditors in
administration, on un
secured only in liquida
tion, and no stay in a
voluntary arrangement
until a proposal is ap
proved.
Secured creditor may prevent
administration order by ap
pointing his own receiver. A
creditor with a ﬁxed or ﬂoating
charge can appoint an adminis
trative receiver to realize the
security and pay the creditor.
Canada Liquidation proceedings
much like Chapter 7 in the
United States
Firms can ﬁle for automatic
stay under the Companies
Creditors Arrangement Act
or the Bankruptcy and Insol
vency Act.
Firm is in control in reorgani
zations while trustee is ap
pointed for liquidation.
Trustee may be appointed to
oversee management in some
reorganizations at the discre
tion of the court.
Stay on all creditors in
reorganization.
Secured creditors have to give
10 days notice to debtor of in
tent to repossess collateral.
Repossession even close to
bankruptcy ﬁling is permitted,
but stayed after ﬁling.
162 Appendix G. International bankruptcy code summary
CreditMetrics™—Technical Document
163
Appendices
Appendix H. Model inputs
Available for free download from the Internet http://jpmorgan.com/ is a data set of all
the elements described in this technical document and necessary to implement the
CreditMetrics methodology. Here, we brieﬂy list what is provided and the format in
which it is available.
CreditMetrics data ﬁles include:
• country/industry index volatilities and correlations,
• yield curves,
• spread curves, and
• transition matrices.
H.1 Common CreditMetrics data format characteristics
In general, CreditMetrics data ﬁles are text (ASCII) ﬁles which use tab characters
(ASCII code 9) as column delimiters, and carriage returns/line feeds as row delimiters.
Every CreditMetrics data ﬁle begins with a header, for example:
The header is followed by a row of column headers, followed by the data.
Cells in the data rows must contain data. If the value is unavailable or not applicable, the
cell should contain the keyword NULL.
H.2 Country/industry index volatilities and correlations
This ﬁle is named indxvcor.cdf. The data represent the weekly volatilities and correla
tions discussed in Chapter 8.
CDFVersion v1.0
Date 02/15/1997
DataType CountryIndustryVolCorrs
CDFVersion v1.0
Date 02/15/1997
DataType CountryIndustryVolCorrs
IndexName Volatility
MSCI Australia Index (.CIAU) 0.0171 1.0000 0.6840 0.6911 0.7343 0.6377
ASX Banks & Finance Index (.ABII) 0.0219 0.6840 1.0000 0.4360 0.4580 0.4436
ASX Media Index (.AMEI) 0.0257 0.6911 0.4360 1.0000 0.5528 0.3525
164 Appendix H. Model inputs
CreditMetrics™—Technical Document
H.3 Yield curves
This ﬁle is named
yldcrv.cdf
. A yield curve is deﬁned by currency . Allowable curren
cies are the standard threeletter ISO currency codes (e.g., CHF, DEM, GBP, JPY, USD).
H.4 Spread curves
Bridge will be the initial data provider for credit spreads. Their contact number is
(1800) 828  8010.
Bridge credit spread data is derived through a compilation of information provided by
major dealers including Citibank, CS First Boston, Goldman Sachs, Liberty Brokerage,
Lehman Brothers, Morgan Stanley, Salomon Brothers and J.P. Morgan. A team of evalu
ators reviews the contributed information on a daily bvasis to ensure accuracy and con
sistency.
This ﬁle is named
sprdcrv.cdf
. A spread curve is deﬁned by a combination of rating sys
tem, rating, and a yield curve (a yield curve being deﬁned as a combination of currency
and asset type). Allowable currencies are the standard 3letter ISO currency codes (e.g.
CHF, DEM, GBP, JPY, USD). Allowable asset types are BOND, LOAN, COMMIT
MENT, RECEIVABLE, and MDI.
Initial data is available only for USD and BOND
H.5 Transition matrices
This ﬁle is named
trnsprb.cdf
. This contains transition probabilities for both Moody’s
major and modiﬁed ratings, S&P major rating transition matrix, and J.P. Morgan derived
matrices estimating longterm ratings behavior. Initially, they will have data for a one
year risk horizon. However, the format supports other horizons.
CDFVersion v1.0
Date 02/15/1997
DataType YieldCurves
Currency CompoundingFrequency Maturity YieldToMaturity
CHF 1 1.0 0.055
CHF 1 2.0 0.05707
CDFVersion v1.0
Date 02/15/1997
DataType SpreadCurves
RatingSystem Rating Currency AssetType CompoundingFrequency Maturity Spread
Moody8 Aaa CHF BOND 1 5.0 0.01118
Moody8 Aaa CHF BOND 1 3.0 0.00866
Moody8 Aaa CHF BOND 1 10.0 0.015811
Moody8 Aaa CHF BOND 1 15.0 0.019365
Moody8 Aaa CHF BOND 1 2.0 0.007071
Moody8 Aaa CHF BOND 1 20.0 0.022361
Appendix H. Model inputs 165
The FromRating and ToRating columns of descriptive rating labels are included for read
ability. CreditMetrics only utilizes the numerical FromRating and ToRating columns.
H.6 Data Input Requirements to the Software Implementation of CreditMetrics
Table H.1
Required inputs for each issuer
Table H.2
Required inputs for each exposure type
CDFVersion v1.0
Date 02/15/97
DataType TransitionProbabilities
RatingSystem FromRank ToRank FromRating ToRating HorizonInMonths Probability
Moody18 0 0 Aaa Aaa 12 0.880784
Moody18 0 1 Aaa Aa1 12 0.050303
Moody18 0 2 Aaa Aa2 12 0.029015
Data Type Description
Issuer name Must be unique.
Credit Rating/Agency Long term rating that applies to the issuer's senior unsecured debt
regardless of the particular seniority class(es) listed as its expo
sure. Each rating has an agency (Moody's, S&P, etcetera)
Market Capitalization Stock price times number of shares outstanding
Country & Industry Proportion of sales assigned to speciﬁed countries and industries.
Issuerspeciﬁc risk Volatility of issuer asset returns not explained by industry/coun
try group(s).
Property Bond Loan Commitment MDI Receivable
Issuer Name x x x x x
Portfolio x x x x x
Currency x x x x x
Asset type x x x x x
Par value x x x
Maturity x x x x
Seniority class x
Recovery rate x x x x x
Recovery rate std x x x x x
Fixed or ﬂoating x x x
Coupon or spread x x x
Coupon frequency x x x
Current line x x
Current drawdown x
Expected drawdown x
Duration x
Expected exposure x
Average exposure x
Forward value x
166
Appendices
Appendix I. Indices used for asset correlations
Asset Category Index
Australia General MSCI Australia Index
Banking and ﬁnance ASX Banks & Finance Index
Broadcasting and media ASX Media Index
Construction and building materials ASX Building Materials Index
Chemicals ASX Chemicals Index
Energy ASX Energy Index
Food ASX Food & Household Goods Index
Insurance ASX Insurance Index
Paper and forest products ASX Paper & Packaging Index
Transportation ASX Transport Index
Austria General MSCI Austria Index
Belgium General MSCI Belgium Index
Canada General MSCI Canada Index
Automobiles Toronto SE Automobiles & Parts Index
Banking and ﬁnance Toronto SE Financial Services Index
Broadcasting and media Toronto SE Broadcasting Index
Construction and building materials Toronto SE Cement & Concrete Index
Chemicals Toronto SE Chemicals Index
Hotels Toronto SE Lodging, Food & Health Index
Insurance Toronto SE Insurance Index
Food Toronto SE Food Stores Index
Electronics Toronto SE Electrical & Electronics Index
Metals mining Toronto SE Metals Mines Index
Energy Toronto SE Integrated Oils Index
Health care and pharmaceuticals Toronto SE Biotechnology & Pharmaceuticals Index
Publishing Toronto SE Publishing & Printing Index
Transportation Toronto SE Transportation Index
Germany General MSCI Germany Index
Automobiles CDAX Automobiles Index
Banking and ﬁnance CDAX Investment Company Index
Chemicals CDAX Chemicals Index
Construction and building materials CDAX Construction Index
Insurance CDAX Insurance Index
Machinery CDAX Machinery Index
Paper and forest products CDAX Paper Index
Textiles CDAX Textiles Index
Transportation CDAX Transport Index
Utilities CDAX Utilities Index
Greece General MSCI Greece Index
Banking and ﬁnance Athens SE Banks Index
Insurance Athens SE Insurance Index
Finland General MSCI Finland Index
Banking and ﬁnance Helsinki SE Banks & Finance Index
Metals mining Helsinki SE Metal Index
Paper and forest products Helsinki SE Forest & Wood Index
Insurance Helsinki SE Insurance & Investment Index
Appendix I. Indices used for asset correlations 167
France General MSCI France Index
Automobiles SBF Automotive Index
Banking and ﬁnance SBF Finance Index
Construction and building materials SBF Construction Index
Energy SBF Energy Index
Food SBF Food Index
Hong Kong General MSCI Hong Kong Index
Banking and ﬁnance Hang Seng Finance Index
Utilities Hang Seng Utilities Index
Indonesia General MSCI Indonesia Index
Italy General MSCI Italy Index
Chemicals Milan SE Chemical Current Index
Banking and ﬁnance Milan SE Financial Current Index
Food Milan SE Food & Groceries Current Index
Paper and forest products Milan SE Paper & Print Current Index
Metals mining Milan SE Mine & Metal Current Index
Japan General MSCI Japan Index
Banking and ﬁnance Topix Banking Index
Broadcasting and media Topix Communications Index
Construction and building materials Topix Construction Index
Chemicals Topix Chemical Index
Electronics Topix Electrical Appliances Index
Food Topix Foods Index
Insurance Topix Insurances Index
Machinery Topix Machinery Index
Metals mining Topix Mining Index
Health care and pharmaceuticals Topix Pharmaceuticals Index
Paper and forest products Topix Pulp and Paper Index
Energy Topix Electric Power and Gas Index
Oil and gas  reﬁning and marketing Topix Oil and Coal Products Index
Textiles Topix Textile Products Index
Transportation Topix Transportation Equipment Index
Korea General MSCI Korea Index
Banking and ﬁnance Korea SE Finance Major Index
Construction and building materials Korea SE Construction Major Index
Chemicals Korea SE Chemical Company Major Index
Food Korea SE Food & Beverage Major Index
Insurance Korea SE Insurance Major Index
Machinery Korea SE Fabricated Metal & Machinery Major Index
Metals mining Korea SE Mining Major Index
Paper and forest products Korea SE Paper Product Major Index
Textiles Korea SE Textile & Wear Major Index
Transportation Korea SE Transport & Storage Major Index
Malaysia General MSCI Malaysia Index
Banking and ﬁnance KLSE Financial Index
Metals mining KLSE Mining Index
Asset Category Index
168 Appendix I. Indices used for asset correlations
CreditMetrics™—Technical Document
Mexico
General MSCI Mexico Index
Transportation Mexican SE Commercial & Transport Index
Metals mining Mexican SE Mining Index
Construction and building materials Mexican SE Construction Index
New Zealand
General MSCI New Zealand Index
Norway
General MSCI Norway Index
Banking and ﬁnance Oslo SE Bank Index
Insurance Oslo SE Insurance Index
Philippines
General MSCI Philippines Index
Metals mining Philippine SE Mining Index
Oil and gas  reﬁning and marketing Philippine SE Oil Index
Poland
General MSCI Poland Index
Portugal
General MSCI Portugal Index
Singapore
General MSCI Singapore Index
Hotels AllSingapore Hotel Index
Banking and ﬁnance AllSingapore Finance Index
Spain
General MSCI Spain Index
Sweden
General MSCI Sweden Index
Banking and ﬁnance Stockholm SE Banking Sector Index
Construction and building materials Stockholm SE Real Estate & Construction Index
Chemicals Stockholm SE Pharmaceutical & Chemical Index
Paper and forest products Stockholm SE Forest Industry Sector Index
Switzerland
General MSCI Switzerland Index
Banking and ﬁnance SPI Banks Cum Dividend Index
Construction and building materials SPI Building Cum Dividend Index
Chemicals SPI Chemical Cum Dividend Index
Electronics SPI Electronic Cum Dividend Index
Thailand
General MSCI Thailand Index
Banking and ﬁnance SET Finance Index
Chemicals SET Chemicals & Plastics Index
Electronics SET Electrical Components Index
Technology SET Electrical Products &Computers Index
Construction and building materials SET Building & Furnishing Materials Index
Energy SET Energy Index
Food SET Food & Beverages Index
Health care and pharmaceuticals SET Health Care Services Index
Insurance SET Insurance Index
Hotels SET Hotel & Travel Index
Machinery SET Machinery & Equipment Index
Metals mining SET Mining Index
Paper and forest products SET Pulp & Paper Index
Publishing SET Printing & Publishing Index
Textiles SET Textile Index
Transportation SET Transportation Index
Asset Category Index
Appendix I. Indices used for asset correlations 169
United Kingdom General MSCI United Kingdom Index
Banking and ﬁnance FTSEA 350 Banks Retail Index
Broadcasting and media FTSEA 350 Media Index
Construction and building materials FTSEA 350 Building Materials & Merchants Index
Chemicals FTSEA 350 Chemicals Index
Electronics FTSEA 350 Electronic & Electrical Equipment Index
Food FTSEA 350 Food Producers Index
Health care and pharmaceuticals FTSEA 350 Health Care Index
Insurance FTSEA 350 Insurance Index
Hotels FTSEA 350 Leisure & Hotels Index
Metals mining FTSEA 350 Extractive Industries Index
Oil and gas  reﬁning and marketing FTSEA 350 Gas Distribution Index
Energy FTSEA 350 Oil Integrated Index
Paper and forest products FTSEA 350 Paper, Packaging & Printing Index
Telecommunications FTSEA 350 Telecommunications Index
Textiles FTSEA 350 Textiles & Apparel Index
Transportation FTSEA 350 Transport Index
United States General MSCI United States Of America Index
Automobiles S&P Automobiles Index
Banking and ﬁnance S&P Financial Index
Broadcasting and media S&P Broadcasting (Television, Radio & Cable)
Construction and building materials S&P Building Materials Index
Chemicals S&P Chemicals Index
Electronics S&P Electronics (Instrumentation)
Energy S&P Energy Index
Entertainment S&P Entertainment Index
Food S&P Foods Index
Health care and pharmaceuticals S&P Health Care Index
Insurance S&P Insurance Composite Index
Hotels S&P LodgingHotels Index
Machinery S&P Machinery (Diversiﬁed)
Manufacturing S&P Manufacturing (Diversiﬁed)
Metals mining S&P Metals Mining Index
Oil and gas  reﬁning and marketing S&P Oil & Gas (Reﬁning & Marketing)
Paper and forest products S&P Paper & Forest Products Index
Publishing S&P Publishing Index
Technology S&P Technology Index
Telecommunications S&P Telecommunications (Long Distance)
Textiles S&P Textiles (Apparel)
Transportation S&P Transport Index
Utilities S&P Utilities Index
South Africa General MSCI South Africa (Gross Dividends Reinvested)
Banking and ﬁnance Johannesburg SE Financial Index
Metals mining Johannesburg SE Mining Holding Index
Asset Category Index
170 Appendix I. Indices used for asset correlations
CreditMetrics™—Technical Document
MSCI
Worldwide Automobiles Automobiles Price Index
Banking and ﬁnance Banking Price Index
Broadcasting and media Broadcasting & Pubs Price Index
Construction and building materials Construction & Housing (US$) Price Index
Chemicals Chemicals Price Index
Electronics Electronic Comps/Inst. Price Index
Energy Energy Sources Price Index
Entertainment Recreation & Other Goods Price Index
Food Food & Household Products Price Index
Health care and pharmaceuticals Health & Personal Care Price Index
Insurance Insurance Price Index
Hotels Leisure & Tourism Price Index
Machinery Machinery & Engineering Price Index
Metals mining Metals Nonferrous Price Index
Paper and forest products Forest Products/Paper Price Index
Telecommunications Recreation & Telecommunications Price Index
Textiles Textiles & Apparel Price Index
Transportation Transport Shipping Price Index
Utilities Utilities Electric & Gas Price Index
MSCI Regional EMF Latin America
Europe 14
Nordic Countries
North America
Paciﬁc
Paciﬁc ex Japan
Asset Category Index
171
Reference
172
CreditMetrics™—Technical Document
173
Glossary of terms
This glossary deﬁnes important terms in
CreditMetrics
.
accounting analytic.
The use of ﬁnancial ratios and fundamental analysis to estimate
ﬁrm speciﬁc credit quality examining items such as leverage and coverage measures,
with an evaluation of the level and stability of earnings and cash ﬂows.
(
See
page 58.)
allowance for loan and lease losses.
An accounting reserve set aside to equate expected
(mean) losses from credit defaults. It is common to consider this reserve as the buffer for
expected losses and some riskbased economic capital as the buffer for unexpected
losses.
(See page 60.)
autocorrelation (serial correlation).
When time series observations have a nonzero
correlation over time. Two empirical examples of autocorrelation are:
• Interest rates exhibit mean reversion behavior and are often negatively autocorre
lated (i.e., an up move one day will suggest a down move the next). But note that
mean reversion does not technically necessitate negative autocorrelation.
• Agency credit ratings typically exhibit move persistence behavior and are positively
autocorrelated during downgrades (i.e., a downgrade will suggest another down
grade soon). But, for completeness, note that upgrades do not better predict future
upgrades – we ﬁnd, they predict a “quiet” period; see also Altman & Kao [92].
(
See
page 32.)
average exposure.
Credit exposure arising from marketdriven instruments will have an
everchanging marktomarket exposure amount. The amount of exposure relevant to
our credit analysis is the timebucketed average exposure in each forward period across
the life of the transaction across all – probability weighted – market rate paths.
(
See
page 49.)
average shortfall.
The expected loss given that a loss occurs, or as the expected loss
given that losses exceed a given level.
(
See
page 137.)
credit exposure.
The amount subject to changes in value upon a change in credit quality
through either a market based revaluation in the event of an up(down)grade or the appli
cation of a recovery fraction in the event of default. (See
page 42
).
commitment.
A legally binding obligation (subject usually both to conditions precedent
and to continuing conditions) to make available loans or other ﬁnancial accommodation
for a speciﬁed period; this includes revolving facilities. Even during publicly known
credit distress, a commit can be legally binding if drawndown before it is formally with
draw for cause.
concentration risk.
Portfolio risk resulting from increased exposure to one obligor or
groups of correlated (e.g., by industry or location) obligors.
(
See
page 6.)
174
CreditMetrics™—Technical Document
correlation.
A linear statistical measure of the comovement between two random vari
ables. A correlation (Greek letter “
ρ
”, pronounced “rho”) will range from +1.0 to 1.0.
Observing “clumps” of ﬁrms defaulting together by industry or geographically is an
example of positive correlation of default events.
(
See
page 35.)
counterparty.
The partner in a credit facility or transaction in which each side takes
broadly comparable credit risk to the other. When a bank lends a company money, the
borrower (not Counterparty) has no meaningful credit risk to the bank. When the same
two agree on an atthemoney forward exchange contract or swap, the company is at risk
if the bank fails just as much as the bank is at risk if the counterparty fails (although for
the opposite movement in exchange or interest rates). After inception, swap positions
often move in/outofthemoney and the relative credit risk changes accordingly.
(
See
page 47.)
covenants.
The terms under which a credit facility will be monitored. Covenants are
most effective when they are speciﬁc measures that state the acceptable limits for change
in the obligor’s ﬁnancial and overall condition. They clearly deﬁne what is meant by
“signiﬁcant” deterioration in the obligor’s credit quality. Financial covenants are more
explicit (and therefore more desirable) than a “material adverse change” clause. Cross
default provisions are common: allowing acceleration of debt repayment.
(
See
page 43.)
credit distress.
A ﬁrm can have many types of credit obligations outstanding. These
may be of all manner of seniority, security and instrument type. In bankruptcy proceed
ings, it is not uncommon for different obligations to realize different recovery rates
including perhaps 100% recovery – zero loss. In our terminology, it is the obligor that
encounters credit distress carrying all of his obligations with him even though some of
these may not realize a true
default
(i.e., some may have zero loss).
(
See
page 65.)
credit exposure.
The amount subject to either changes in value upon credit quality
up(down)grade or loss in the event of default.
(
See
page 42.)
credit quality.
Generally meant to refer to an obligor’s relative chance of default, usu
ally expressed in alphabetic terms (e.g., Aaa, Aa, A, etc.). CreditMetrics makes use of an
extended deﬁnition that includes also the volatility of up(down)grades.
credit scoring
. Generically, credit scoring refers to the estimation of the relative likeli
hood of default of an individual ﬁrm. More speciﬁcally, this is a reference to the appli
cation of linear discriminant analysis to combine ﬁnancial rations to quantitatively
predict the relative chance of default.
(
See
page 57.)
ρ
X Y ,
COV
X Y ,
σ
X
σ
Y
⋅

X
i
X – ( ) Y
i
Y – ( )
i 1 =
N
∑
X
i
X – ( )
2
i 1 =
N
∑
Y
i
Y – ( )
2
i 1 =
N
∑
⋅
 = =
Glossary 175
Reference
current exposure.
For marketdriven instruments, the amount it would cost to replace a
transaction today should a counterparty default. If there is an enforceable netting agree
ment with the counterparty, then the current exposure would be the net replacement cost;
otherwise, it would be the gross amount.
default probability.
The likelihood that an obligor or counterparty will encounter credit
distress within a given time period. “Credit distress” usually leads to either an omitted
delayed payment or distressed exchange which would impair the value to senior unse
cured debt holders. Note that this leaves open the possibilities that:
• Subordinated debt might default without impairing senior debt value, and
• Transfers and clearing might continue even with a senior debt impairment.
(
See
page 65.)
dirty price.
Inclusion of the accrued value of the coupon in the quoted price of a bond
For instance, a 6% annual coupon bond trading at par would have a dirty price of $106
just prior to coupon payment. CreditMetrics estimates dirty prices since the coupon is
paid in nondefault states but assumed not paid in default.
(
See
page 10.)
distressed exchange.
During a time of credit distress, debt holders may be effectively
forced to accepted securities in exchange for their debt claim – such securities being of a
lower value than the nominal present value of their original claim. They may have a
lower coupon, delayed sinking funds, and/or lengthened maturity. For historical estima
tion of default probabilities, this would count as a default event since it can signiﬁcantly
impair value. In the U.S., exchange offers on traded bonds may be either registered with
the SEC or unregistered if they meet requirements under Section 3(a)(9) of the Securities
Act of 1933. Refer to Asquith, Mullins & Wolff [89].
(
See
page 65.)
duration.
The weighted average term of a security’s cash ﬂows. The longer the dura
tion, the larger the price movement given a 1bp change in the yield.
expected excession of a percentile level.
For a speciﬁed percentile level, the expected
loss given that the loss is more extreme than that percentile level.
(
See
page 137.)
exposure.
The amount which would be lost in a default given the worst possible
assumptions about recovery in the liquidation or bankruptcy of an obligor. For a loan or
fully drawn facility, this is the full amount plus accrued interest; for an unused or partly
used facility it is the full amount of the facility, since the worst assumption is that the
borrower draws the full amount and then goes bankrupt.
• Exposure is not usually a statistical concept; it does not make any attempt to assess
the probability of loss, it only states the amount at risk.
• For marketdriven instruments, (e.g., foreign exchange, swaps, options and deriva
tives) a proxy for exposure is estimated given the volatility of underlying market
rates/prices. See Loan Equivalent Exposure.
facility.
A generic term which includes loans, commitments, lines, letters, etc. Any
arrangement by which a bank accepts credit exposure to an obligor.
(
See
page 79.)
176
CreditMetrics™—Technical Document
fallen angels.
Obligors having both relatively high percentage risk and relatively large
exposure, whose large exposures were created when their credit ratings were better, but
who now have much higher percentage risk due to recent downgrades.
ISDA.
(Institutional Swap Dealers Association) A committee sponsored by this organi
zation was instrumental in drafting an industry standard under which securities dealers
would trade swaps. Included in this was a draft of a master agreement by which institu
tions outlined their rights to net multiple offsetting exposures which they might have to a
counterparty at the time of a default.
issuer exposure.
The credit risk to the issuer of traded instruments (typically a bond, but
also swaps, foreign exchange, etc.). Labeling credit spread volatility as either market or
credit risk is a question of semantics. CreditMetrics addresses market price volatility as
it is caused by changes in credit quality.
joint probabilities.
Standalone obligors have some likelihood of each possible credit
quality migration. Between two obligors there is some likelihood of each possible joint
credit quality migration. The probabilities are commonly inﬂuences by the correlation
between the two obligors.
(
See
page 36.)
kurtosis. C
haracterizes relative peakedness or ﬂatness of a given distribution compared
to a normal distribution. It is the fourth moment of a distribution.
Since the unconditional normal distribution has a kurtosis of 3, excess kurtosis is deﬁned
as
Kx3.
See
leptokurtosis.
leptokurtosis (fat tails)
. The property of a statistical distribution to have more occur
rences far away from the mean than would be predicted by a Normal distribution. Since
a normal distribution has a kurtosis measure of 3, excess kurtosis is deﬁned as
Kx3 > 0
.
A credit portfolio loss distribution will typically be leptokurtotic given positive obligor
correlations or coarse granularity in the size / number of exposures. This means that a
downside conﬁdence interval will be further away from the mean than would be
expected given the standard deviation and skewness.
letter of credit.
A promise to lend issued by a bank which agrees to pay the addressee,
the “beneﬁciary”, under speciﬁed conditions on behalf of a third party, also known as the
“account party”.
(
See
page 46)
.
There are different types of letters of credit. A
ﬁnancial
letter of credit (also termed a
standby letter of credit) is used to assure access to funding without the immediate need
for funds and is triggered at the obligor’s discretion. A
project
letter of credit is secured
by a speciﬁc asset or project income. A
trade
letter of credit is typically triggered by a
non credit related (and infrequent) event.
K
X
N
2
2N – 3 +
N 1 – ( ) N 2 – ( ) N 3 – ( )

X
i
x –
σ
x

¸ ,
¸ _
4
i 1 =
N
∑
¹ ¹
' ;
¹ ¹
3
N 1 – ( ) N 3 – ( )
N N 2 – ( ) N 3 – ( )
 – =
Glossary 177
Reference
liquidity. There are two separate meanings:
• At the enterprise level, the ability to meet current liabilities as they fall due; often
measures as the ratio of current assets to current liabilities.
• At the security level, the ability to trade in volume without directly moving the mar
ket price; often measured as bid/ask spread and daily turnover.
loan exposure. The face amount of any loan outstanding plus accrued interest plus. See
dirty price.
marginal standard deviation. Impact of a given asset on the total portfolio standard
deviation. (See page 129.)
marginal statistic. A statistic for a particular asset which is the difference between that
statistic for the entire portfolio and that for the portfolio not including the asset.
marketdriven instruments. Derivative instruments that are subject to counterparty
default (e.g., swaps, forwards, options, etc.). The distinguishing feature of these types of
credit exposures is that their amount is only the net replacement cost – the amount the
position is inthemoney – rather than a full notional amount. (See page 47).
market exposure. For marketdriven instruments, there is an amount at risk to default
only when the contract is inthemoney (i.e., when the replacement cost of the contract
exceeds the original value). This exposure/uncertainty is captured by calculating the net
ted mean and standard deviation of exposure(s).
Markov process. A model which deﬁnes a ﬁnite set of “states” and whose next progres
sion is determinable solely by the current state. A transition matrix model is an example
of a Markov process. (See page 71.)
mean. A statistical measure of central tendency. Sum of observation values divided by
the number of observations. It is the ﬁrst moment of a distribution. There are two types
of means. A mean calculated across a sample from a population is referred to as ,
while means calculated across the entire population – or means given exogenously – are
referred to as µ, pronounced “mu.” (See page 15.)
mean reversion. The statistical tendency in a time series to gravitate back towards a
long term historical level. This is on a much longer scale than another similar measure,
called autocorrelation; and these two behaviors are mathematically independent of one
another.
migration. Credit quality migration describes the possibility that a ﬁrm or obligor with
some credit rating today may move to (or “migrate”) to potentially any other credit rating
– or perhaps default – by the risk horizon. (See page 24.)
X
x
1
N

x
i
i 1 =
N
∑
=
178
CreditMetrics™—Technical Document
migration analysis. The technique of estimating the likelihood of credit quality migra
tions. See transition matrix.
moments (of a statistical distribution). Statistical distributions show the frequency at
which events might occur across a range of values. The most familiar distribution is a
Normal “Bell Shaped” curve. In general though, the shape of any distribution can be
described by its (inﬁnitely many) moments.
1. The ﬁrst moment is the mean which indicates the central tendency.
2. The second moment is the variance which indicates the width.
3. The third moment is the skewness which indicates any asymmetric “leaning”
either left or right.
4. The fourth moment is the kurtosis which indicates the degree of central “peaked
ness” or, equivalently, the “fatness” of the outer tails.
monotinicity. See rank order.
move persistence. The statistical tendency in a time series to move on the next step in
the same direction as the previous step (see also, positive autocorrelation).
netting. There are at least three types of netting:
closeout netting: In the event of counterparty bankruptcy, all transactions or all of a
given type are netted at market value. The alternative would allow the liquidator to
choose which contracts to enforce and which not to (and thus potentially “cherry pick”).
There are international jurisdictions where the enforceability of netting in bankruptcy
has not been legally tested.
netting by novation: The legal obligation of the parties to make required payments under
one or more series of related transactions are canceled and a new obligation to make only
the net payment is created.
settlement or payment netting: For cash settled trades, this can be applied either bilater
ally or multilaterally and on related or unrelated transactions.
notional amount. The face amount of a transaction typically used as the basis for inter
est payment calculations. For swaps, this amount is not itself a cash ﬂow. Credit expo
sure arises – not against the notional – but against the present value (market replacement
cost) of inthemoney future terminal payment(s).
obligor. A party who is in debt to another: (i) a loan borrower; (ii) a bond issuer; (iii) a
trader who has not yet settled; (iv) a trade partner with accounts payable; (v) a contractor
with unﬁnished performance, etc.; see Counterparty. (See page 5.)
option theoretic. An approach to estimating the expected default frequency of a partic
ular ﬁrm. It applies Robert Merton’s modeloftheﬁrm which states that debt can be
valued as a put option of the underlying asset value of the ﬁrm. (See page 36.)
originator. The ﬁnancial institution that extends credit on a facility which may later be
held by another institution through, for instance, a loan sale.
Glossary 179
Reference
peak exposure. For marketdriven instruments, the maximum (perhaps netted) exposure
expected with 95% conﬁdence for the remaining life of a transaction. CreditMetrics
does not utilize this ﬁgure because it is not possible to aggregate tail statistics across a
portfolio, since it is not the case that these “peaks” will all occur at the same time.
percent marginal standard deviation. Expression in percent terms of the impact of a
given asset on the total portfolio standard deviation. (See page 129.)
percentile level. A measure of risk based on the speciﬁed conﬁdence level of the portfo
lio value distribution: e.g., the likelihood that the portfolio market falls below the 99
th
percentile number is 1%. (See page 16.)
pricing grid. A schedule of credit spreads listed by credit rating that are applied to
either a loan or CreditSensitive Note (CSN) upon an up(down)grade of the obligor or
issuer. If the spreads are speciﬁed at market levels, then such terms reduce the volatility
of value across all nondefault credit quality migrations by keeping the instrument close
to par. (See page 67.)
rank order. A quality of data often found across credit rating categories where values
consistently progress in one direction – never reversing direction. Mathematicians term
this property of data, monotonicity. (See page 66.)
receivables. Non interest bearing short term extensions of credit in the normal course of
business, “trade credit,” that are at risk to the extent that the customer may not pay its
obligation in full. (See page 42).
revolving commitment (revolver). A generic term referring to some facility which a
client can use – or refrain from using – without canceling the facility.
sector loadings. For correlation analysis, a ﬁrm or industry group is said to be depen
dent upon underlying economic factors or “sectors” such as: (i) the market as a whole,
(ii) interest rates, (iii) oil prices, etc. As two industries “load” – are inﬂuenced by – com
mon factors, they will have a higher correlation between them.
serial correlation. See autocorrelation.
skewness. A statistical measure which characterizes the asymmetry of a distribution
around its mean. Positive skews indicate asymmetric tail extending toward positive val
ues (righthand side). Negative skewness implies asymmetry toward negative values
(lefthand side). It is the third moment of a distribution.
The distribution of losses across a credit portfolio will be positively skewed if there is
positive correlation between obligors or the size / number of exposures is coarsely gran
ular. This means that the conﬁdence interval out on the downside tail will be further
S
x
N
N 1 – ( ) N 2 – ( )

X
i
x –
σ
X

¸ ,
¸ _
3
i 1 =
N
∑
=
180
CreditMetrics™—Technical Document
away from the mean than would be expected given the portfolio’s standard deviation
alone.
standalone standard deviation. Standard deviation of value for an asset computed
without regard for the other instruments in the portfolio. (See page 129.)
standard deviation. A statistical measure which indicates the width of a distribution
around the mean. A standard deviation (Greek letter “σ,” pronounced “sigma”) is the
square root of the second moment of a distribution.
The distribution of losses across a credit portfolio will (typically) have a standard devia
tion which is much larger than its mean and yet negative losses are not possible. Thus, it
is not meaningful to think of a standard deviation as being a +/ range within which will
lie X% of the distribution – as one would naturally do for a normal distribution. (See
page 15.)
standalone percent standard deviation. Standalone standard deviation expressed as a
percentage of the mean value for the given asset. (See page 129.)
standby letter of credit. See letter of credit.
state of the world. A credit rating migration outcome; a new credit rating arrived at the
risk horizon. This can be either for a single obligor on a standalone basis or jointly
between two obligors. (See page 24.)
stochastic. Following a process which includes a random element. (See page 70.)
trade credit. See “receivables.”
transition matrix. A square table of probabilities which summarize the likelihood that a
credit will migrate from its current credit rating today to any possible credit rating – or
perhaps default – in one period. (See page 25.)
unexpected losses. A popular term for the volatility of losses but also used when refer
ring to the realization of a large loss which, in retrospect, was unexpected. (See
page 60.)
valueatrisk (VaR). A measure of the maximum potential change in value of a portfolio
of ﬁnancial instruments with a given probability over a preset horizon. (See page 5.)
σ
x
1
N 1 –

X
i
x – ( )
2
i 1 =
N
∑
=
Glossary 181
Reference
variance. A statistical measure which indicates the width of a distribution around the
mean. It is the second moment of a distribution. A related measure is the standard devi
ation, which is the square root of the variance. (See page 16.)
VAR
x
σ
x
2 1
N 1 –

X
i
x – ( )
2
i 1 =
N
∑
= =
182
CreditMetrics™—Technical Document
183
Reference
Bibliography
Alici, Yurt. “Neural Networks in Corporate Failure Prediction: The UK Experience,”
Working Paper, University of Exeter (1995).
Altman, Edward I. “Rating Migration of Corporate Bonds: Comparative Results and
Investor/Lender Implications,” Working Paper, New York University Salomon Center,
(Dec1996).
Altman, Edward I. “Corporate Bond and Commercial Loan Portfolio Analysis,” Work
ing Paper, (Sep1996).
Altman, Edward I. “Defaulted Bonds: Demand, Supply and Performance, 19871992,”
Financial Analysts Journal
, (May/Jun1993), pp. 5560.
Altman, Edward I. “Revisiting the High Yield Bond Market,” Working Paper Series,
New York University Salomon Center (Feb1992).
Altman, Edward I. “Measuring Corporate Bond Mortality and Performance,”
The Jour
nal of Finance
, (Sep1989), pp. 909922.
Altman, Edward I. “The Anatomy of the HighYield Bank Market,”
Financial Analysts
Journal
, (Jul/Aug1987), pp. 1225.
Altman, Edward I. “Financial Ratios, Discriminant Analysis and the Prediction of Cor
porate Bankruptcy”.
Journal of Finance
, (Sep1968), pp. 589609.
Altman, Edward I., and Joseph C. Bencivenga. “A Yield Premium Model for the High
Yield Debt Market,”
Financial Analysts Journal
, (Sep/Oct1995), pp. 4956.
Altman, Edward I., and Allan C. Eberhart. “Do Seniority Provisions Protect Bondhold
ers’ Investments?”
Journal of Portfolio Management
, (Summer1994), pp. 6775.
Altman, Edward I., and Robert G.Haldeman. “Valuation, Loss Reserves and Pricing of
Commercial Loans,” Working Paper Series, New York University Salomon Center
(May1992).
Altman, Edward I., and Duen Li Kao. “The Implications of Corporate Bond Ratings
Drift,”
Financial Analysts Journal
, (May/Jun1992), pp. 6475.
Altman, Edward I., and Duen Li Kao. “Rating Drift in High Yield Bonds,”
The Journal
of Fixed Income
, (Mar1992), pp 1520.
Altman, Edward I., and Duen Li Kao. “Examining and Modeling Corporate Bond Rating
Drift,” Working Paper Series, New York University Salomon Center (1991).
Altman, Edward I., Giancarlo Marco, and Franco Varetto. “Corporate Distress Diagno
sis: Comparisons Using Linear Discriminant Analysis and Neural Networks,” Working
Paper Series, New York University Salomon Center (Dec 1993).
184 Bibliography
CreditMetrics™—Technical Document
Altman, Edward I., and Scott A. Nammacher. “The Default Rate Experience of High
Yield Corporate Debt,”
Financial Analysts Journal
, (Jul/Aug1985), pp. 2541.
Altman, Edward I., and Paul Narayanan. “Business Failure Classiﬁcation Models: An
International Survey,” Working Paper, New York University, (May 1996).
Asarnow, Elliot. “Best Practices in Loan Portfolio Management,”
Journal of Lending
and Credit Risk Management
, (Mar1996), pp. 1424.
Asarnow, Elliot. “Measuring the Hidden Risks in Corporate Loans,”
Commercial Lend
ing Review
, Vol. 10, No. 1, (Winter 199495), pp. 2432.
Asarnow, Elliot, and David Edwards. “Measuring Loss on Defaulted Bank Loans: A 24
Year Study,”
The Journal of Commercial Lending
, (Mar1995).
Asarnow, Elliot, and James Marker. “Historical Performance of the U.S. Corporate Loan
Market: 19881993,”
The Journal of Commercial Lending
, Vol. 10 No. 2, (Spring 1995),
pp.1332.
Asquith, Paul, Robert Gertner, and David Scharfstein. “Anatomy of Financial Distress:
An Examination of JunkBond Issuers,”
Quarterly Journal of Economics
, (Aug1994),
pp. 625657.
Asquith, Paul, David W. Mullins, and Eric D. Wolff. “Original Issue High Yield Bonds:
Aging Analysis of Defaults, Exchanges, and Calls,”
The
Journal of Finance
, (Sep1989),
pp. 923952.
Austin, Donald G. “Use Migration Analysis to Reﬁne Estimates of Future Loan Losses,”
Commercial Lending Review
, Vol. 7, No. 2, (Spring 1992), pp. 3443.
Banks, Eric.
Volatility and Credit Risk in the Capital Markets,
(MacMillan Publishers
Ltd., Merrill Lynch, Tokyo, 1993).
Barclay, Michael J., and Clifford W. Smith. “The Priority Structure of Corporate Liabil
ities,”
Journal of Finance
, Vol. L, No. 3, (Jul1995), pp. 899917.
Bennett, Thomas L., Stephen F. Esser, and Christian G. Roth. “Corporate Credit Risk
and Reward,” Working Paper, Miller Anderson & Sherrerd (1993).
Bensoussan, Alain, Michael Crouhy, and Dan Galai. “Stochastic Equity Volatility
Related to the Leverage Effect,”
Applied Mathematical Finance
, Vol. 1, No. 1, (Sep
1994), pp. 6385.
Bernstein, Peter L.
Against the Gods: The Remarkable Story of Risk.
(New York: John
Wiley & Sons, Inc., 1996).
Blume, Marshall E., Howard Butcher III, and Donald B. Keim. “The Risk and Return of
LowGrade Bonds: An Update,”
Financial Analysts Journal
, (Sep/Oct1991), pp. 8589.
185
Reference
Blume, Marshall E., Donald B. Keim, and Sandeep A. Patel. “Returns and Volatility of
LowGrade Bonds 19771989,”
The Journal of Finance
, Vol. XLVI, No. 1, (Mar1991),
pp. 4974.
Cantor, Richard, and Frank Packer. “Determinants and Impact of Sovereign Credit Rat
ings,”
Economic Policy Review
, Vol. 2, No. 2, Federal Reserve Bank of New York, (Oct
1996), pp. 3753.
Cantor, Richard, and Frank Packer. “Multiple Ratings and Credit Standards: Differences
of Opinion in the Credit Rating Industry,” Federal Reserve Bank of New York, Research
Paper #9527, (Dec1995).
Carty, Lea V., and Dana Lieberman. “Corporate Bond Defaults and Default Rates 1938
1995,” Moody’s Investors Service, Global Credit Research, (Jan1996).
Carty, Lea V., and Dana Lieberman. “Defaulted Bank Loan Recoveries,” Moody’s
Investors Service, Global Credit Research, Special Report (Nov1996).
Chen, Yehning, J. Fred Weston, and Edward I. Altman. “Financial Distress and Restruc
turing Models,”
Financial Management
, Vol. 24, No. 2, (Summer1995), pp. 5775.
Cook, Douglas O., and Lewis J. Spellman. “Firm and Guarantor Risk, Risk Contagion,
and the Interﬁrm Spread among Insured Deposits,”
Journal of Financial and Quantita
tive Analysis
, (Jun1996), pp. 265281.
Cooper, Ian A., and Antonio S. Mello. “The Default Risk of Swaps,”
The Journal of
Finance
, Vol. XLVI, No. 2, (Jun1991), pp. 597619.
Crabbe, Leland. “A Framework for Corporate Bond Strategy,”
The Journal of Fixed
Income
, (Jun1995), pp. 1525.
Crabbe, Leland, and Mitchell A. Post. “The Effect of a Rating Downgrade on Outstand
ing Commercial Paper,”
The Journal of Finance
, Vol. XLIX, No. 1, (Mar 1994),
pp. 3956.
Cumby, Robert E., and Martin D.D. Evans. “The Term Structure of Credit Risk: Esti
mates and Speciﬁcation Tests,” Working Paper, Georgetown University and New York
University (Revised May1995).
Das, Sanjiv Ranjan, and Peter Tufano. “Pricing CreditSensitive Debt When Interest
Rates, Credit Ratings and Credit Spreads are Stochastic,”
The Journal of Financial Engi
neering
, Vol 5, No. 2, (Jun96), pp. 161198.
DeGroot, Morris H.
Probability and Statistics,
Second Edition
.
Addison Wesley Pub
lishing Co., Reading, Massachusetts (1986).
Duffee, Gregory R. “Estimating the price of default risk,” Working Paper, Federal
Reserve Board (Sep1995).
186 Bibliography
CreditMetrics™—Technical Document
Duffee, Gregory R. “On Measuring Credit Risks of Derivative Instruments,” Finance
and Economics Discussion Series, Federal Reserve Board, Washington, D.C., (Sep
1994).
Dufﬁe, Darrell, and Ming Huang. “Swap Rates and Credit Quality,”
The Journal of
Finance
, Vol. LI, No. 3, (Jul1996), pp. 921949.
Dutta, Soumintra, and Shashi Shekhar. “Bond Rating: A NonConservative Application
of Neural Networks,” IEEE International Conference on Neural Networks, (Jul1988),
pp. II443458.
Eberhart, Allan C., William T. Moore, and Rodney L. Roenfeldt. “Security Pricing and
Deviations from the Absolute Priority Rule in Bankruptcy Proceedings,”
The Journal of
Finance
, Vol. XLV, No. 5, (Dec1990), pp. 14571469.
Eberhart, Allan C., and Richard J. Sweeney. “Does the Bond Market Predict Bankruptcy
Settlements?”
The Journal of Finance
, Vol. XLVII, No. 3, (Jul1992), pp. 943980.
Episcopos, Athanasios, Andreas Pericli, and Jianxun Hu. “Commercial Mortgage
Default: A Comparison of the Logistic Model with Artiﬁcial Neural Networks,” Work
ing Paper, Clarkson University, Smith Barney Inc., City University of New York (1995).
Foss, Gregory W. “Quantifying Risk in the Corporate Bond Markets,”
Financial Ana
lysts Journal
, (Mar/Apr1995), pp. 2934.
Franks, Julian R., and Walter N. Torous. “An Empirical Investigation of U.S. Firms in
Reorganization,”
The Journal of Finance
, (Jul1989), pp. 747769.
Fridson, Martin S., and Yan Gao. “Primary versus Secondary Pricing of HighYield
Bonds,”
Financial Analysts Journal
, (May/Jun1996), pp. 2027.
Ginzburg, Alex, Kevin J. Maloney, and Ram Willner. “Risk Rating Migration and the
Valuation of Floating Rate Debt,” Working Paper, Citibank, Dartmouth College, Sanford
C. Bernstein & Co. (Nov1993).
Gollinger, Terri L., and John B. Morgan. “Calculation of an Efﬁcient Frontier for a
Commercial Loan Portfolio,”
Journal of Portfolio Management
, (Winter 1993),
pp. 3946.
Golubchin, Leonid. “The Pure Economic Theory of Default Risk for Sovereign Debt,”
Working Paper, City University of New York (Preliminary Draft, 29Dec94).
Golubchin, Leonid. “The ARCH Structure of the Sovereign Debt Risk Spread,” Work
ing Paper, City University of New York (Preliminary Draft, 29Dec94).
Hamilton, Robert, Barry Howcroft, Keith Pond, and Liu Zhonghua. “Predicting the Fate
of Insolvent Companies in Administrative Receivership — Using Statistical and Neural
Classiﬁcation Methods,” Working Paper, Loughborough University, (Dec1995).
Hradsky, Gregory T., and Robert D. Long. “HighYield Default Losses and the Return
Performance of Bankrupt Debt,”
Financial Analysts Journal
, (Jul/Aug1989), pp. 3849.
187
Reference
Hurley, William J., and Lewis D. Johnson. “On the Pricing of Bond Default Risk,”
Jour
nal of Portfolio Management
, (Winter 1996), pp. 6670.
Hyndman, Carl. “Are Bank Regulators Ready for Credit Scoring of Commercial
Loans?”
Commercial Lending Review
, (Winter 94/95), pp. 9295.
Jarrow, Robert A., David Lando, and Stuart M. Turnbull. “A Markov Model for the
Term Structure of Credit Risk Spreads,” Forthcoming, Review of Financial Studies,
(revised Aug1995).
Jarrow, Robert A., and Stuart M. Turnbull. “Pricing Derivatives on Financial Securities
Subject to Credit Risk,”
The Journal of Finance
, Vol. L. No. 1, (Mar1995), pp. 5385.
Jónsson, Jón G., and Martin S. Fridson. “Forecasting Default Rates on HighYield
Bonds,”
The Journal of Fixed Income
, (Jun1996), pp. 6977.
Kahn, Ronald N. “FixedIncome Risk Modeling in the 1990s,”
Journal of Portfolio
Management
, (Fall 1995), pp. 94101.
Kealhofer, Stephen. “Managing Default Risk in Portfolios of Derivatives,”
Derivative
Credit Risk: Advances in Measurement and Management
, (Renaissance Risk Publica
tions, 1995).
Kerling, Matthias. “Corporate Distress Diagnosis — An International Comparison,”
University of Freiburg, (1995).
Leibowitz, Martin L. Stanley Kogelman, and Lawrence R. Bader. “The Spread Curve
and the Risk/Return Decision,”
The Journal of Fixed Income
, (Jun1995), pp. 4350.
Leland, Hayne E. “Corporate Debt Value, Bond Covenants, and Optimal Capital Struc
ture,”
The Journal of Finance
, Vol. XLIX, No. 4, (Sep1994), pp. 12131252.
Leland, Hayne E., and Klaus Bjerre Toft. “Optimal Capital Structure, Endogenous
Bankruptcy, and the Term Structure of Credit Spreads,”
The Journal of Finance
, Vol. LI,
No. 3, (Jul1996), pp. 9871019.
Longstaff, Francis A., and Eduardo S. Schwartz. “A Simple Approach to Valuing Risky
Fixed and Floating Rate Debt,”
The Journal of Finance
, Vol. L, No. 3, (Jul1995), pp.
789819.
Longstaff, Francis A., and Eduardo S. Schwartz. “Valuing Credit Derivatives,”
The
Journal of Fixed Income
, (Jun1995), pp. 612.
Lucas, Douglas J. “The Effectiveness of Downgrade Provisions in Reducing Counter
party Credit Risk,”
The Journal of Fixed Income
, (Jun1995), pp. 3241.
Lucas, Douglas J. “Default Correlation and Credit Analysis,”
The Journal of Fixed
Income
, (Mar1995), pp. 7687.
Maden, Dilip B., and Haluk Unal. “Pricing the Risks of Default,” Working Paper, Uni
versity of Maryland (Oct1996).
188 Bibliography
CreditMetrics™—Technical Document
McAllister, Patrick H., and Douglas A. McManus. “Diversiﬁcation and Risk in Bank
ing: Evidence from Ex Post Returns,” Finance and Economics Discussion Series, Federal
Reserve Board, Washington, D.C. (Jun1992).
McAllister, Patrick H., and John J. Mingo. “Commercial Loan Risk Management,
CreditScoring, and Pricing: The Need for a New Shared Database,”
The Journal of
Commercial Lending
, Vol. 76, No. 9, (1994), pp. 622.
Merton, Robert C. “On the Pricing of Corporate Debt: The Risk Structure of Interest
Rates,”
The Journal of Finance
, Vol. 29, (1974), pp. 449470.
Meyer, Douglas W. “Using Quantitative Methods to Support CreditRisk Management,”
Commercial Lending Review
, Vol. 11 No. 1, (Winter 199596), pp. 5470.
Nielsen, Soren S., and Ehud I. Ronn. “The Valuation of Default Risk in Corporate
Bonds and Interest Rate Swaps,” Working Paper, University of Texas at Austin (Oct
1994, revised 13Feb96).
Queen, Maggie, and Richard Roll. “Firm Mortality: Using Market Indicators to Predict
Survival,”
Financial Analysts Journal
, (May/Jun1987), pp. 926.
Rajan, Raghuram G., and Luigi Zingales. “What Do We Know about Capital Structure?
Some Evidence from International Data,”
The Journal of Finance
, Vol. L, No. 5, (Dec
1995), pp. 14211460.
Longerstaey, Jacques, and Peter Zangari.
RiskMetrics™—Technical Document
. 4th Edi
tion. (New York: Morgan Guaranty Trust Co., 1996)
Sarig, Oded, and Arthur Warga. “Some Empirical Estimates of the Risk Structure of
Interest Rates,”
The Journal of Finance
, Vol. XLIV, No. 5, (Dec1989), pp. 13511360.
Scherr, Frederick C. “Optimal Trade Credit Limits,” Financial Management, Vol. 25,
No. 1, (Spring1996), pp. 7185.
Shane, Hilary. “Comovements of Lowgrade Debt and Equity Returns of Highly Lever
aged Firms,”
The Journal of Fixed Income
, (Mar1994), pp. 7989.
Shimko, David C., Naohiko Tejima, and Donald R. Van Deventer. “The Pricing of
Risky Debt When Interest Rates are Stochastic,”
The Journal of Fixed Income
, (Sep
1993), pp. 5865.
Shleifer, Andrei, and Robert W. Vishny. “Liquidation Values and Debt Capacity: A Mar
ket Equilibrium Approach,”
The Journal of Finance
, Vol. XLVII, No. 4, (Sep1992), pp.
13431366.
Skinner, Frank S. “A Trinomial Model of Bonds with Default Risk,”
Financial Analysts
Journal
, (Mar/Apr1994), pp. 7378.
Smith, L. Douglas, and Edward C. Lawrence. “Forecasting losses on a liquidating long
term loan portfolio,”
Journal of Banking & Finance
, Vol. 19, (1995), pp. 959985.
189
Reference
Sorensen, Eric H., and Thierry F. Bollier. “Pricing Swap Default Risk,”
Financial Ana
lysts Journal
, (May/Jun1994), pp. 2333.
Standard & Poor’s.
CreditWeek
, (April 15, 1996), pp. 4452.
Stevenson, Bruce G., and Michael W. Fadil. “Modern Portfolio Theory: Can It Work for
Commercial Loans?”
Commercial Lending Review
, Vol. 10, No. 2, (Spring 1995),
pp. 412.
Strang, Gilbert.
Linear Algebra and Its Applications.
3rd Edition
.
(Orlando: Saunders
College Publishing, 1988).
Swank, Thomas A., and Thomas H. Root. “Bonds in Default: Is Patience a Virtue?”
The
Journal of Fixed Income
, (Jun1995), pp. 2631.
Tyree, Eric K., and J. A. Long. “Assessing Financial Distress with Probabilistic Neural
Networks,” Working Paper, City University of London (1994).
Wagner, Herbert S. III. “The Pricing of Bonds in Bankruptcy and Financial Restructur
ing,”
The Journal of Fixed Income
, (Jun1996), pp. 4047.
Ward, David J., and Gary L. Griepentrog. “Risk and Return in Defaulted Bonds,”
Finan
cial Analysts Journal
, (May/Jun1993), pp. 6165.
190 Bibliography
CreditMetrics™—Technical Document
191
Index
Numerics
Ward & Griepentrog
78
A
Alici [95]
59
allowance for loan and lease losses
60
Altman
[87]
58, 65
[88]
58
[89]
24
[92]
58
Altman & Bencivenga [95]
58
Altman & Haldeman [92]
58, 133
Altman & Kao
[91]
60
[92]
60, 173
[92b]
32
Altman & Kishore [96]
26, 78
Altman & Nammacher [85]
58
Altman, Marco & Varetto [93]
59
Asarnow & Edwards [95]
79
Asarnow & Marker [95]
44, 45, 62
Asarnow [96]
62
Asquith, Mullins & Wolff [89]
58, 65, 175
Austin [92]
60
autocorrelation
32
average shortfall
137
average value
16
B
Bank for International Settlements
140
bellshaped distributions
17
Bennett, Esser & Roth [93]
62
Bernstein [96]
139
beta distribution
77
beta distribution.
80
bonds
18, 21, 43, 58, 62, 78
C
Carty & Lieberman [96a]
iv, 26, 58, 60, 61, 78, 82, 160
Carty & Lieberman [96b]
79
commitments to lend
18
concentration risk
6
conﬁdence interval
16
calculation
30
conﬁdence level
16
correlation
of two binomials
157
Crabbe [95]
60
credit derivatives
7
credit distress
65, 66
credit exposure
42
credit quality
65
credit rating systems
65
credit risk
computing on different horizons
32
in a portfolio
15
limits
135
portfolio
38
pricing of
57
credit risk measures
15
credit scoring
58
CreditMetrics
steps to quantify risk
24
D
Das & Tufano [96]
57, 60, 67, 133
debt classes
66
default event
deﬁnition
65
mean estimate
65
default probability
65, 71
discriminant analysis
59
distressed exchanges
58, 65
distribution of portfolio value, modeling
8
Dun & Bradstreet
64, 81
Dutta & Shekhar [88]
59
E
Eberhart & Sweeney [92]
78
Eberhart, Moore & Roenfeldt [90]
133
economic capital assessment
138, 139
Episcopos, Pericli & Hu [95]
59
expected default frequencies
64
exposure reduction
140
exposure size, absolute
133
F
fallen angels
134, 138
Fons [91]
59
Foss [95]
133
FourFifteen
8
Fridson & Gao [96]
133
FX forwards
8
G
Ginzburg, Maloney & Willner [93]
57, 60
Gollinger & Morgan [93]
64, 81
Greenspan, Alan
140
192
CreditMetrics™—Technical Document
H
historical tabulation
67
Hurley & Johnson [96]
133
J
jackkniﬁng
149
Jarrow & Turnbull [95]
133
Jarrow, Lando & Turnbull [96]
57, 60
joint likelihoods
36
joint probabilities
36
Jónsson & Fridson [96]
59
K
Kealhoffer [95]
85
Keiser [94]
161
Kerling [95]
59
KMV Corporation
59, 64, 81
L
letters of credit
19, 46
limit setting
140
Lo Pucki & Triantis [94]
161
loan commitment
43
loans
21, 62, 79
logistic regression
59
logit analysis
59
Lucas
[95a]
158
[95b]
60
M
Madan & Unal [96]
133
marginal risk
35, 40
marginal standard deviation
129
marginal statistic
118
market volatilities 8
marketdriven exposure uncertainty 8
marketdriven instruments 19, 47
Markov process 72, 75
Markowitztype analysis 63
McAllister, Patrick H. 93
mean reversion 32
Merton [74] 36, 57, 85, 133
Meyer [95] 60
migration analysis iv, 60
Monte Carlo
estimation 116
simulation 113
Monte Carlo simulation 137
Moody’s Investors Service 58
mutually exclusive outcomes 155
N
Neilsen & Ronn [96] 133
netting 7
neural network techniques 59
O
optionality 48
optiontheoretic approach 59
P
percent marginal standard deviation 129
percentile level 118, 137
performance evaluation 140
portfolio credit risk 7
portfolio effects 81
principal components analysis 59
probability of default. See Default event
probit analysis 59
R
Rajan & Zingales [95] 161
receivables 18, 42
Recoveries
mean estimate 77
uncertainty 77
recovery rate 77
recovery rate distribution 79
recovery rate estimates 79
risk
distinction between market and credit 8
idiosyncratic 98
marginal 40
risk level, statistical 133
RiskMetrics iii
S
S&P CreditWeek [96] 58
Sarig & Warga [89] 133
scenario generation 113
scenarios
default 116
nondefault 116
selforganizing feature maps 59
Shimko, Tejima & VanDeventer [93] 133
simulation 113, 118
Skinner [94] 133
Smith & Lawrence [95] 60
Sorenson & Bollier [94] 133
standalone percent standard deviation 129
standalone standard deviation 129
Standard & Poor’s 58
standard deviation 15, 137
standardized asset returns 92
Index 193
Reference
states of the world 24, 113
Stevenson & Fadil [95] 64, 81
Strang [88] 115
subordinated debt 66
Swank & Root [95] 78
T
time horizon
choosing 31
trade credit 18, 42
transition matrix 20, 25, 66, 71, 75
Tyree & Long [94] 59
U
unexpected losses 60
V
valueatrisk 5
W
Wagner [96] 62
White [93] 161
Z
Zeta Services, Inc. 59, 64, 81
194
CreditMetrics™—Technical Document
195
Part III: Applications
196
CreditMetrics™—Technical Document
197
Part III: Applications
198
CreditMetrics™—Technical Document
199
CreditMetrics™—Technical Document
Look to the J.P. Morgan site on the Internet located at
http://jpmorgan.com for updates of examples illustrating
points from this technical document or useful new tools.
Numerical examples are implemented in Excel spreadsheets.
These Excel spreadsheets are intended as a demonstration of the
CreditMetrics credit risk management methodology. They have
been designed as an educational tool and should not be used for
the risk estimation of actual portfolio positions. Separately,
J.P. Morgan sells software which embodies the CreditMetrics
methodology. If you have any questions about the use of these
spreadsheets contact your local J.P. Morgan representative or:
New York Greg M. Gupton (1212) 6488062
gupton_greg@jpmorgan.com
London Guy Coughlan (44171) 3255384
coughlan_g@jpmorgan.com
Singapore Michael Wilson (65) 3269901
wilson_mike@jpmorgan.com
CreditMetrics™—Technical Document
First Edition, 1997
page 200
CreditMetrics™ is based on, but differs signiﬁcantly from, the credit risk management systems developed by J.P. Morgan for its own use. J.P. Morgan does not warrant any
results obtained from the use of the CreditMetrics™ data, methodology, documentation or any information derived from the data (collectively the “Data”) and does not guarantee
its sequence, timeliness, accuracy, completeness or continued availability. The Data is calculated on the basis of historical observations and should not be relied upon to predict
future credit upgrades, downgrades, defaults or market movements. Examples are for illustrative purposes only; actual risks will vary depending on speciﬁc circumstances.
Additional information is available upon request. Information herein is believed to be reliable, but J.P. Morgan does not warrant its completeness or accuracy. Opinions and estimates constitute our judgment and are subject to change without
notice. Past performance is not indicative of future results. This material is not intended as an offer or solicitation for the purchase or sale of any ﬁnancial instrument. J.P. Morgan may hold a position or act as market maker in the ﬁnancial
instruments of any issuer discussed herein or act as advisor or lender to such issuer. Morgan Guaranty Trust Company is a member of FDIC and SFA. Copyright 1997 J.P. Morgan & Co. Incorporated. Clients should contact analysts at and
execute transactions through a J.P. Morgan entity in their home jurisdiction unless governing law permits otherwise.
CreditMetrics™ Products
Introduction to CreditMetrics™:
An abbreviated doc
ument which broadly describes the CreditMetrics™
methodology for measuring portfolio credit risk.
CreditMetrics™ – Technical Document:
A manual de
scribing the CreditMetrics™ methodology for estimating
credit risks. It fully speciﬁes how we construct the vola
tility of value due to credit quality changes for both stand
alone exposures and portfolios of exposures. It also dis
closes our approach to estimating credit exposures by in
strument type and a method of estimating correlations of
credit quality comovements. Finally, the manual de
scribes the format of the data set.
CreditMetrics™ Monitor:
A semiannual publication
which will discuss broad credit risk management issues,
statistical questions as well as new software implementa
tions and enhancements.
CreditMetrics™ data set:
A set of historical statistics
and results of academic and industry studies which will
be updated periodically.
All the above can be downloaded from the Internet at
http://www.jpmorgan.com/RiskManagement/CreditMet
rics
CreditManager™ PC Program
: A desktop software
tool that implements the methodology of CreditMetrics
and produces valueatrisk reports and other analysis of
credit risk such as those outlined in the CreditMetrics
documents. CreditManager can be purchased from J.P.
Morgan and any of the the cosponsors.
Trouble accessing the Internet?
If you encounter any
difﬁculties in either accessing the J.P. Morgan home page
on http://www.jpmorgan.com or downloading the Credit
Metrics™ data ﬁles, you can call (1800) JPMINET in
the United States.
Worldwide CreditMetrics™ Contacts
For more information about Credit
Metrics
™, please contact
the authors or any cosponsors listed below:
J.P. Morgan
Americas (1212) 6483461
cmx_amer@jpmorgan.com
Europe (44171) 3258007
cmx_euro@jpmorgan.com
Asia pacific (852) 29735646
cmx_asia@jpmorgan.com
Cosponsors
Bank of America Janet M. Tavakoli (1312) 8284732
Philip Basil (44171) 6344482
Walter Bloomenthal (1312) 8281668
Bank of Montreal Barry Campbell (1416) 8674809
Loretta Hennessey (1212) 6051541
BZW Jo Ousterhout (1212) 4126893
Michael Dyson (44171) 9563045
Loren Boston (852) 29032588
Deutsche Morgan Grenfell Hugo Bänziger (44171) 5452562
KMV Corporation David Nordby (1415) 7563337
edfs@kmv.com
Swiss Bank Corporation Robert Gumerlock (411) 2395739
robert.gumerlock@swissbank.com
Linda Bammann (1212) 3351085
linda.bammann@swissbank.com
Union Bank of Switzerland Hei Wai Chan (1212) 8215547
nycnh@ny.ubs.com
CreditMetrics™—Technical Document Copyright © 1997 J.P. Morgan & Co. Incorporated. All rights reserved. CreditMetrics™ is a trademark of J.P. Morgan in the United States and in other countries. It is written with the symbol ™ at its ﬁrst occurance in the publication, and as CreditMetrics thereafter.
CreditMetrics™ — Technical Document
iii
This book
Authors:
Greg M. Gupton Morgan Guaranty Trust Company Risk Management Research (1212) 6488062 gupton_greg@jpmorgan.com Christopher C. Finger Morgan Guaranty Trust Company Risk Management Research (1212) 6484657 ﬁnger_christopher@jpmorgan.com Mickey Bhatia Morgan Guaranty Trust Company Risk Management Research (1212) 6484299 bhatia_mickey@jpmorgan.com
This is the reference document for CreditMetrics™. It is meant to serve as an introduction to the methodology and mathematics behind statistical credit risk estimation, as well as a detailed documentation of the analytics that generate the data set we provide. This document reviews: • the conceptual framework of our methodologies for estimating credit risk; • the description of the obligors’ credit quality characteristics, their statistical description and associated statistical models; • the description of credit exposure types across “marketdriven” instruments and the more traditional corporate ﬁnance credit products; and • the data set that we update periodically and provide to the market for free. In the interest of establishing a benchmark in a ﬁeld with as little standardization and precise data as credit risk measurement, we have invited ﬁve leading banks, Bank of America, BZW, Deutsche Morgan Grenfell, Swiss Bank Corporation, and Union Bank of Switzerland, and a leading credit risk analytics ﬁrm, KMV Corporation, to be cosponsors of CreditMetrics. All these ﬁrms have spent a signiﬁcant amount of time working on their own credit risk management issues, and we are pleased to have received their input and support in the development of CreditMetrics. With their sponsorship we hope to send one clear and consistent message to the marketplace in an area with little clarity to date. We have also had many fruitful dialogues with professionals from Central Banks, regulators, competitors, and academics. We are grateful for their insights, help, and encouragement. Of course, all remaining errors and omissions are solely our responsibility. How is this related to RiskMetrics™?
We developed CreditMetrics to be as good a methodology for capturing counterparty default risk as the available data quality would allow. Although we never mandated during this development that CreditMetrics must resemble RiskMetrics, the outcome has yielded philosophically similar models. One major difference in the models was driven by the difference in the available data. In RiskMetrics, we have an abundance of daily liquid pricing data on which to construct a model of conditional volatility. In CreditMetrics, we have relatively sparse and infrequently priced data on which to construct a model of unconditional volatility. What is different about CreditMetrics? Unlike market risks where daily liquid price observations allow a direct calculation of valueatrisk (VaR), CreditMetrics seeks to construct what it cannot directly observe: the volatility of value due to credit quality changes. This constructive approach makes CreditMetrics less an exercise in ﬁtting distributions to observed price data, and more an exercise in proposing models which explain the changes in credit related instruments.
iv
Preface
And as we will mention many times in this document, the models which best describe credit risk do not rely on the assumption that returns are normally distributed, marking a signiﬁcant departure from the RiskMetrics framework. In the end, we seek to balance the best of all sources of information in a model which looks across broad historical data rather than only recent market moves and across the full range of credit quality migration — upgrades and downgrades — rather than just default. Our framework can be described in the diagram below. The many sources of information may give an impression of complexity. However, we give a stepbystep introduction in the ﬁrst four chapters of this book which should be accessible to all readers.
Exposures
User Portfolio Market volatilities Exposure distributions
Value at Risk due to Credit
Credit Rating Seniority Credit Spreads Present value bond revaluation
Correlations
Ratings series, Equities series Models (e.g., correlations) Joint credit rating changes
Rating migration likelihoods
Recovery rate in default
Standard Deviation of value due to credit quality changes for a single exposure
Portfolio Value at Risk due to Credit
One of our fundamental techniques is migration analysis, that is, the study of changes in the credit quality of names through time. Morgan developed transition matrices for this purpose as early as 1987. We have since built upon a broad literature of work which applies migration analysis to credit risk evaluation. The ﬁrst publication of transition matrices was in 1991 by both Professor Edward Altman of New York University and separately by Lucas & Lonski of Moody’s Investors Service. They have since been published regularly (see Moody’s Carty & Lieberman [96a]1 and Standard & Poor’s Creditweek [15Apr96]) and are also calculated by ﬁrms such as KMV. Are RiskMetrics and CreditMetrics comparable? Yes, in brief, RiskMetrics looks to a horizon and estimates the valueatrisk across a distribution of historically estimated realizations. Likewise, CreditMetrics looks to a horizon and constructs a distribution of historically estimated credit outcomes (rating migrations including potentially default). Each credit quality migration is weighted by its likelihood (transition matrix analysis). Each outcome has an estimate of change in value (given by either credit spreads or studies of recovery rates in default). We then aggregate volatilities across the portfolio, applying estimates of correlation. Thus, although the relevant time horizon is usually longer for credit risk, with CreditMetrics we compute credit risk on a comparable basis with market risk.
1
Bracketed numbers refer to year of publication.
CreditMetrics™—Technical Document
each with a deﬁned default probability. Just as a chain is only as strong as its weakest link. . Part I Risk Measurement Framework This section is for the general practitioner. We ﬁrst review the current academic context within which we developed our credit risk framework. Most prior work has been on the estimation of the relative likelihoods of default for individual ﬁrms. We wish to estimate the volatility of value due to changes in credit quality. These both embody the same modeling framework and Part II Part III 2 These assessments may be agency debt ratings. (ii) residual value estimates and their uncertainties. In our view. or any other approach. but rather by ﬁlling in what we believe is lacking. We review the statistical assumptions needed to describe discrete credit events. their mean expectations. We then look at how these credit statistics can be estimated to describe what happened in the past and what can be projected in the future. The second is a simulation approach which estimates the full distribution of value changes. it is only one link in the long chain of modeling and estimation that is necessary to fully assess credit risk (volatility) within a portfolio. however. we apply our framework across different exposures and across a portfolio. and how to interpret the results. This document is organized into three parts that address subjects of particular interest to our diverse readers. and not the ﬁnal output. not just the expected loss. By example. The ﬁrst is an analytic calculation of the mean and standard deviation of value changes. How is this document organized? What CreditMetrics is not One need not read and fully understand the details of this entire document to understand CreditMetrics. and correlations.Preface v We have sought to add value to the market’s understanding of credit risk estimation. Applications We discuss two implementations of our portfolio framework for estimating the volatility of value due to credit quality changes. a user’s internal ratings. there is content accessible to all readers. We provide a practicable framework of how to think about credit risk. how to apply that thinking in practice. and (iii) credit quality correlations across the portfolio. Model Parameters Although this section occasionally refers to advanced statistical analysis. We begin with an example of a single bond and then add more variation and detail. that these assessments are only inputs to CreditMetrics. it is also important to diligently address: (i) uncertainty of exposure such as is found in swaps and forwards. Moody’s and S&P have long done this and many others have started to do so. volatilities. We have designed CreditMetrics to accept as an input any assessment of default probability2 which results in ﬁrms being classiﬁed into discrete groups (such as rating categories). as important as default likelihood estimation is. It is important to realize. not by replicating what others have done before. the output of a statistical default prediction model.
P. CreditMetrics has been developed by the Risk Management Research Group at J. data and software implementation as we receive client and academic comments.P. limit setting. as well as professionals at other banks and academic institutions who offered input at various levels. CreditMetrics™—Technical Document . Morgan. Morgan who participated in this project. We thank numerous individuals at J. In particular. Acknowledgments We would like to thank our cosponsors for their input and support in the writing and editing of this document. Also. Special mention must go to Greg M. and whose work has inﬂuenced many of the methods presented here. this document could not have been produced without the contributions of our consulting editor. We also discuss how the results can be used in portfolio management. We welcome any suggestions to enhance the methodology and adapt it further to the changing needs of the market. which has been a pioneer in developing portfolio approaches to credit risk. Future plans We expect to update this Technical Document regularly. We apologize for any omissions. we thank the KMV Corporation. We intend to further develop our methodology. We encourage academic studies and are prepared to supply data for wellstructured projects. Gupton who conceived of this project and has been working on developing the CreditMetrics approach at JPMorgan for the last four years.vi Preface produce comparable results. and economic capital allocation. Margaret Dunkle.
5 Chapter 3.2 3.3 1.4 4.1 5.3 4.1 4. 3.3 Chapter 4.1 2.2 2.3 Chapter 6.4 2.2 1.3 6. 5.5 1.1 6.5 Risk Measurement Framework Introduction to CreditMetrics The portfolio context of credit Types of risks modeled Modeling the distribution of portfolio value Different credit risk measures Exposure type differences Data issues Advanced modeling features Standalone risk calculation Overview: Risk for a standalone exposure Step #1: Credit rating migration Step #2: Valuation Step #3: Credit risk estimation Choosing a time horizon Portfolio risk calculation Joint probabilities Portfolio credit risk Marginal risk Differing exposure types Receivables Bonds and loans Loan commitments Financial letters of credit (LCs) Marketdriven instruments 5 5 8 8 15 17 20 21 23 23 24 26 28 31 35 36 38 40 41 42 43 43 46 47 Part II Chapter 5.4 1. 6.4 Model Parameters Overview of credit risk literature Expected losses Unexpected losses A portfolio view Default and credit quality migration Default Credit quality migration Historical tabulation Longterm behavior 57 57 60 63 65 65 66 67 70 77 77 80 81 81 Chapter 7.2 5. 4.6 1. 2.vii Table of Contents Part I Chapter 1.3 2.2 6. 1.1 Estimating recovery rates 7.7 Chapter 2.1 3. Credit quality correlations 8.1 1.1 Finding evidence of default correlation . Recovery rates 7.2 4.2 Distribution of recovery rate Chapter 8.
2 Marginal standard deviation Chapter 10.5 Direct estimation of joint credit moves Estimating credit quality correlations through bond spreads Asset value model Estimating asset correlations 83 84 85 92 Part III Applications Chapter 9. 11.4 8. Inferring default correlations from default volatilities 155 157 161 Appendix G.2 12. Derivation of risk across mutually exclusive outcomes Appendix F. 10. Derivation of the product of N random variables Appendix E.3 11. Analytic standard deviation calculation Appendices Appendix B.3 Chapter 11.viii Table of contents 8. Indices used for asset correlations 159 163 166 Glossary of terms Bibliography Index Reference 173 183 191 CreditMetrics™ —Technical Document . 12.3 12.1 Threeasset portfolio 9. International bankruptcy code summary Appendix H. Derivation of the correlation of two binomials 149 147 153 Appendix D. Precision of simulationbased estimates Appendix C.1 10.4 Chapter 12.3 8.2 10.1 11.1 12. Model inputs Appendix I.2 11.2 8. Analytic portfolio calculation 9.4 Simulation Scenario generation Portfolio valuation Summarizing the results Portfolio example The example portfolio Simulation results Assessing precision Marginal risk measures Application of model outputs Prioritizing risk reduction actions Credit risk limits Economic capital assessment Summary 107 107 110 113 113 116 117 121 121 122 125 129 133 133 135 138 140 Appendix A.
3 Table 6.8 Table 2.4 Table 1.1 Table 4.1 Table 10.2 Table 4.8 Table 6.1 Table 6.5 Table 3.7 Table 1.7 Table 6.9 Table 8.1 Table 8. 69 KMV oneyear transition matrices as tabulated from expected default frequencies70 Average cumulative default rates (%) 71 Imputed transition matrix which best replicates default rates 72 Resulting cumulative default rates from imputed transition matrix (%) 73 73 “BB barrier” probabilities calculated from Table 6.3 Table 8.8 Table 8.6 Table 1.5 Table 8.9 Table 6. “par”) 26 Example oneyear forward zero curves by credit rating category (%) 27 Possible oneyear forward values for a BBB bond plus coupon 28 Calculating volatility in value due to credit quality changes 28 Joint migration probabilities with zero correlation (%) 36 Joint migration probabilities with 0.2 Table 5.2 Table 1.2 Table 4.2 Table 10.5 Table 5.1 Table 3.4 Table 2.2 Table 6.2 Table 9. i.3 Table 10.7 Table 8.5 Table 6.3 Table 1.R.3 Probability of credit rating migrations in one year for a BBB 9 Calculation of yearend values after credit rating migration from BBB ($) 10 Distribution of value of a BBB par bond in one year 11 Yearend values after credit rating migration from singleA ($) 12 All possible 64 yearend values for a twobond portfolio ($) 12 Probability of credit rating migrations in one year for a singleA 13 Yearend joint likelihoods (probabilities) across 64 different states (%) 14 Oneyear transition matrix (%) 20 Oneyear transition matrix (%) 25 Recovery rates by seniority class (% of face value.4 Table 4.6 matrix (%) Imputed transition matrix with default rate rank order constraint 74 Estimate of debt market proﬁle across credit rating categories 75 Achieving a closer ﬁt to the longterm steady state proﬁle 76 Recovery statistics by seniority class 78 Inferred default correlations with conﬁdence levels 82 Historically tabulated joint credit quality comovements 84 Historically tabulated joint credit quality comovement (%) 84 One year transition probabilities for a BB rated obligor 87 Threshold values for asset return for a BBB rated obligor 88 Transition probabilities and asset return thresholds for A rating 89 Joint rating change probabilities for BB and A rated obligors (%) 90 Countries and respective index families 94 Industry groupings with codes 95 Countryindustry index availability 96 Volatilities and correlations for countryindustry pairs 98 Transition probabilities (%) 107 Instrument values in future ratings ($mm) 108 Values of a twoasset portfolio in future ratings ($mm) 109 Transition probabilities (%) 114 Asset return thresholds 114 Correlation matrix for example portfolio 115 .6 Table 8.11 Table 9.3 Table 6.ix List of Tables Table 1.1 Table 8.5 Table 1.6 Table 6.1 Table 5.3 Table 4.1 Table 9.p.) 43 Average usage of commitments to lend 45 Example estimate of changes in drawdown 45 Revaluations for $20mm initially drawn commitment 46 Value of swap at the risk horizon in each rating state 51 Moody’s corporate bond average cumulative default rates (%) 58 Credit quality migration likelihoods for a BBB in one year 60 Volatility of historical default rates by rating category 61 Moody’s Investors Service: Oneyear transition matrix 68 Standard & Poor’s oneyear transition matrix – adjusted for removal of N.3 Table 2.10 Table 8.1 Table 1.e.30 asset correlation (%) 38 Fee on undrawn portion of commitment (b.2 Table 8.6 matrix (%) 74 “BB barrier” probabilities calculated from Table 6.11 Table 7.1 Table 2..10 Table 6.2 Table 2.4 Table 6.4 Table 8.
3 Table 11.4 Table 11.1 Table 11.2 Table 11.5 Table G.1 Table H.5 Table 10.x List of tables Table 10.6 Table 11.1 Table H.4 Table 10.2 Scenarios for standardized asset returns Mapping return scenarios to rating scenarios Valuation of portfolio scenarios ($mm) Example portfolio Asset correlations for example portfolio Percentiles of future portfolio values ($mm) Portfolio value statistics with 90% conﬁdence levels ($mm) Standard deviation of value change Summary of international bankruptcy codes Required inputs for each issuer Required inputs for each exposure type 115 116 117 121 122 125 126 130 161 165 165 CreditMetrics™—Technical Document .
9 Evolution of confidence bands for 0.1 Chart 5.1 Risk versus size of exposures within a typical credit portfolio Chart 12.2 Chart 3.4 Evolution of confidence bands for portfolio mean ($mm) Chart 11.1 Chart 1.3 Histogram of future portfolio values – lower 5% of scenarios Chart 11.xi List of Charts Chart 1.1 Chart 2.2 Chart 8.6 Evolution of confidence bands for 5th percentile ($mm) Chart 11.1 Chart 3.3 Chart 2.1 Chart 11.1 percentile ($mm) Chart 11.1 Chart 7.1 Chart 5.1 Chart 11.2 Chart 8.2 Possible risk limits for an example portfolio 7 11 14 23 24 35 37 37 41 60 63 67 75 79 80 86 88 91 92 118 123 123 124 126 127 127 128 128 129 131 134 135 .2 Comparison of distribution of credit returns and market returns Distribution of value for a 5year BBB bond in one year Distribution of value for a portfolio of two bonds Our ﬁrst “road map” of the analytics within CreditMetrics Examples of credit quality migrations (oneyear risk horizon) Our second “road map” of the analytics within CreditMetrics Model of firm value and its default threshold Model of ﬁrm value and generalized credit quality thresholds Our ﬁnal “road map” of the analytics within CreditMetrics Credit migration Construction of volatility across credit quality categories Model of ﬁrm value and migration Achieving a closer ﬁt to the longterm steady state proﬁle Distribution of bank facility recoveries Example beta distributions for seniority classes Credit rating migration driven by underlying BB ﬁrm asset value Distribution of asset returns with rating change thresholds Probability of joint defaults as a function of asset return correlation Translation of equity correlation to default correlation Frequency plot of portfolio scenarios Histogram of future portfolio values – upper 85% of scenarios Histogram of future portfolio values – scenarios between 95th and 65th percentiles Chart 11.7 Evolution of conﬁdence bands for 1st percentile ($mm) Chart 11.3 Chart 4.2 Chart 6.10 Marginal risk versus current value for example portfolio Chart 12.1 Chart 6.5 Evolution of confidence bands for standard deviation ($mm) Chart 11.8 Evolution of confidence bands for 0.2 Chart 1.3 Chart 8.2 Chart 3.2 Chart 7.4 Chart 10.1 Chart 8.5 percentile ($mm) Chart 11.
xii List of charts CreditMetrics™—Technical Document .
1 Risk Measurement Framework Part I .
2 CreditMetrics™—Technical Document .
commitments to lend. swaps and forwards. This twobond “portfolio” will serve to illustrate all the methodology we need to calculate credit risk across a portfolio of any size. We emphasize that our risk modeling framework is general. which will be central to our treatment of risk at the portfolio level. Part I is organized into the following four chapters: • Chapter 1: Introduction to CreditMetrics. In this chapter. • Chapter 3: Portfolio risk calculation. In this chapter. we provide details of how CreditMetrics estimates credit risk for a single bond. methodology and data requirements of CreditMetrics. we discuss how CreditMetrics addresses other instruments such as: receivables. ﬁnancial letters of credit. we extend the credit risk calculation to a portfolio containing two bonds and introduce the notion of correlations. loans. We emphasize the basic ideas and illustrate them by means of simple examples. and we discuss the data necessary to extend it to other exposure types. Using simple examples of one. In this chapter. • Chapter 2: Standalone risk calculation. • Chapter 4: Differing exposure types.3 Overview of Part I This section describes the risk measurement framework used in CreditMetrics. Later in Parts II and III. Part I: Risk Measurement Framework . For simplicity. we explain the ideas. we have limited our discussion in the previous two chapters to bonds. we discuss the merits and challenges of pursuing a quantitative portfolio approach to measuring credit risk. we give a more detailed treatment of CreditMetrics including methodology and data issues. Again. We discuss how we directly calculate the standard deviation of value due to credit quality changes. we illustrate the credit risk calculation for this portfolio with the help of a simple example. We give a summary of what we hope to achieve and the scope of our application.and twobond portfolios. In this chapter.
4 CreditMetrics™—Technical Document .
As credit exposures have multiplied. In general. deployment and management of credit risk taking across both a portfolio and marginal transactions. This is far from the more normally distributed market risks that VaR models typically address. These measures are consistent with the – perhaps more familiar – valueatrisk models which are used for market risks. The result of our efforts will be measures of valueatrisk due to credit quality changes. These measures will assist in the evaluation. Credit risk arises because the bond’s value in one year can vary depending on the credit quality of its issuer. Value changes will be relatively small with minor up(down)grades. Also. Of course. However. Introduction to CreditMetrics CreditMetrics is a tool for assessing portfolio risk due to changes in debt value caused by changes in obligor credit quality. • illustrate the resulting risk assessment with the simple example of a single bond. • discuss the beneﬁts and challenges to a portfolio approach and use a twobond example to show how we address a full portfolio. we assess the valueatrisk (VaR) – the volatility of value – not just the expected losses. but also by upgrades and downgrades in credit quality. Globally. we step through our CreditMetrics methodology and data in a survey fashion to give the broad picture of what we hope to achieve. the need for more sophisticated risk management techniques for credit risk has also increased. credit risk can be managed – as it has been – by more rigorous enforcement of traditional credit processes such as stringent underwriting standards. risk managers are increasingly seeking to quantify and integrate the overall credit risk assessment within a VaR statement which captures exposure to market.5 Chapter 1. Importantly. Speciﬁcally. limit enforcement and counterparty monitoring.1 The portfolio context of credit Credit risk has become perhaps the key risk management challenge of the late 1990s. and • summarize the data required for any credit instrument. but could be substantial – 50% to 90% are common – if there is a default. suppose we invest in a bond. For example. we assess risk within the full context of a portfolio. We address the correlation of credit quality moves across obligors. . • extend our focus to speciﬁc credit instruments other than bonds. rating change. and default risks. 1. institutions are taking on an increasing amount of credit risk. we know that the value of this bond will decline with a downgrade or default of its issuer – and appreciate if the credit quality of the obligor improves. We include changes in value caused not only by possible default events. In this chapter. we will: • establish the link between the process of credit quality migration and the resulting changes in debt value. This allows us to directly calculate the diversiﬁcation beneﬁts or potential overconcentrations across the portfolio.
The decision to take on ever higher exposure to an obligor will meet ever higher marginal risk – risk that grows geometrically with the concentration on that name.). as well as related market.). Another important reason to take a portfolio view of credit risk is to more rationally and accountably address portfolio diversiﬁcation. by location. risks.1. instrument type.6 Chapter 1. sector.1 There are also more practical reasons for a more quantitative approach to credit risk: • Financial products have become more complex.. Traditionally. similar additional exposure to an equally rated obligor who has relatively little existing exposure will entail less risk. CreditMetrics™—Technical Document . but also the uncertainty of loss. are increasingly looking to internal economic models for a better understanding of a bank's credit risk. Finally. ﬁxed exposure limits do not recognize the relationship between risk and return. The growth of derivatives activity has created uncertain and dynamic counterparty exposures that are signiﬁcantly more challenging to manage than the static exposures of more traditional instruments such as bonds or loans. Some bank regulators. The BIS riskbased capital guidelines do not distinguish high quality and welldiversiﬁed portfolios from low quality and concentrated portfolios. such a model creates a framework within which to consider concentrations along almost any dimension (industry. 1. Managers can then make risk versus return tradeoffs with knowledge of not only the expected credit losses.1 The need for a portfolio approach The primary reason to have a quantitative portfolio approach to credit risk management is so that we can more systematically address concentration risk. Over time. Introduction to CreditMetrics In the end. by capturing portfolio effects (diversiﬁcation beneﬁts and concentration risks) and recognizing that risk accelerates with declining credit quality. Furthermore. etc. Indeed. A more quantitative approach such as that presented here allows a portfolio manager to state credit lines and limits in units of marginal portfolio volatility. a portfolio credit risk methodology can be the foundation for a rational riskbased capital allocation process. 1 A capital measure reﬂecting these economic factors is a fundamental departure from the capital adequacy measures mandated for bank regulation by the Bank for International Settlements ("BIS"). Conversely. However. etc.g. but offer a relatively small marginal contribution to overall portfolio risk due to diversiﬁcation beneﬁts. Endusers and providers of these instruments need to identify such exposures and understand their credit. country. such names may be individually risky. Concentration risk refers to additional portfolio risk resulting from increased exposure to one obligor or groups of correlated obligors (e. Intuitive – but arbitrary – exposurebased credit limits have been the primary defense against unacceptable concentrations of credit risk. positions can be taken to best utilize risktaking capacity – which is a scarce and costly resource. a better understanding of the credit portfolio will help portfolio managers to better identify pockets of concentration and opportunities for diversiﬁcation. recognizing that the BIS regulatory capital regime can create uneconomic decision incentives and misleading presentation of the level of a bank's risk. portfolio managers have relied on a qualitative feel for the concentration risk in their credit portfolios. by industry.
The ﬁrst problem is that equity returns are relatively symmetric and are well approximated by normal or Gaussian distributions. Credit returns are characterized by a fairly large likelihood of earning a (relatively) small proﬁt through net interest earnings (NIE). 1. fundamental differences between credit risks and equity price risks make equity portfolio theory problematic when applied to credit portfolios. There are two problems. Thus. 1. we discussed why a portfolio approach to credit risk is necessary. • Improved liquidity in secondary cash markets and the emergence of credit derivatives have made possible more active management of credit risk based on rational pricing. we need more than just the mean and standard deviation to fully understand a credit portfolio’s distribution. In the following section. posted collateral. and netting.1 The portfolio context of credit 7 • The proliferation of credit enhancement mechanisms: thirdparty guarantees. Chart 1. we discuss why estimating portfolio credit risk is a much harder problem than estimating portfolio market risk. modern portfolio theory has taken enormous strides in its application to equity price risks. coupled with a (relatively) small chance of losing a Part I: Risk Measurement Framework . Above. makes it increasingly necessary to assess credit risk at the portfolio level as well as at the individual exposure level.2 Challenges in estimating portfolio credit risk Modeling portfolio risk in credit portfolios is neither analytically nor practically easy. In contrast. or credit events such as upgrades.1. the two statistical measures – mean (average) and standard deviation of portfolio value – are sufﬁcient to help us understand market risk and quantify percentile levels for equity portfolios. We can best understand these in the context of a portfolio model that also explicitly accounts for credit quality migrations. For instance.Sec. margin arrangements.1 Comparison of distribution of credit returns and market returns Typical market returns Typical credit returns Losses 0 Gains This long downside tail of the distribution of credit returns is caused by defaults. credit returns are highly skewed and fattailed (see Chart 1. However. Thus.1). downgrades or default. • Innovative new credit instruments explicitly derive value from correlation estimates. Proper due diligence standards require that institutions thoroughly review existing risks before hedging or trading them.
but is in no way limited to. CreditManager™. or (ii) proposing a model to capture credit quality correlations that has more readily estimated parameters. bonds. we will leave until Section 1. probabilities of migrations between these categories. For these instruments. For credit quality.2 Types of risks modeled A distinction is often drawn between “market” and “credit” risk. The ﬁrst step. These include the marketdriven volatility of credit exposures like swaps. 1. Potential remedies include either: (i) assuming that credit correlations are uniform across the portfolio. Groups of ﬁrms which KMV has assigned similar expected default frequencies could just as easily be used as “rating categories. we mean any grouping of ﬁrms of similar credit quality. we intend to address much of this difﬁculty. the distinction is not always clear (e.P.2 Typically. In summary.4 the calculation of standard deviations and percentile levels. 2 As a matter of implementation. By “rating categories”. Introduction to CreditMetrics fairly large amount of investment. we will be satisﬁed to obtain the distribution of outcomes.g. In addition. CreditMetrics estimates portfolio risk due to credit events. revaluation upon an up(down)grade.3 Modeling the distribution of portfolio value In this section. CreditMetrics allows us to capture certain market risk components in our risk estimates.. we begin to introduce some key modeling components: speciﬁcation of which rating categories3 to employ. Morgan software product called FourFifteen™ which uses RiskMetrics data sets of market volatility and correlation to analyze market risk. However. In other words. Across a large portfolio.8 Chapter 1. volatility of credit exposure due to FX moves). measuring risk across a credit portfolio is as necessary as it is difﬁcult. there is likely to be a blend of these two forces creating the smooth but skewed distribution shape above. forwards. the estimation of marketdriven exposure is performed in a J. the correlations can be directly estimated by observing highfrequency liquid market prices. For equities.” 3 CreditMetrics™—Technical Document . is to state exactly what risks we will be treating. the categories used by rating agencies. The second problem is the difﬁculty of modeling correlations. the software implementation of CreditMetrics. then. since credit is generally viewed over a larger horizon. But increasingly. For this section. However. marketdriven exposure estimates should match the longer credit risk horizon. it measures the uncertainty in the forward value of the portfolio at the risk horizon caused by the possibility of obligor credit quality changes – both up(down)grades and default. can accept marketdriven exposures from any source. and valuation in default. 1. This includes. volatility of value due to credit quality changes is increased by this further volatility of credit exposure. the lack of data makes it difﬁcult to estimate any type of credit correlation directly from history. With the CreditMetrics methodology. and to a lesser extent. market volatilities are estimated over a daily or monthly risk horizon.
Refer to Section 2.18 Note that there is a 86. or • the issuer defaults. and (ii) the probabilities of each outcome.1 Probability of credit rating migrations in one year for a BBB Yearend rating AAA AA A BBB BB B CCC Default Probability (%) 0. We derive these from historical rating data. or CCC. we know precisely the probabilities that this bond will end up in one of the seven rating categories (AAA through CCC) or defaults at the end of one year.30 1. These probabilities are shown in Table 1.95 86. Now we must obtain the value of the bond under 4 Throughout Part I.4 The second is to compute risk over a one year horizon.g. That is.95% for a rating change to singleA).18% likelihood of default.Sec. Of course. we make two choices. Consider a single BBB rated bond which matures in ﬁve years. we assume that the probabilities are known.1 Obtaining a distribution of values for a single bond To begin. So far we have speciﬁed: (i) each possible outcome for the bond’s yearend rating. 1. The ﬁrst is to utilize the Standard & Poor’s rating categories and corresponding transition matrices. There is a smaller likelihood of a rating change (e. we will consistently follow one set of credit quality migration likelihoods to aid clarity of exposition. Part I: Risk Measurement Framework . Table 1. Let us ﬁrst list all possible credit outcomes that can occur at the end of the year due to credit events: • the issuer stays at BBB at the end of the year. or A or down to BB. therefore we are interested in characterizing the range of values that the bond can take at the end of that period. Our risk horizon is one year.3 Modeling the distribution of portfolio value 9 1. • the issuer migrates up to AAA. other risk horizons may certainly be appropriate. This set of migration likelihoods happens to be taken from Standard & Poor’s.33 5. Each outcome above has a different likelihood or probability of occurring.5 for a discussion of how to choose a risk horizon. and a 0. For the purposes of this example.93% likelihood that the bond stays at the original rating of BBB.. 5. There are however a variety of data providers and we in no way wish to give the impression that we endorse one over any other. let us use S&P’s rating categories. for a bond starting out as BBB. which we will discuss at the end of the chapter. For now. AA. B.3.12 0.93 5.1.17 0.02 0.
First.00 – Forward Value 103. Say the face value of this bond is $100 and the coupon rate is 6%.10 Chapter 1.1). To obtain the value of the bond assuming an upgrade to singleA.00 6. Assuming annual coupons for our example. we have obtained the probabilities or likelihoods for the original BBB bond to be in any given rating category in one year (Table 1. Conversely. To illustrate. as well as the principal payment of $100 at maturity.10 77.13% in this example).00 6.66 101. we show the coupon payment received.00 6. we have also obtained the values of the bond in these rating categories (Table 1. the value decreases upon rating downgrade or default. the value of the bond increases if there is a rating upgrade. We leave aside the details of this calculation until the next chapter. consider our ﬁveyear BBB bond. In Table 1.13 Let us summarize what we have achieved so far.66 107.00 6. There is also a rise in value as the BBB remains BBB which is commonly seen when the credit spread curve is upward sloping. Four coupon payments ($6 each) remain. in the nondefault state. Further. Note that as expected. Introduction to CreditMetrics each of the possible rating scenarios. CreditMetrics™—Technical Document . as shown in Table 1. The information in Tables 1.37 103. the total value is due to a recovery rate (51. which we discuss in detail in the next chapter.3.2 is now used to specify the distribution of value of the bond in one year.2).19 108. at the end of one year we receive a coupon payment of $6 from holding the bond. In the default state.02 98.00 6. the forward bond value. Table 1.2. This zero curve is different for each forward rating category. and the total value of the bond (sometimes termed the dirty price of the bond).1 and 1.10 102.37 109.64 51. we discount these ﬁve cash ﬂows (four coupons and one principal) with interest rates derived from the forward zero singleA curve. we must ﬁnd the new present value of the bond’s remaining cash ﬂows at its new rating.00 6. The discount rate that enters this present value calculation is read from the forward zero curve that extends from the end of the risk horizon to the maturity of the bond. What value will the bond have at yearend if it is upgraded to singleA? If it is downgraded to BB? To answer these questions.13 Total Value 109.10 83. We want to ﬁnd the value of the bond at yearend if it upgrades to singleA.55 102.2 Calculation of yearend values after credit rating migration from BBB ($) Rating AAA AA A BBB BB B CCC Default Coupon 6.64 51.55 96. Here we just note that the calculations result in the following values at yearend across all possible rating categories.02 92.
000 50 60 70 80 CCC 90 100 AAA 110 AA A BBB Revaluation at risk horizon In the next section.075 BB 0. 1.2 Distribution of value for a 5year BBB bond in one year Frequency 0.3 Distribution of value of a BBB par bond in one year AAA AA A BBB BB B CCC Default Yearend rating Value ($) 109.18 The value distribution is also shown graphically in Chart 1. to characterize the distribution of values. 1.17 0. Now.Sec. as shown in Table 1.2.050 0. Chart 1.19 108. we deﬁne the credit risk estimate for this value distribution. The distribution of value tells us the possible values the bond can take at yearend.3. This bond pays annual coupons at the rate of 5%. Just as in the onebond case.93 5.900 0.100 0. we will discuss how we can generalize this probability distribution to a portfolio with more than just one instrument. we already know that the BBB bond can have one of the eight values at yearend. Now we will add a singleA three year bond to this portfolio.37 109. First. We want to obtain the distribution of values for this twobond portfolio in one year.33 5.2 Obtaining a distribution of values for a portfolio of two bonds So far we have illustrated the treatment of a standalone ﬁveyear BBB bond.95 86. Note that in the chart the horizontal axis represents the value.30 1.64 51.13 Probability (%) 0. and the probability or likelihood of achieving these numbers.02 98. we can calculate the corresponding year Part I: Risk Measurement Framework . and the vertical axis represents the probability.3 Modeling the distribution of portfolio value 11 Table 1. however. Similarly.10 83.2.66 107. we need to specify the portfolio’s possible yearend values and the probability of achieving these values.025 B Default 0.55 102.02 0.12 0.
8 • 8) different values.71 198.40 189. These values range from $102.86 215. By similarly calculating the values of the portfolio in the other states.59 106.73 186. as an example.15 96.10 83.77 102.25 214.5 shows the portfolio taking on 64 possible values at the end of a year depending on the credit rating migration of the two bonds.13 Next. Thus.96 215.79 158.59 if it upgrades to AAA.39 88.94 203.49 214.59 101.12 Chapter 1.77 BBB 103.50 160.00 5.02 98.15 214. Thus the portfolio as a whole has a value of $215.4.43 A 101.78 215.52 B 88.83 214.37 109.71 51. CreditMetrics™—Technical Document .96. the yearend value of the original BBB bond is $109. the portfolio can take on 64 (that is. From Table 1.64 98.2.52 212.04 208.72 AAA 106.19 108.55 102.26 190.94 157.64 103.13 160.4 Yearend values after credit rating migration from singleA ($) Yearend rating AAA AA A BBB BB B CCC Default Coupon 5. we leave the details of the calculation to the next chapter.5. Further.74 189.49 106.2 and 1.49 101.51 204. Table 1.28 156. We obtain the portfolio’s value at the risk horizon in each of the 64 states by simply adding together the values for the individual bonds.23 157. Since either of the bonds can have any of eight values in one year as a result of rating migration.69 190.00 5. Again.41 199.79 154.23 134.90 197.5.13 Total Value 106. we obtain the results shown in Table 1.00 5. which reads 215.28 BB AAA AA A BBB BB B CCC Default Obligor #1 (BBB) 109. Table 1.81 172.33 204. consider the top left cell in Table 1.67 215.00 5.03 152.30 213.00 – Forward Value 101.15 101.59 215.49 215.64 51. the yearend value of the original singleA bond is $106.68 153. we combine the possible values for the individual bonds (Tables 1.05 208. Introduction to CreditMetrics end values for the singleA rated bond.15 212.96 213.66 203.15 149.19 207.25 186. from Table 1.01 214.66 107.59) in the ﬁrst of 64 states.4.71 51.14 208.5 All possible 64 yearend values for a twobond portfolio ($) Obligor #2 (singleA) 105.35 139.30 105.37 196.64 215.39 73.37 + $106.32 159.58 210.37 if it upgrades to AAA.00 5.59 190. This cell corresponds to the outcome that both the BBB and singleA bonds upgrade to AAA at the end of the year.96 (= $109.39 210.4) to obtain the yearend values for the portfolio as a whole. but simply state the results in Table 1.85 208.26 So Table 1.13 157.76 210.49 185.34 211.81 210.17 201.61 204.13 106.84 CCC Default 51.00 5.26 (when both bonds default) to $215.96 (when both bonds are upgraded to AAA).30 100.08 197.68 215.70 205.30 215.62 AA 106.
0). the remaining piece in the puzzle is the likelihood or probability of observing these values.93% (the probability of BBB bond staying BBB from Table 1.27 91. a correlation equal to 0. this picture is simplistic. While doing this.09 2. we can also specify that they reﬂect some desired correlation (i. Unfortunately.6 Probability of credit rating migrations in one year for a singleA Yearend rating AAA AA A BBB BB B CCC Default Probability (%) 0. Thus. for example.Sec. because they are affected at least in part by the same macroeconomic factors. 5 By joint likelihoods. the ﬁrst obligor downgrades to BB while the second obligor remains at A.6. we derive these likelihoods from historical rating data.6. the likelihood that both bonds will maintain their original rating at yearend is simply equal to the product of 86. We will brieﬂy touch on the data issues at the end of the chapter.1 we showed the eight likelihoods for the BBB bond to be in each rating category in one year.1 and 1. Here we just note the numbers as they are given to us. Part I: Risk Measurement Framework .6) which is equal to 79.52 0. In reality. it is extremely important to account for correlations between rating migrations in an estimation of the risk on a portfolio.e. we mean the chance that the two obligors undergo a given pair of rating migrations.1 and 1..3 Modeling the distribution of portfolio value 13 We have illustrated the different possible values for the portfolio at the end of the year. In this case the joint likelihood is simply a product of the individual likelihoods from Tables 1.5 in one year. We must now estimate the 64 joint likelihoods5 so that we can calculate the volatility of value in our twobond example. the corresponding likelihoods for the singleA rated bond are displayed in Table 1.6. for example. Those 64 “joint” likelihoods must reconsile with each set of eight likelihoods which we have seen for the bonds on a standalone basis.01 0.05 5. So we must estimate the likelihood of observing each of the 64 states of Table 1. Table 1. In Table 1. To obtain the portfolio value distribution.15%.60 0. but leave aside the details for later (see Chapter 6). Similarly. assuming independence. These joint likelihoods must satisfy the constraint of summing to the standalone likelihoods in Tables 1. We introduce our model for correlations in Chapter 3 and describe the model in detail in Chapter 8. This is simple if the rating outcomes on the two bonds are independent of each other.1) and 91. Thus.74 0. 1.05% (the probability of a singleA bond staying as singleA from Table 1.06 Again. the rating outcomes on the two bonds are not independent of each other.
00 0. for the case mentioned above where the two bonds maintain their original ratings at the end of the year.7 1.44 79.69 4. Table 1. the actual joint likelihood value is 79.00 0.00 0. By plotting the likelihoods in Table 1.00 0.01 0. we obtain the portfolio value distribution shown in Chart 1. We remark that this becomes inconvenient. to do this in practice as the size of the portfolio grows.3 Obtaining a distribution of values for a portfolio of more than two bonds In our examples of one and two bond portfolios.4 Revaluation at risk horizon 211.55 0.01 0.57 0.17 0. we simply assume a correlation equal to 0.00 0.92 0.3 172.00 0.02 0.11 0.29 5.74 BB AAA AA A BBB BB B CCC Default Obligor #1 (BBB) AAA 0.7 Yearend joint likelihoods (probabilities) across 64 different states (%) Obligor #2 (singleA) BBB 0.52 0.00 91. from Table 1.93 5.00 0.95 86.00 0.5 on the same graph.00 0.00 0.00 0.00 0.02 0.33 5.04 0.01 Default 0.00 0.64 0.01 0.00 0.81 0.39 1.02 0.04 5.04 0.8 215.02 0.30 1.00 0.00 0.4 203.14 Chapter 1.00 0. Noting that for CreditMetrics™—Technical Document .26 B CCC 0.00 0.00 0. For other portfolio states.00 0. Introduction to CreditMetrics Here.12 0.00 0.69%. For example.00 0.02 0.18 0.7 below.3 Distribution of value for a portfolio of two bonds Probability 80% 70% 10% 0% 102.09 0.00 0.27 AA 0.04 0.08 4. and ﬁnally impossible. Speciﬁcally.02 0. Chart 1.7 we know the likelihoods of achieving each of these 64 values.00 0.06 We now have all the data with which to specify the portfolio value distribution.00 0.00 0.05 0.00 0. From Table 1.18 2.00 0.19 0.3 and take the resulting joint likelihoods as given.5 we know all the different 64 values that the portfolio can have at the end of a year.00 0.13 A 0.09 0.47 0.04 0. we have been able to specify the entire distribution of values for the portfolio.00 0.3.07 0.01 0.3.00 0.00 0.7 and the values in Table 1. the joint likelihood values are as shown in Table 1.
2 and 1.4 Different credit risk measures 15 a three asset portfolio. the larger the standard deviation. 1. The result of such an approach is an estimate of the portfolio distribution.5.1] Mean = p 1 ⋅ V 1 + p 2 ⋅ V 2 + … + p 64 ⋅ V 64 where p1 refers to the probability or likelihood of being in State 1 at the end of the risk horizon. Because of this exponential growth in the complexity of the portfolio distribution. For a ﬁve asset portfolio. and in general. Both measures reﬂect the portfolio value distribution and aid in quantifying credit risk. this number jumps to 32. We remark that it is always possible to compute some portfolio risk measures analytically. To calculate the standard deviation. 1. the two risk measures reﬂect potential losses from the same portfolio distribution. Neither is “best. With this background. 8 times 8 times 8) possible joint rating states. we next provide two alternative measures of credit risk that we use in CreditMetrics. for a portfolio with N assets. the likelihoods of each state are shown in Table 1. To illustrate the standard deviation calculation. 1. and the greater the risk. for larger portfolios. This is obtained by multiplying the values with the corresponding probabilities and then adding the resulting values. The greater the dispersion around the average value.” They both contribute to our understanding of the risk. they are different measures of credit risk.768.1 Credit risk measure #1: standard deviation The standard deviation is a symmetric measure of dispersion around the average portfolio value. the average value can be written as: [1. we ﬁrst must obtain the mean value for the portfolio. However.3. If the portfolio values are expressed in dollars. The credit risk in a portfolio arises because there is variability in the value of the portfolio due to credit quality changes.4. We emphasize that the credit risk model underlying both of these risk measures is the same. Part I: Risk Measurement Framework . Therefore. we utilize a simulation approach. Loosely speaking. there are 8N possible joint rating states. we expect any credit risk measure to reﬂect this variability. there are 512 (that is. regardless of the portfolio size. the greater the absolute amount at credit risk. the greater the dispersion in the range of possible values. Therefore. and V1 refers to the value in State 1.4 Different credit risk measures CreditMetrics can calculate two measures commonly used in risk literature to characterize the credit risk inherent in a portfolio: standard deviation and percentile level. which for large portfolios looks more like a smooth curve and less like the collections of a few discrete points in Charts 1. and discuss these in the following section. Mathematically. we again refer to our twobond portfolio. A simulation is very much like the preceding example except that outcomes are sampled at random across all the possibile joint rating states. and the values corresponding to these states are displayed in Table 1.7. this standard deviation calculation also results in a dollar amount.Sec. For this portfolio.
For instance. For now. By comparison. there is no particular percentile level that is “best” (5%.75% tail is 7. once we have calculated the 1st percentile level.2] (Standard Deviation)2= p1 · (V1 –Mean)2 + p2 · (V2 –Mean)2 +. Because the standard deviation statistic is a symmetric measure of dispersion. the standard deviation calculation is computationally simple and efﬁcient. is calculated as: [1.). which is also known as the “variance. simply note that since we do not have to rely on simulation to obtain the distribution of portfolio values.2 Credit risk measure #2: percentile level We deﬁne this second measure of risk as a speciﬁed percentile level of the portfolio value distribution.1). Carrying out the above calculation for our example portfolio. these distances for a normal distribution are 2. 1. the likelihood that the actual portfolio value is less than this number is only 1%.70 standard deviation below the average (the 99. We discuss this pairwise calculation in the remainder of this section.90 standard deviation below the average).4. Introduction to CreditMetrics Performing this simple calculation for our portfolio with the data from Table 1. distinguish in our example that the maximum upside value is only 0. and depends mostly on how the risk measure will be applied. it does not itself distinguish in our example between the gains side versus the losses side of the distribution. we ﬁnd that the average value for the portfolio is $213. Thus the 1st percentile level number provides us with a probabilistic lower bound on the yearend portfolio value.33 and 2. The interpretation of the percentile level is much simpler than the standard deviation: the lowest value that the portfolio will achieve 1% of the time is the 1st percentile..” The squareroot of this value is the standard deviation. Now the standard deviation. The interpretation of standard deviation is difﬁcult here because credit risk is not normally distributed.+ p64 · (V64–Mean)2 Note that the above expression yields the squared standard deviation value. we can operate pairwise across all pairs in the portfolio.63.5%.81 standard deviations respectively. 1%. we ﬁnd that the portfolio standard deviation is $3.25 standard deviations below the average. etc. The length of this downside tail could be characterized by its length in standard deviations..5 and Table 1.70 standard deviations above the average while the maximum downside value is 33. Rather. 0. Thus. for instance.16 Chapter 1. This is consistent with our earlier comment that the standard deviation measures the dispersion of the individual values around the average value.35. To calculate the standard deviation we do not have to specify the entire distribution of portfolio values. Of course. it is not possible to look up distribution probabilities in a normal table. the 99% tail is 1. It cannot. CreditMetrics™—Technical Document . Therefore.7. which measures the dispersion between each potential migration value (V’s) and this average value. The distribution of credit value is likely to have a long tail on the “loss” side and limited “gains” (see Chart 1. The individual terms in the expression are of the form (Vi –Average). The particular level used is the choice of the portfolio manager.
Sec. 1.5 Exposure type differences
17
For normal distributions (or any other known distribution which is completely characterized by its mean and standard deviation), it is possible to calculate percentile levels from knowledge of the standard deviation. Unfortunately, normal distributions are mostly a characteristic of market risk.6 In contrast, credit risk distributions are not typically symmetrical or bellshaped. In particular, the distributions display a much fatter lower tail than a standard bellshaped curve, as illustrated in Chart 1.1. Since we cannot assume that credit portfolio distributions are normal, nor can we characterize them according to any other standard distribution (such as the lognormal or Studentt), we must estimate percentile levels via another approach. To calculate a percentile level, we must ﬁrst specify the full distribution of portfolio values. For portfolios consisting of more than two exposures, this requires a simulation approach, which may be timeconsuming. Our approach will be to generate possible portfolio scenarios at random according to a Monte Carlo framework. While the generation of scenarios may be time consuming, once we obtain these scenarios, the calculation of the 1st percentile level is simple. To do this, we ﬁrst sort the portfolio values in ascending order. Given these sorted values, the 1st percentile level is the one below which there are exactly 1% of the total values. So if the simulation generates 10,000 portfolio values, the 1st percentile level is the 100th largest among these. Percentile levels may have more meaning for portfolios with many exposures, where the portfolio can take on many possible values. We may still consider our example portfolio with two bonds, however. For this portfolio, we estimate the 1st percentile to be $204.40. Note that this amount is $9.23 (= $213.63 – $204.40) less than the mean portfolio value. Thus, using the 1st percentile, we estimate the amount at credit risk to be $9.23, while using (one) standard deviation, we estimate this value at $3.35. Thus we see that the two measures give different values and so must be interpreted differently. These different computational requirements introduce a tradeoff between using the standard deviation and using the percentile level. The percentile level is intuitively appealing to use, because we know precisely what the likelihood is that the portfolio value will fall below this number. On the other hand, it is often much faster to compute the standard deviation. Users should evaluate this tradeoff carefully and use the risk measure that best ﬁts their purpose. Further discussion of this issue is presented in Chapter 12. 1.5 Exposure type differences
Up to this point our examples have used bonds, but the concepts that we have described in this chapter are equally applicable to other exposure types. The other exposure types we consider are receivables, loans, commitments to lend, ﬁnancial letters of credit and marketdriven instruments such as swaps and forwards. Recall from Section 1.3 that we derive both of our credit risk measures from the portfolio value distribution. Two components characterize this distribution. The ﬁrst is the likelihood of being in any possible portfolio state. The second is the value of the portfolio in each of the possible states. Only the calculation of future values is different for different instrument categories. The likelihoods of being in each credit quality state are the same for all instrument categories since these are tagged to the obligor rather than to each of its
6
See, for example, RiskMetrics™—Technical Document, 4th Edition, 1996.
Part I: Risk Measurement Framework
18
Chapter 1. Introduction to CreditMetrics
obligations. In the remainder of this section, we brieﬂy discuss the different exposure types in CreditMetrics. We provide a more detailed treatment in Chapter 4. 1.5.1 Receivables Many commercial and industrial ﬁrms will have credit exposure to their customers through receivables, or trade credit. We suggest that the risk in such exposures be addressed within this same framework. It will commonly be the case that receivable will have a “maturity” which is shorter than the risk horizon (e.g., one year or less). This would simplify matters in that there would be no need to revalue the exposure upon up(down)grade. But even if revaluing is necessary, the credit risk is – in concept – no different than the risk in a comparable bond issued to the customer, and so it can be revalued accordingly. 1.5.2 Bonds and loans For bonds, as we discussed in Section 1.3, the value at the end of the risk horizon is the present value of the remaining cash ﬂows. These cash ﬂows consist of the remaining coupon payments and the principal payment at maturity. To discount the cash ﬂows, we use the discount rates derived from the forward zero curve for each speciﬁc rating category. This forward curve is calculated as of the end of the risk horizon. We treat loans in the same manner as bonds, revaluing in each future rating state by discounting future cash ﬂows. This revaluation accounts for the change in the value of a loan which results from the likelihood changing that the loan will be repaid fully. 1.5.3 Commitments A loan commitment is a facility which gives the obligor the option to borrow at his own discretion. In practice, this essentially means both a loan (equal to the amount currently drawn on the line) and an option to increase the amount of the loan up to the face amount of the facility. The counterparty pays interest on the drawn amount, and a fee on the undrawn amount in return for the option to draw down further. For these exposures three factors inﬂuence the revaluation in future rating states: • the amount currently drawn; • expected changes in the amount drawn that are due to credit rating changes; and • the spreads and fees needed to revalue both the drawn and undrawn portions. All of these factors may be affected by covenants speciﬁc to a particular commitment. The details of commitment revaluation and typical covenants are discussed in Section 4.3.
CreditMetrics™—Technical Document
Sec. 1.5 Exposure type differences
19
1.5.4 Financial letters of credit A ﬁnancial or standby letter of credit is treated as an off balance sheet item until it is actually drawn. When it is drawn down its accounting treatment is just like a loan. However, the obligor can draw down at his discretion and the lending institution typically has no way to prevent a drawdown even during a period of obligor credit distress. Thus, for risk assessment purposes, we argue that the full nominal amount should be considered “exposed.” This means that we suggest treating a ﬁnancial letter of credit – whether or not any portion is actually drawn – exactly as a loan. Note that there are other types of letters of credit which may be either securitised by a speciﬁc asset or project or triggered only by some infrequent event. The unique features of these types of letters of credit are not currently addressable within the current speciﬁcation of CreditMetrics. 1.5.5 Marketdriven instruments For instruments whose credit exposure depends on the moves of underlying market rates, such as swaps and forwards, revaluation at future rating states is more difﬁcult. The complexity for these instruments comes from the fact that if a swap, for example, is marked to market and is currently outofthemoney to us, then a default by the counterparty does not inﬂuence the swap’s value, since we will still make the payments we owe on the swap.7 On the other hand, if the swap is inthemoney to us, then we expect payments, and do not receive the full amount in the case of a counterparty default. So in general, the credit exposure at any time to a marketdriven instrument is the maximum of the transaction’s net present value or zero. The methodology we propose for marketdriven instruments is applicable to single instruments, such as swaps or forwards, or to groups of swaps, forwards, bonds, or other instruments whose exposures can be netted. Thus, any set of cash ﬂows which are settled together (typically, these will all be exposures to the same counterparty) can be considered as one marketdriven instrument. In cases of default, we estimate the future value of marketdriven instruments using the expected exposure of the instrument at the risk horizon. This expected exposure depends both on the current market rates and their volatilities. In nondefault states, the revaluation consists of two parts: the present value of future cashﬂows, and the amount we might lose if the counterparty defaults at some future time. The second part, the expected loss, depends on the average marketdriven exposure over the remaining life of the instrument (which is estimated in a similar fashion to the expected exposure mentioned above), the probability that the counterparty will default over the same time (which is determined by the credit rating at the risk horizon), and the recovery rate in default. Details of this methodology and a discussion of the various exposure calculations appear in Section 4.5.
7
The exact settlement will depend on the covenants particular to the swap, but this is a reasonable assumption for explanatory purposes.
Part I: Risk Measurement Framework
20
Chapter 1. Introduction to CreditMetrics
1.6 Data issues
Given a choice of which rating system (that is, what groupings of similar credits) will be used, CreditMetrics requires three types of data: • likelihoods of credit quality migration, including default likelihoods; • likelihoods of joint credit quality migration; and • valuation estimates (e.g. bonds revalued at forward spreads) at the risk horizon given a credit quality migration. Together, these data types result in the portfolio value distribution, which determines the absolute amount at risk due to credit quality changes. 1.6.1 Data required for credit migration likelihoods We showed these individual likelihoods for BBB and singleA rating separately in Tables 1.1 and 1.6 respectively, but this information is more compactly represented in matrix form as shown below in Table 1.8. We call this table a transition matrix. The ratings in the ﬁrst column are the starting or current ratings. The ratings in the ﬁrst row are the ratings at the risk horizon. For example, the likelihoods in Table 1.8 corresponding to an initial rating of BBB are represented by the BBB row in the matrix. Further, note that each row of the matrix sums to 100%.
Table 1.8 Oneyear transition matrix (%) Initial rating AAA AA A BBB BB B CCC Rating at yearend (%) 0.06 0.64 5.52 86.93 7.73 0.43 1.30 BBB 0.12 0.06 0.74 5.30 80.53 6.48 2.38 BB
90.81 0.70 0.09 0.02 0.03 0 0.22
AAA
8.33 90.65 2.27 0.33 0.14 0.11 0
AA
0.68 7.79 91.05 5.95 0.67 0.24 0.22
A
0 0.14 0.26 1.17 8.84 83.46 11.24
B
0 0.02 0.01 0.12 1.00 4.07 64.86
CCC
0 0 0.06 0.18 1.06 5.20 19.79
Default
Source: Standard & Poor’s CreditWeek (15 April 96)
Transition matrices can be calculated by observing the historical pattern of rating change and default. They have been published by S&P and Moody’s rating agencies, and can be computed based on KMV’s studies, but any provider’s matrix is welcome and usable within CreditMetrics.8 The transition matrix should, however, be estimated for the same time interval as the risk horizon over which we are interested in estimating risks. For instance, a semiannual risk horizon would use a semiannual rather than oneyear transition matrix.
8
As we discuss later in Chapter 6, adjustments due to limited historical data may sometimes be desirable.
CreditMetrics™—Technical Document
there are three generic types: 1. 3. where the amount subject to credit risk is itself driven by market rates. we may propose a model for how the credit ratings of multiple names evolve together.Sec. CreditManager™. and estimate the requisite correlation parameters for the model. but it is also uncertain what the loss will be in the event of a default. Second. Therefore. 1. loans. We present mean and standard deviation estimates for recoveries in Chapter 7. which is generally not true. we need to do one of two things. such as the current version of J. Standard deviation estimates of recovery value also are available from public research. Marketdriven instruments. Thus. One such provision is for cases such as swaps and forwards. Morgan’s FourFifteen™. First. It is interesting to note that an obligor’s exposures across instruments can be estimated on a netted basis. we require the drawn and undrawn portions for a loan commitment and the spread/fees for both portions.P. recoveries in the event of default are notoriously uncertain. This is because using the product as joint likelihood implicitly assumes that the pairwise rating outcomes are independent of each other. our software implementation of CreditMetrics.7 Advanced modeling features CreditMetrics incorporates provisions to model additional parameters that make the credit risk estimate more precise. require an examination of exposures which is detailed in Section 4. we allow for the treatment of recoveries as random quantities.6.5. Coupon rate and term of maturity are required for: receivables.2 Data required for joint likelihood calculations Individual likelihoods are just one component of the portfolio joint likelihood. takes marketdriven exposures as an import from an external source. Thus. These have been detailed in Section 1. and bonds in order to revalue them. Part I: Risk Measurement Framework . We discuss several approaches to estimating joint likelihoods in Chapter 8.7 Advanced modeling features 21 1.5. including swaps. not only is it uncertain in these cases whether a counterparty will default or experience a change in credit quality. Estimates of the exposures in these cases rely on market rates and volatilities. as mentioned before. In addition to (1) above.3 Data required for portfolio value calculation Each instrument type requires sufﬁcient data to calculate the change in value for each possible credit quality migration. Earlier we stated that the joint likelihood is not simply the product of the individual likelihoods. to have joint likelihoods. and to a lesser extent bonds. we may historically tabulate joint credit rating moves just as we historically tabulated single credit rating moves in the transition matrix.6. 1. In general. 2. Thus. 1. letters of credit. Separately. these are provided in the CreditMetrics data set. forwards.
22 Chapter 1. Introduction to CreditMetrics CreditMetrics™—Technical Document .
23
Chapter 2.
Standalone risk calculation
This chapter illustrates the methodology used by CreditMetrics for calculating the credit risk for a single or standalone exposure. In Chapter 1, we summarized this methodology with the help of a BBB rated bond. Here, we discuss in detail each of the steps outlined in Chapter 1, using the same BBB example for illustration. Speciﬁcally: • we describe an individual obligor and how his credit rating implies both a default likelihood and the likelihoods for possible credit quality migrations; • we describe a credit exposure and how its seniority standing implies a loss rate (that is, loss in the event of default); • we describe credit spreads over the default free yield and their implication for the bond value upon up(down)grade in credit quality; and • we assemble all of these pieces to estimate volatility of value due to credit quality changes. 2.1 Overview: Risk for a standalone exposure
There are three steps to calculating the credit risk for a “portfolio” of one bond, as illustrated in Chart 2.1 below: • Step 1: The senior unsecured credit rating of the bond’s issuer determines the chance of the bond either defaulting or migrating to any possible credit quality state at the risk horizon. • Step 2: The seniority of the bond determines its recovery rate in the case of default. The forward zero curve for each credit rating category determines the value of the bond upon up(down)grade. Both of these aid revaluation of the bond. • Step 3: The likelihoods from Step 1 and the values from Step 2 then combine in our calculation of volatility of value due to credit quality changes.
Chart 2.1 Our ﬁrst “road map” of the analytics within CreditMetrics
Value at Risk due to Credit
Credit Rating Seniority Credit Spreads Present value bond revaluation
Rating migration likelihoods
Recovery rate in default
Standard Deviation of value due to credit quality changes for a single exposure
24
Chapter 2. Standalone risk calculation
Readers who are familiar with RiskMetrics will see that the framework for credit risk shown above is different from the market risk framework. This is because the quality and availability of credit data are generally much different. Therefore, we construct what we cannot directly observe. In the process, we model the mechanisms of changes in value rather than try to observe value changes. In the following sections, we detail each step used in CreditMetrics to quantify the risk of a standalone exposure. We illustrate these steps with our senior unsecured 5year BBB rated bond. This bond pays an annual coupon at the rate of 6%. We also include this example in the CHAP01.XLS Excel spreadsheet, which is available on our web site location (http://www.jpmorgan.com). The calculations performed in this chapter assume a risk horizon of one year. This choice is somewhat arbitrary. However, at the end of this chapter we discuss some of the issues surrounding the choice of this risk horizon. 2.2 Step #1: Credit rating migration
Credit Rating Rating migration
In our model, risk comes not only from default but also from changes in value due to up(down)grades. Thus, it is important for us to estimate not only the likelihood of default but also the chance of migrating to any possible credit quality state at the risk horizon. So we view default as just one of several “states of the world” that may exist for this credit one period from now. The likelihood of any credit rating migration in the coming period is conditioned on the senior unsecured credit rating of the obligor.1 Chart 2.2 shows the credit quality migration likelihoods for obligors currently rated A, AAA, and BBB. For our BBB bond, the rightmost diagram is applicable.
Chart 2.2 Examples of credit quality migrations (oneyear risk horizon)
Currently A rated. 0.09% AAA 2.27% A A 91.05% A 5.52% B B B 0.74% B B 0.26% B 0.01% CCC 0.06% D 100.00% Currently AAA rated. AAA 90.81% AAA 8.33% A A 0.68% A 0.06% B B B 0.12% B B 0.00% B 0.00% CCC 0.00% D 100.00% Currently BBB rated. 0.02% AAA 0.33% A A 5.95% A BBB 86.93% B B B 5.30% B B 1.17% B 0.12% CCC 0.18% D 100.00%
A
Chart 2.2 says, for example, that there is a 5.30% chance that a BBB rated credit will downgrade to a BB rating within one year. There are several common patterns among the three examples. Intuitively, we see that the most likely credit rating one year from now is the current credit rating. The next most likely ratings are one letter grade above or below. The only absolute rule about credit quality migrations is that the likelihoods
1
There are some academic studies which condition the estimation of default likelihood upon not only the current credit rating but also whether the speciﬁc debt issue is new: see for instance Altman [89]. While this is sound and has its applications, we believe that many users will have dealings with established – not just new – obligors.
CreditMetrics™—Technical Document
Sec. 2.2 Step #1: Credit rating migration
25
must sum to 100% since these are all the “states of the world” that are possible. Rather than showing each rating’s credit quality migration likelihoods separately, it is often convenient to think of them in a square table, or transition matrix, as shown below in Table 2.1.
Table 2.1 Oneyear transition matrix (%) Initial Rating
AAA AA A
Rating at yearend (%) 0.06 0.64 5.52 86.93 7.73 0.43 1.30
BBB
AAA AA A BBB BB B CCC
90.81 0.70 0.09 0.02 0.03 0 0.22
8.33 90.65 2.27 0.33 0.14 0.11 0
0.68 7.79 91.05 5.95 0.67 0.24 0.22
0.12 0.06 0.74 5.30 80.53 6.48 2.38
BB
0 0.14 0.26 1.17 8.84 83.46 11.24
B
0 0.02 0.01 0.12 1.00 4.07 64.86
CCC
Default 0 0 0.06 0.18 1.06 5.20 19.79
Source: Standard & Poor’s CreditWeek (15 April 96)
To read this table, ﬁnd today’s credit rating on the left and follow along that row to the column which represents the rating at the risk horizon. For instance, the leftmost bottom ﬁgure of 0.22% says that there is a 0.22% likelihood that a CCC rated credit will migrate to AAA at the end of one year. We derived the transition matrix in Table 2.1 from rating migration data published by S&P. Thus the leftmost bottom ﬁgure of 0.22% means that 0.22% of the time (over the 15year history from which this data was tabulated) a CCCrated credit today migrated to AAA in one year. Of course, migrating from CCC to AAA within one year is highly unusual and likely represents only one instance in the historical data.2 This presents a practical problem: results based on limited data are subject to estimation errors. Later, in Chapter 6, we discuss the anticipated longterm behavior of credit migrations, which would tend to mitigate this estimation noise. As we have mentioned, it is possible to create transition matrices for any system of grouping similar credits. Again, we refer to these groupings loosely as rating categories. Regardless of how the rating categories are constructed and of how many categories there are, it is necessary to specify the default likelihood for each category, and the likelihoods that ﬁrms in one category migrate to any other. In addition, as we will see in the following section, for the purposes of revaluation, it is necessary to provide a credit spread to correspond to each category as well.
2
The only adjustment we made to S&P’s data was for the “nolongerrated” migrations. The CCC row in this transition matrix, sourced from S&P, is based upon 561 ﬁrm/years worth of observation with 79 occurrences of a transition to “no longer rated.” Across all rows in this transition matrix, there are more than 25,000 ﬁrm/years worth of observation, with most being in the BBBtoAA rows.
Part I:. Risk Measurement Framework
26
Chapter 2. Standalone risk calculation
2.3 Step #2: Valuation In Step 1, we determined the likelihoods of migration to any possible credit quality states at the risk horizon. In Step 2, we determine the values at the risk horizon for these credit quality states. Value is calculated once for each migration state; thus there are (in this example) eight revaluations in our simple onebond example. These eight valuations fall into two categories. First, in the event of a default, we estimate the recovery rate based on the seniority classiﬁcation of the bond. Second, in the event of up(down)grades, we estimate the change in credit spread that results from the rating migration. We then perform a present value calculation of the bond’s remaining cash ﬂows at the new yield to estimate its new value. 2.3.1 Valuation in the state of default
Seniority Recovery rate in default
If the credit quality migration is into default, the likely residual value net of recoveries will depend on the seniority class of the debt. In CreditMetrics, we offer several historical studies of this dependence.3 Table 2.2 below summarizes the recovery rates in the state of default as reported by one of the available studies.
Recovery rates by seniority class (% of face value, i.e., “par”) Seniority Class Senior Secured Senior Unsecured Senior Subordinated Subordinated Junior Subordinated Mean (%) 53.80 51.13 38.52 32.74 17.09 Standard Deviation (%) 26.86 25.45 23.81 20.18 10.90
Table 2.2
Source: Carty & Lieberman [96a] —Moody’s Investors Service
In this table, we show the mean recovery rate (middle column) as well as the standard deviation of the recovery rate (last column). Our example BBB bond is senior unsecured. Therefore, we estimate its mean value in default to be 51.13% of its face value – which in this case we have assumed to be $100. Also from Table 2.2, the standard deviation of the recovery rate is 25.45%. 2.3.2 Valuation in the states of up(down)grade If the credit quality migration is to another letter rating rather than to default, then we must revalue the exposure by other means. To obtain the values at the risk horizon corresponding to rating up(down)grades, we perform a straightforward present value bond revaluation. This involves the following steps:
3
Credit Spreads PV bond revaluation
There is also a recent study (see Altman & Kishore [96]) which conditions recovery rates on industry participations of the obligor in addition to seniority class.
CreditMetrics™—Technical Document
27 8. let us calculate the value V of the bond at the end of one year assuming that the bond upgrades to singleA.12 5. Part I:.60 3.78 4. Assume that the forward zero curves for each rating category has been given to us.+ .78 8.73 4. After completing these calculations for different rating categories.05 15.= 108.93% ) ( 1 + 5. Obtain the forward zero curves for each rating category.52 Year 4 First.05 Year 1 4. let us determine the cash ﬂows which result from holding the bond position.1] 6 6 6 6 V = 6 + .22 4. To calculate the value of the bond in a rating category other than singleA.63 7.02 7.4.10 5.03 Year 3 5.32 4.25 6. the bond pays $6 each at the end of the next four years.65 3.17 5.+ .3 below. we obtain the values in Table 2.02 15.17 4.66 2 3 4 ( 1 + 3. Recall that our example bond pays an annual coupon at the rate of 6%. Now.3 Example oneyear forward zero curves by credit rating category (%) Category AAA AA A BBB BB B CCC 3. We show an example in Table 2.55 6. assuming a face value of $100. These forward curves are stated as of the risk horizon and go to the maturity of the bond. Therefore. revalue the bond's remaining cash ﬂows at the risk horizon for each rating category. and pays annual coupons at the rate of 6%. At the end of the ﬁfth year. we would substitute the appropriate zero rates from the table.02 Year 2 4. we use the forward zero rates for the singleA rating category from Table 2.72 4. 2. the bond pays a cash ﬂow of face value plus coupon.32% ) ( 1 + 4. Using these zero curves.72% ) ( 1 + 4. Risk Measurement Framework . 2.32 5.67 6. Let us illustrate the above steps with the help of our BBB bond example.03 14.3. Recall that this bond has a ﬁveyear maturity. This calculation is described by the formula below: [2. which equals $106 in this case.Sec.52 13. Table 2.+ .3 Step #2: Valuation 27 1.93 5.32% ) In the above formula.
are shown in Table 2.19 108.37 109.28 0.12 0.47 1.5 0.17 0.6598 5. respectively.66 107.41 (5.96) $107.09 (55.4 Step #3: Credit risk estimation Standard deviation of value due to credit changes for a single exposure We now have all the information that we need to estimate the volatility of value due to credit quality changes for this one exposure on a standalone basis.10 6.64 Default 51. which we obtain from Steps 1 and 2 respectively. Standalone risk calculation Possible oneyear forward values for a BBB bond plus coupon AAA AA A BBB BB B CCC Table 2.13 2.19 108.55 102. First we consider the calculation of the standard deviation.28 Chapter 2. That is.2.1853 1.64 51.49 0.15 (8.0146 0. 2.30 1.57 93.6358 8.0010 0. Here we use these two columns to calculate the risk estimate.10 83.37 109.33 5. we have to ﬁrst obtain the average value (the mean).10 83.66 107. and 2.3.4.93 5. Calculating volatility in value due to credit quality changes Yearend rating AAA AA A BBB BB B CCC Default Probability of state (%) 0.06) 1.13 Mean = Probability weighted value ($) Difference of Probability value from weighted difference mean ($) squared 0.99) 1.5 below.45) 0.02 98.02 2. CreditMetrics™—Technical Document .99 Table 2.02 0.9477 $2.4 Yearend rating Value ($) 109.36 2.1474 0. These likelihoods and values.18 New bond value plus coupon ($) 109.02 98.9446 0. we know the likelihood of all possible outcomes – all up(down)grades plus default – and the distribution of value within each outcome.55 102.3592 0.10 (23.1 Calculation of standard deviation as a measure of credit risk Recall from Chapter 1 that there are two useful measures of credit risk that one can use: standard deviation and percentile level.09 Variance = Standard deviation = The ﬁgures in the ﬁrst two columns – likelihoods of migration and value in each state – have been discussed in Sections 2.46 5. For this.95 86.
95% ⋅ 108. The standard deviation then measures the dispersion between the individual values and this mean. which includes the $6.12% ⋅ 83. let us present the standard deviation calculation in a manner that will help us later to incorporate recovery rate uncertainty. As shown in Table 2. In particular.64 + 2 0.02% ⋅ 109. (This is the expected recovery rate for a senior unsecured bond from Table 2.66 + 86. this is estimated at $107. This uncertainty adds to the overall credit risk of holding the bond position.64 + 0. Risk Measurement Framework .00 coupon in all nondefault states.37 + 0.5. we calculate the mean µ and the standard deviation σ using the formulae below: s s µ Total = i=1 ∑ pµ i i σ Total = i=1 ∑ p µ – µ Total i i 2 2 [2.13 = 107.Sec.19 2 + 2 5. we pointed out that there is an uncertainty or standard deviation associated with this recovery rate.17% ⋅ 98.10 2 + 2 0.30% ⋅ 102.33% ⋅ 109.12% ⋅ 83.02% ⋅ 109.95% ⋅ 108. Before we describe how to account for this recovery rate uncertainty. we used a recovery value of $51. This (or some scale of this) is one measure of the absolute amount that is at credit risk. In the above calculation of standard deviation.02 2 + 1.99.09 The above formula is overly simple in that it allows the bond to only take on a mean value within each state.18% ⋅ 51. We incorporate this added uncertainty as follows: Part I:.5 respectively).13 = 2.30% ⋅ 102.66 + 2 86. the bond can take on a distribution of values within each state.) While discussing the results presented in this table. Let pi be the probability of being in any given state and µi be the value within each state (the ﬁrst and second columns of Table 2.37 2 + 0. we observe that the standard deviation of value changes due to credit is $2.55 + = 5.17% ⋅ 98.93% ⋅ 107.13 for the case of default. In general.93% ⋅ 107.19 + 5.09.4 Step #3: Credit risk estimation 29 Note that the mean is just the probabilityweighted average of the values across all rating categories including default.55 + = 5.10 + 0.2.18% ⋅ 51. Given this.99 2 – 107.02 + 1. there is well documented uncertainty surrounding the recovery rate in default. 2. After completing the calculations.2] 0.09 0.33% ⋅ 109.
note the zero values in the standard deviation formula.18% ⋅ 51.13 = 107.12% ⋅ 83. For the sake of illustration.02 + 1. The other risk measure is the percentile level. we hope to allow credit spread volatility in future versions of CreditMetrics.13 + 25. There are good reasons why different users should use different percentile levels. percentile levels are more meaningful statistics for large portfolios.17% ⋅ 98.66 + 0 ) + 2 2 86.18% ⋅ ( 51. where the portfolio can take on many different portfolio values. Finally. Again.19 2 + 0 2 ) + 2 2 5. As we have mentioned before.33% ⋅ ( 109.12% ⋅ ( 83.10 + 0. however. where we add a component σi representing the uncertainty in recovery value in the defaulted state i = 8. For now. The only difference is in the standard deviation calculation.02% ⋅ ( 109.93% ⋅ ( 107. This inclusion of the uncertainty in recovery rates increases the standard deviation from $2. It is also the CreditMetrics™—Technical Document . This is the level below which our portfolio value will fall with probability 1%. refer to Appendix D. Say we are interested in determining the 1st percentile level for our bond.55 + = 5.18 (a 6.19 + 5.4.37 + 0.02 2 + 0 2 ) + 1.66 + 86.09 0.99 to $3.64 + 0 ) + 2 2 0.3] 0.30% ⋅ ( 102.18 2 –107.55 + 0 ) + = 5.09 Note that the expected value or mean calculation remains the same as before.95% ⋅ 108. we concentrate on the calculation of the 1st percentile level. we expect an uncertainty in value in the other rating up(down)grade states.64 + 0. These zeros represent the uncertainty of value in the up(down)grade states. Standalone risk calculation µ Total = i=1 ∑ s pµ i i σ Total = i=1 ∑ s p ( µ i + σ i ) – µ Total i 2 2 2 [2. This would be caused by the uncertainty of credit spreads within each credit rating category.93% ⋅ 107.02% ⋅ 109. For a derivation of this formula for the standard deviation.33% ⋅ 109.95% ⋅ ( 108.2 Calculation of percentile level as a measure of credit risk Standard deviation is just one of two useful credit risk measures.30 Chapter 2.17% ⋅ ( 98.32% increase). we have set this credit spread uncertainty to zero since it is unclear what portion of it is systematic versus diversiﬁable.45 ) = 3.30% ⋅ 102.37 2 + 0 2 ) + 0. 1% is not the only percentile level we advocate for the reader. If we ever have sufﬁcient data to resolve this issue. Just as there is an uncertainty in value in the default state.10 2 + 0 2 ) + 2 2 0. 2.
12%). This now exceeds 1%. We therefore stop here and read off the corresponding value from the B row. that there is nothing about the CreditMetrics methodology that requires a oneyear horizon. in order to provide an example. it is necessary to perform simulations to compute percentile levels. This value.. Indeed. Part I:. It is important to note.1 Should there be one horizon or many? The choice of time horizon for risk measurement and risk management is not clear because there is no explicit theory to guide us. Risk Measurement Framework . credit relationships. so we move up to the CCC state. This is less than 1%. The managers for each securitytype (e. CCC. the one thing that is clear is that comparisons between alternatives must be made at the same risk horizon.10. we discuss issues surrounding the choice of risk horizon. This is $8.5 Choosing a time horizon Much of the academic and credit agency data is stated on an annual basis.99 below the mean value. loans versus swaps) may wish to see their security type calculated at their own risk horizon. and common lack of credit hedging instruments can all lead to prolonged riskmitigating actions.30% (sum of 0. Let us go through the procedure: The likelihood of being in the defaulted state is 0. Nonetheless. However. The choice of risk horizon raises two practical questions: • Should a practitioner use only one risk horizon or many? • Is there any ﬁrm basis for saying that any one particular horizon is best? 2. The value at which this running total ﬁrst becomes equal to or greater than 1% is the 1st percentile level. Illiquidity. 2.Sec. is the 1st percentile level value. Below. Table 2. This is a convention rather than a requirement. for any portfolio with much more than two assets).17% (sum of 0. or B is now equal to 2. it may be that interest rate swaps are more liquid than loans. so we move up again. the risk estimates for these different subportfolios cannot be aggregated if there is a mismatch in time horizons.5. For instance. However. So far we have used an arbitrary risk horizon of one year. this time to B rating state. the state of default. The combined likelihood of being in default. One of the common arguments in favor of multiple credit risk horizons is that they allow us to calculate risk at horizons tailored to each credit security type. which is equal to $98.18% and 0.17%). Many different security types bear credit risk.18%.30% and 1. we compute percentile levels for our single bond.g. it is difﬁcult to support the argument than any one particular risk horizon is best. 2.5 Choosing a time horizon 31 case that for these large portfolios (in fact. We keep a running total of the likelihoods as we move up. We start from the bottom of the table. and move upwards towards the AAA rating state. The combined likelihood of being in default or CCC state is 0. This is also less than 1%.5 displays the likelihood that our bond will be in any given credit rating at the risk horizon and the value at each credit rating.
commitments. some markets tend to exhibit mean reversion (that is. Since relative risk measurements will likely drive decisions. Standalone risk calculation 2. however.2 Which horizon might be “best”? Almost any risk measurement system is better at stating relative risk than it is at stating absolute risk. An analogy is driving a car. perhaps every second and every meter. On the other side. Altman & Kao [92b] ﬁnd that there is positive autocorrelation in S&P downgrades. Note. looking at the S&P rating data. this methodology ignores the issue of autocorrelation in the credit quality changes over multiple time horizons.5. the issue of time period interdependencies can also arise for credit quality migrations. the natural turnover due to the ongoing maturity and reinvestment of positions provides appreciable room for riskmitigating action even for highly illiquid instruments. risk stated over the coming year can guide riskmitigating actions.5. and also ﬁnd that an upgrade tends to lead to a “quiet” period. A car’s instrument panel serves perfectly well when reporting speed at kilometers per hour even though driving decisions are made far more often. that this ﬁnding applies in particular to the S&P rating system. must be restated to the new risk horizon. ﬁnancial letters of credit. (For example.3 Computing credit risk on different horizons Two CreditMetrics modeling parameters must change to address different risk horizons: • the credit instrument revaluation formulas change to perform the revaluation computation for the alternate time horizon. autocorrelation prevents us from translating daily volatilities to monthly or yearly volatilities in a simple way. and • the likelihoods of credit quality migration.32 Chapter 2. a twoyear transition matrix could be obtained by multiplying the oneyear transition matrix with itself. and other credit assessment approaches are not necesCreditMetrics™—Technical Document . Thus. the choice of risk horizon is not likely to make an appreciable difference. any given risk horizon is likely to lead to the same qualitative decisions. Regrettably. Although these actions may differ among institutions. however. The key element to any risk information system is the resulting riskmitigating actions. a tendency for prices to return to some longterm stable level). etc. We conﬁrm this. as shown in the transition matrix. the risk horizon is not likely to be signiﬁcantly less than a quarter for a bank with loans. So too. 2. This issue of autocorrelation surfaces for market risk calculations also. A nonzero autocorrelation would indicate that successive credit quality moves are not statistically independent between adjoining periods. recalculating risk at a longer horizon can still provide guidance to changes in relative risk. Even if riskmitigating actions are performed daily. using as a convention a one year risk horizon – not unlike the convention of annualized interest rates – is common. For instance. One way of doing the latter is simply to multiply the shorthorizon transition matrices to obtain the transition matrix for a longer horizon. For instance. so a downgrade implies a higher likelihood of a downgrade in the following period.) Unfortunately.
Sec. Part I:. 2. In this chapter we have discussed the essence of the CreditMetrics methodology. We discuss these and other issues surrounding transition matrices in Chapter 6. Risk Measurement Framework .5 Choosing a time horizon 33 sarily subject to this problem. The next two chapters extend our framework across a portfolio of exposures and across different exposure types beyond a simple bond.
34 Chapter 2. Standalone risk calculation CreditMetrics™—Technical Document .
1 Our second “road map” of the analytics within CreditMetrics Value at Risk due to Credit Credit Rating Seniority Credit Spreads Present value bond revaluation Correlations Ratings series. senior unsecured. we point out that Bond #1 is the one for which we estimated the credit risk on a standalone basis in Chapter 2. and • we discuss the calculation of marginal risk estimation. we extend our methodology to a “portfolio” of two exposures. • we extend our credit risk calculation for standalone exposure (discussed in Chapter 2) to the multiple exposure case.35 Chapter 3. Equities series Models(e. ﬁveyear maturity • Bond #2: A rated. but in greater detail. Here. determine how often losses occur in multiple exposures at the same time. Note that there is one signiﬁcant addition. Our volatility of value – our risk – will be lower if . Chart 3. we must estimate joint likelihoods in the credit quality comovements. for example. The chapter is organized as follows: • we elaborate on the joint likelihoods in the credit quality comovements. senior unsecured. we explained the methodology used by CreditMetrics to obtain the credit risk for a standalone exposure. which identiﬁes overconcentrations within a portfolio and thus suggests potential riskmitigating actions. we discuss the required steps to calculate credit risk across a portfolio with an example portfolio consisting of the following two speciﬁc bonds: • Bond #1: BBB rated. 5% annual coupon. 6% annual coupon.1 for a standalone exposure. we discuss the same steps as in Chapter 1. Correlation will. The reader can compare this chart with the prior chapter’s corresponding Chart 2. We must now estimate the contribution to risk brought by the effects of nonzero credit quality correlations.g. correlations) Joint credit rating changes Rating migration likelihoods Recovery rate in default Standard Deviation of value due to credit quality changes for a single exposure Understanding joint likelihoods will allow us to properly account for the portfolio diversiﬁcation effects.. threeyear maturity This example portfolio is the same as the one that we considered in Chapter 1 when we were highlighting the steps in the calculation of portfolio credit risk. Portfolio risk calculation In Chapter 2. We now update the “road map” for CreditMetrics in Chart 3. Also. Thus. For clarity.1 to show the additional work needed to address a portfolio. Here.
00 0. The simplest way of obtaining the joint likelihoods is to just assume that these are the product of the individual likelihoods. as shown in Table 3.27 0. However.05 0.12 0. calculation will be true only for the simplest case where the two obligors’ credit rating changes are statistically independent. we will discuss several different methods for determining joint likelihoods of credit quality migrations.00 0.00 0.93% (the likelihood that the BBB rated bond remains a BBB) and 91. In Chart 3.02 0.06 0.00 79.74 0. we need to consider all possible combinations of states between the two obligors.00 0.30 5.02 0.00 0.00 0.00 0.33 5. in Chapter 8.01 Chance both Chance a BBB retain current rating remains at BBB By repeating this type of calculation for all the 64 states we then ﬁll the joint likelihood table shown below. BBB 5.00 0.15%.00 0.00 0.64 0.00 0. we will introduce in Chapter 8 a model which links ﬁrm asset value to ﬁrm credit rating.15 4. Portfolio risk calculation the correlation between credit events is lower.1 Joint migration probabilities with zero correlation (%) Obligor #1 (BBB) AAA 0.06 0. Later.08 0.95 86.00 0.00 0.05% Chance an A remains at A CCC 0.2 we illustrate a framework for thinking about default as a function of the underlying (and volatile) value of the ﬁrm.05 0.06 0.17 0.93% ⋅ 91.00 0.00 0.00 0.52 0.11 0.02 0. with the major ratings from AAA to CCC.00 0.01 0.00 B 0.05% (the likelihood that the singleA rated bond remains a singleA). Rather.03 0. This is the product of 86. Thus.09 0.00 0.00 0. and is often referred to as the option theoretic valuation of CreditMetrics™—Technical Document .00 0.00 A 91. This framework was ﬁrst proposed by Robert Merton (see Merton [74]).00 0.29 0.01 0. Now we are interested in two obligors considered together.00 0. However.00 AA 2.01 0.01 0. There are eight times eight or sixtyfour possible states to which the two credits might migrate at the risk horizon.14 1. we do not elaborate here on the connection between correlation and joint likelihoods.33 4.00 0.26 0.12 0.00 Assuming a zero correlation like this is too simplistic.80 0.98 0.04 0. we assume that the joint likelihoods are given to us. there are eight possible outcomes for an obligor’s credit quality in one year.42 79.00 0.00 0. 3.01 0.93 5.18 BB 0.23 0.00 0.1 below.01 0. the joint likelihood that the two obligors maintain their initial ratings is equal to 79.83 1.16 Obligor #2 (singleA) AAA AA A BBB BB B CCC Default 0.1 Joint probabilities We have seen that.02 0. Table 3.00 Default 0. For us to estimate this joint risk. In order to capture this effect.01 0.30 1.04 0.15% = 86. It is unrealistic since these credit movements are affected in part by the same macro economic variables.00 0.36 Chapter 3. We touch brieﬂy on this model here.
3.3 Model of ﬁrm value and generalized credit quality thresholds BBB BB B CCC Default Firm remains BBB A AA AAA Lower Value of BBB firm at horizon date Higher Part I: Risk Measurement Framework . We treat them as input parameters. It builds upon Black and Scholes option pricing model by stating that the credit risk component of a ﬁrm’s debt can be valued like a put option on the value of the underlying assets of the ﬁrm. Chart 3.Sec. The ﬁrm’s asset value relative to these thresholds determines its future rating. It is unimportant to CreditMetrics how default likelihoods are estimated. CreditMetrics assumes that each obligor will be labeled with a credit rating.2 Model of ﬁrm value and its default threshold Default threshold Default Scenarios Lower Value of the firm Higher We do not suggest here that default likelihoods must be estimated based on the volatility of underlying ﬁrm value. there are credit rating up(down)grade thresholds as well. Chart 3. The generalization involves stating that in addition to the default threshold. The Merton model can be easily extended to include rating changes. which in turn will be associated with a default likelihood.3. which we refer to as the default threshold. Under the Merton model. underlying ﬁrm value is random with some distribution.1 Joint probabilities 37 debt. as illustrated in Chart 3. If the value of assets should happen to decline so much that the value is less than amount of liabilities outstanding. then it will be impossible for the ﬁrm to satisfy its obligations and it will thus default.
09 0.00 0.00 0.00 0.04 BB 0. we need two types of information for each of the 64 joint states between two obligors: joint likelihoods and revaluation estimates. The effect of correlation is generally to increase the joint probabilities along the diagonal drawn through their current joint standing (in this case.95 86.06 0.52 0.17 0.00 0. 2.00 0. to calculate the volatility of value due to credit quality changes.02 0.30 asset correlation (%) Obligor #1 (BBB) AAA 0.00 A 91.07 0.02 0.19 0.01 0. this approach is developed in detail in Section 8.12 0.13 Obligor #2 (singleA) BBB 5.00 0.00 AAA AA A BBB BB B CCC Default 0.47 0.00 0.04 0. CreditMetrics™—Technical Document . The sum of each column or each row must equal the chance of migration for that obligor standing alone. 3.26 0.00 0.30 1.44 79. we show how we extend our credit risk calculation from the standalone exposure case to the multiple exposure portfolio case. The most likely outcome is that both obligors simply remain at their current credit ratings.00 0. the sum of the last row must be 0.04 0.55 0.39 1.2 Portfolio credit risk As mentioned in Chapter 1.18%.00 0.18 There are at least four interesting features in the joint likelihood table above: 1.00 0.2 Joint migration probabilities with 0.00 0.08 4.38 Chapter 3.57 0.00 0.05 0. With this discussion of the joint likelihood.00 0. we discuss the revaluation of the two exposures – given the credit quality migration – for each of the 64 states.00 0. the likelihoods of joint migration become rapidly smaller as the migration distance grows.01 0.00 0. 4.29 5.04 0.69 4.00 0. For instance.01 0.00 0.00 0.00 0.64 0. and can build the joint probabilities for two obligors from both this and a knowledge of the correlation between the two obligors’ ﬁrm values.00 0. Table 3.00 0. The probabilities across the table necessarily sum to 100%. we present joint likelihoods which result from an application of this model. Portfolio risk calculation In the end.4. Again. Here. In Table 3.09 0.02 0. In fact. In the previous section.33 5.27 0.01 0. 3. We will see that these data are then combined for two obligors in a fashion very similar to what we have already shown for one obligor.00 Default 0.18 0.00 0.00 0. we have a link between the underlying ﬁrm value and the ﬁrm’s credit rating.00 CCC 0.02 0.00 0.00 0.74 0.92 0.02 0. Speciﬁcally. we covered the joint likelihoods across comovements in credit quality. we turn our attention next to the credit risk calculation for our example twobond portfolio.81 0.02 0.93 5. through BBBA).11 0.00 0.01 B 0.00 0.04 0.2.00 AA 2. which is the default likelihood for Obligor #1 (BBB) in isolation.
At this point. we combine these variances with the variances for individual assets and arrive at a portfolio standard deviation. that is.5. It is only when either (or both) obligors suffer a downturn that the change in value becomes great.Sec. We next focus on the calculation of the two risk measures for the portfolio. We then consider each pair of assets as a subportfolio and compute its variance using the methods described in this section. we use the same formula in the standalone exposure case of Chapter 2. The only difference is that now we have 64 possible states rather than just eight in the standalone case. Part I: Risk Measurement Framework . although we have only discussed the calculation of standard deviation for a twoasset portfolio. Note from the calculation that the mean and standard deviation for the portfolio are $213. Assume that we are interested in calculating the 1st percentile level. 3. we ﬁrst identify all pairs of assets. Indeed. We illustrate this standard deviation calculation for the twobond portfolio below: S = 64 Mean: µ Total = [3.49. but the risk – as measured by standard deviations – is much less than the summed individuals due to diversiﬁcation. namely the standard deviation and the percentile level. the value is relatively ﬂat across most of Table 1. It is again noteworthy that the greatest potential for changes in value are on the downside..3 showing the frequency distribution of values. Also.55 and a standard deviation of $1. the percentile level. the above calculation ignores the additional contribution to the portfolio risk from the uncertainty (i. Finally.09 and $2. we next calculate a second measure of credit risk. For an arbitrary portfolio.35} i= 1 The probabilities and the values that enter the standard deviation calculation are read off Tables 3.22 2 2 {std.e. Again. This is easy since the value that the portfolio takes on in each pairwise credit rating class is simply the sum of the individual values. This same data was illustrated in Chart 1.1] Variance: 2 ∑ p i µ i = 213. The details of this calculation are discussed in Chapter 9 and Appendix A. For the singleA bond the comparable statistics are a mean of $106.63 and $3. Recall from Chapter 2 that the mean and standard deviation of our BBB bond were $107. we have presented all of the components necessary to calculate the standard deviation for any portfolio. After having calculated the portfolio standard distribution. As far as the portfolio standard deviation is concerned. the standard deviation) in the recovery rate value. for simplicity. is 3.99 respectively. we point out that there is no ﬁxed rule to prefer any given percentile level over another.35 respectively.63 S = 64 i= 1 σ Total = ∑ p i µ i – µ Total = 11.2 Portfolio credit risk 39 We now must determine the 64 possible values of the portfolio at the risk horizon. dev.2 and 1. The portfolio values in each possible joint rating state are given in Table 1. So we see that the means or expected values sum directly.5.5 respectively.
and with which exposures we beneﬁt due to diversiﬁcation. This is because of the diversiﬁcation effect that is in turn caused by the fact that the yearend values of the individual bonds are not perfectly correlated. they both serve the purpose of measuring an exposure’s risk contribution to a portfolio. In the next section. This percentile level is therefore $8. Since in this case. ﬁnancial letters of credit and marketdriven instruments such as swaps and forwards.2.63 and a 1st percentile level of 204. accounting for the effects of diversiﬁcation. we introduce the concept of marginal risk. Once the singleA rated bond is added. the more relevant calculation is the marginal increase to the portfolio risk that would be created by adding a new bond to it. The marginal standard deviation of this second bond is therefore equal to $0. others deﬁne it to be the marginal impact on portfolio risk of increasing an exposure by some small amount.99 below the mean.99.24) and the standalone risk ($3. Recall that the BBBrated bond had a mean value of $107. which is equal to $0.99.55.39 below the mean value of $106.15. the singleA rated bond has a 1st percentile level value of $103.23 and $8. While the two deﬁnitions do differ.40. loans. we show how CreditMetrics treats other asset types: receivables. Of course.10. Recall that the standard deviation of our onebond (BBBrated) portfolio in Chapter 2 was $2.09 and a 1st percentile level value of $98. This percentile level is $9. We can now calculate the marginal risk of the singleA rated bond as the difference between $9. Where we deﬁne marginal risk to be the contribution of one asset to the total portfolio risk. Let us ﬁrst illustrate the calculation of marginal risk by using the standard deviation as a risk measure.23 below the mean. This concept enables us to understand where the risks are concentrated in the portfolio.36.35 once we added the second.99. CreditMetrics™—Technical Document . On a standalone basis. This ﬁnishes our discussion of the calculation of the credit risk measures for the example portfolio of two bonds. This concludes our discussion of marginal risk. or lack thereof.49. the twobond portfolio has a mean of $213.35 and $2. which represents the difference between $3. Thus. the portfolio has no more than two assets. In the next chapter. it is not possible to calculate percentile levels analytically – we would have to perform a simulation. again with the caveat that this approach is most appropriate for large portfolios. singleA rated bond. We describe next how we extend our marginal risk calculation to percentile levels. loan commitments. We remark that marginal risk statistics are sometimes deﬁned in a slightly different way. we may simply examine the probabilities and values shown in Table 3. 3. which is $1. for larger portfolios.3 Marginal risk We saw in Chapter 2 how the credit risk can be calculated for an individual bond on a stand alone basis.40.23 below the mean value. and obtain a 1st percentile level number of $204.39) is again due to diversiﬁcation. The portfolio standard deviation increased to $3. Note that this marginal standard deviation is much smaller than the standalone standard deviation of the second bond. This is $9.24. the decision to hold a bond or not is likely to be made within the context of some existing portfolio. However. which is $3.40 Chapter 3. Portfolio risk calculation Recall from Chapter 1 and Chapter 2 that this calculation is quite simple. This difference between the marginal risk ($0. All we have to do is to ﬁnd the portfolio value such that the likelihoods of all the values less than this sum to 1%.
we have included the following generic exposure types: 1. . Rather. correlations) Joint credit rating changes Rating migration likelihoods Recovery rate in default Standard Deviation of value due to credit quality changes for a single exposure Portfolio Value at Risk due to Credit Before we describe how we treat different exposure types in CreditMetrics. We show below in Chart 4. here we only need to describe how to revalue each new exposure type in the event of a rating change or default. As a matter of implementation. CreditMetrics is capable of estimating most any credit risk type limited only by the data available to revalue exposures upon up(down)grade and default. noninterest bearing receivables (trade credit).. In Chapter 2.1 the ﬁnal “road map” for credit risk analytics within CreditMetrics. ﬁnancial letters of credit. the ﬁrst step is identical as for bonds.1 Our ﬁnal “road map” of the analytics within CreditMetrics 5. Chart 4. Thus. however. 4. 3. Equities series Models(e. As discussed in Chapter 1.) Exposures User Portfolio Market volatilities Exposure distributions Value at Risk due to Credit Credit Rating Seniority Credit Spreads Present value bond revaluation Correlations Ratings series. and This chapter explains how CreditMetrics addresses different exposure types. the analysis consisted of two steps: (i) specifying the likelihoods and joint likelihoods of obligors experiencing a credit quality change. Thus.2 for a variety of exposure types. 2. etc. Note that this map now includes the provision for different exposure types. commitments to lend.g. let us summarize the CreditMetrics analytics shown in the chart above. bonds and loans. For each of the additional exposure types covered in this chapter. In Chapter 3. we extended this standalone credit risk methodology to a portfolio of two bonds. forwards. the goal of this chapter is to provide the reader with methods for constructing versions of Table 1. Recall that for bonds. we discussed the case of a standalone bond exposure.41 Chapter 4. marketdriven instruments (swaps. CreditMetrics is not limited to bonds. Differing exposure types So far we have demonstrated the methodology used in CreditMetrics with the help of portfolios consisting of only bonds. and (ii) calculating the new values given each possible rating change at the risk horizon.
we treat receivables in the same way as we treated the bonds in the previous chapters. are at risk to changes in credit quality of their customers. or there is a default. • The joint likelihoods of credit quality movements for any pair of obligors is estimated through a treatment of the correlation between obligors. In concept.. We know of no systematic study of recovery rate experience for corporate receivables and so suggest that users take senior unsecured bond recovery experience as a guide. it is not even necessary to revalue in different rating categories. It is necessary. and revalue the cashﬂow based on the bond spreads in each rating category. FX or interest rates). Either the payment is made. In practice. where the exposure depends on the movement of market variables (e.). then. we revalue the exposure at $1mm in each nondefault state. in the case of a $1mm receivable due in nine months. • The transaction details and the forward zero rates for each credit rating category determine the changes in value of each transaction upon obligor up(down)grade. Thus. • The portfolio standard deviation of changes in value due to credit quality changes are then calculated directly – in closed form. and we revalue based on some recovery rate. etc.42 Chapter 4.000 (= 30% times $1mm) in default. where the risk horizon is one year and the recovery rate is. we must estimate this amount for: (i) commitments. In this situation. where the exposure may change due to additional drawdowns. Often. and (ii) marketdriven instruments (swaps. to consider such receivables on a comparable basis with any other risky credit instrument. If there are more applicable spreads available. we treat the cashﬂow as if it were a zero coupon bond paying on the receivable date. Differing exposure types • The credit exposure is the amount subject to either changes in value upon credit quality up(down)grade or loss in the event of default. forwards. speciﬁc to receivables. 4. CreditMetrics™—Technical Document . In the following sections. For receivables which become due beyond the risk horizon. it would certainly be reasonable to use these in place of the bond spreads for the purposes of this revaluation. and we “revalue” at the receivable’s face amount.1 Receivables Corporations which do not hold loans or bonds as exposures may still be subject to credit risk through payments due from their customers. • The obligor’s current longterm senior unsecured credit rating indicates its likelihood of credit quality migration that will be applied to all the obligor’s obligations. a receivable will be due before the risk horizon. we discuss how CreditMetrics treats different exposure types. • The seniority standing and instrument type of each transaction indicates the recovery rate of that exposure in default. and at $300. Non interest bearing receivables (also called trade credit).g. say 30%.
loan commitments can dynamically change the portfolio composition. we must therefore account for the changes in value to both portions. Part I: Risk Measurement Framework . and a fee is paid on the undrawn portion. and the increasing likelihood of repayment if the obligor is upgraded. we treat a loan as a par bond. the bond’s value was taken to be a recovery fraction multiplied by the face value of the bond. The drawn portion is revalued exactly like a loan. For example. it is quite appropriate for each institution to utilize its own pricing for this revaluation rather than using generic spreads in a downloadable data set.1 Fee on undrawn portion of commitment (b. The amount drawn down at the risk horizon is closely related to the credit rating of the obligor (see Table 4. Interest is paid on the drawn portion. Typical fees on the undrawn portion are presented in Table 4. The drawdown on the loan commitment is the amount currently borrowed. In concept. we discuss the differences between the two approaches. if an obligor deteriorates. if its prospects improve. it is likely to draw down additional funds. To this we add the change in value of the undrawn portion. As a practical matter.3 Loan commitments 4.2 Bonds and loans A loan commitment is composed of a drawn and undrawn portion. When we revalue a loan commitment given a credit rating change. Thus. bonds may be treated in the same way as the marketdriven instruments described in Section 4. As an alternative. each lending institution is the best judge of its own pricing for loan commitments. 4. We generated a forward curve based on the credit spread curve for that rating.1. In the case of default. This revaluation upon up(down)grade accounts for the decreasing likelihood that the full amount of the loan will be repaid as the obligor undergoes rating downgrades. and discounted the future cash ﬂows of the bond. it is unlikely to need the extra borrowings.Sec. Table 4.2 Bonds and loans 43 In Chapter 2.) Yearend rating AAA AA A BBB BB B CCC Fee: undrawn portion 3 4 6 9 18 40 120 Because loan commitments give the obligor the option of changing the size of a loan.2). 4. In that section. revaluing the loan using loan forward curves upon up(down)grade and applying a loan recovery rate to the principal amount in the case of default. On the other hand.4.p. we described how to revalue bonds in each future rating state using the forward interest rate curves for each credit rating.
the commitment’s revaluation estimate at the risk horizon will depend on both: • estimates of the change in amount drawn due to credit quality changes. The worst possible case for a commitment is that the counterparty draws down the full amount and then defaults. we will assume the credit spreads of Section 1. Asarnow & Marker (A&M) [95] have examined the average drawdown (of normally unused commitment) in the event of default (see Table 4. given one source of data on commitment usage. it is necessary to estimate not only the amount of the commitment which will be drawn down in the case of default.44 Chapter 4. the change in value of a commitment in each possible credit rating migration. In order to model commitments more accurately. we present a framework to calculate. In each credit rating. It is intuitive. We will address each of these in turn. In practice. but this is certainly an area where each ﬁrm should apply its own experience. Differing exposure types Note that it is not uncommon for loan commitments to have covenants that can reduce the credit risk. Part of this estimation. to treat a commitment as if it were a loan. and hence. Thus. the most conservative. and • estimates of the change in value for both the drawn and undrawn portions. CreditMetrics™—Technical Document . if the loan rate not only ﬂoats with interest rate levels but also has credit spreads which change upon up(down)grade – a repricing grid – then the value of the facility will remain essentially unchanged across all up(down)grade categories. but also the amount which will be drawn down (or paid back) as the counterparty undergoes credit rating changes. it has been seen that commitments are not always fully drawn in the case of default. with principal equal to the full commitment line. we will rely on one study of commitment usage. For example. We may use this information to estimate how much of the undrawn amount of our commitment will be drawn in the case of a default. and from a risk perspective. it will draw an additional 71% of the $80mm undrawn amount.2. Thus. so that the change in value of any drawn amount due only to changes in credit quality (and neglecting changes in commitment usage) will be the same as for bond example in that section. In the remainder of this section.p. For ease of illustration. Our ﬁrst task is to estimate changes in drawdown given each possible credit rating change.8mm. can be directly taken from a published study. We do this via an example.8mm. that the risk on a commitment is less than the risk of a fully drawn loan. Our example will have $20mm currently drawn down with the remaining $80mm undrawn and charged at a fee of 6 b.2).3. if our counterparty defaults. For purposes of explanation in this section. then. this results in an estimated drawdown in default of $76. Added to the current drawn amount of $20mm. Consider a threeyear $100mm commitment to lend at a ﬁxed rate (on the drawn portion) of 6% to a currently A rated obligor. the risk will have been reduced because the only volatility of value remaining will be the potential loss in the event of default. This is certainly the simplest approach to commitments. or $56. the drawdown in default.
6% 4. the undrawn portions moved from 95. A&M found that the average draw increased from 4. In other words.[80. Referring to Table 4.0 20.0 20.9 55. One suggestion of this behavior is also evidenced in the A&M study. consider the change to BBB from our initial rating of singleA.6% 20.4 7.3.0 20.6 59.3 Loan commitments 45 Table 4.0 20. As an example. we see for instance that when the obligor downgrades to BB.0% between singleA and BBB.0 56.3 Example estimate of changes in drawdown Yearend rating AAA AA A BBB BB B CCC Default Current Drawdown 20. each institution will substitute its own study based on its unique experience.0 20.2 Average usage of commitments to lend Credit rating AAA AA A BBB BB B CCC Average commitment usage 0. We present estimates of change in drawdown for all possible migrations in Table 4.8 It is against this estimate of the new drawdown amounts at the risk horizon that we now apply our revaluation estimates – to both the drawn and undrawn portions.9% (=16. Table 4. with our example initial drawdown of 20%.0 0. We see that the average commitment usage varies directly with credit rating.6 79.7% 75.0 Change in Drawdown 19.4% to 80. we estimate that an additional 12.0% 46. 4.4%]). We offer the following method as a suggestion to initiate discussion rather than as a deﬁnitive result.4.4 69.0 32.0 20.4 49. The following table estimates the changes in drawdown attributable to credit rating changes.0% or a reduction of 16.1% (= 100% .0% Usage of the normally unused commitment in the event of default 69% 73% 71% 65% 52% 48% 44% Source: Asarnow & Marker [95] We also expect that there will be a credit rating related change in drawdown in all nondefault states.Sec.9 35.8 Estimate of New Drawdown 0. corresponding to the counterpart paying back some amount. We fully anticipate that.6% to 20.0 20. the change in Part I: Risk Measurement Framework . Thus.0%/95. How to best apply this information is open to question. we actually estimate a negative change in drawdown.1% 1.0 20.6 13. Note that in cases of upgrades.0 12.1% * 80%) will be drawn down in the case of a migration to BBB. as a matter of implementation.8% 63.0 76.
7%. we collect fees on $35.6 0. • it is possible to have negative revaluations greater than the current drawdown.) In default.04 0. an outright loan would be inefﬁcient since its proceeds may sit in the obligor’s hands underemployed.4 7.8 Change in fees ($mm) 0. as seen in Table 4.0 0.94 39.3 3. Thus in the case of downgrade to BB. The change in value of the fees1 is also given in Table 4. the full 6 b.01 0.0 0. or $1. Recall that at present. but we ignore this effect for this example.5mm due to credit spread widening on the ﬁxed rate dawn portion.3 0.01 0. The calculation of credit risk for the commitment is now simply a matter of applying the techniques of the previous two chapters.0 76.11 1.03 12.p. as well as the total change in value for the commitment. and • covenants that reset the drawdown spread upon an up(down)grade would reduce the volatility of value in all nondefault states – keeping value close to par.52 3. (This $35.1 2. so too do the fees we collect. we are collecting 6 b. What is typically needed in this case is a ﬁnancial letter of credit. Table 4.00 0.05 Total Value Change ($mm) 0.5 4. Differing exposure types value is negative 2.85 Several observations are in order: • the expected percentage drawn down in default is the most important factor.8 39.7 1.4mm less. Thus we have obtained the revaluations estimates in each future credit rating state.0 20.01 0.02 0.4 69.3 12. With this type of off balance sheet access to 1 To be precise.8 Change in Change in Value (%) Value ($mm) 0.03 0.5 0.9 51. In such a case. CreditMetrics™—Technical Document .10 0.4mm corresponds to the additional draw in this case.4 Financial letters of credit (LCs) There are times when an obligor may desire to have the option to borrow even if there is no immediate need to borrow.1 0.3. As the undrawn amount changes. we lose the all of the fees which we were receiving.9 55. on the full $80mm.p.4. • fees have a relatively small impact on the revaluations. 4.46 Chapter 4. for all possible rating migrations. the change in fees should also account for a new discount function corresponding to the new rating.0 32.6 79.3 0. on the undrawn amount of $80mm. There is also a change in the value of the fees collected on the undrawn amount.01 0.0 16.01 0.4 Revaluations for $20mm initially drawn commitment Yearend rating AAA AA A BBB BB B CCC Default Drawdown at Yearend 0.
Sec. that swaps based on the same interest rate would tend to go in. We argue that such an exposure is comparable in risk to an outright term loan Indeed. the market typically prices these instruments comparable to loans. etc. that is.1 Credit risk calculation for swaps Swaps are treated within CreditMetrics consistent with the way bonds and loans are treated. bonds. Our discussion focuses on the example of swaps. but to capture the most crucial inﬂuences of market volatilities to the credit risks of marketdriven instruments.e.). 2. This optionality stems from the fact that we face a loss on the transaction if the counterparty defaults only if we are inthemoney (i. In these transactions.5. credit risk and market risk components are intimately coupled because of an inherent optionality. We suggest that LCs be treated identically to loans. 4. We remark that to treat products like swaps in full detail. Thus.5 Marketdriven instruments We have so far described how CreditMetrics calculates the credit risk for receivables. The swap transaction is outofthemoney for the counterparty. there will also be full exposure just like a loan. the counterparty owes money on the swap transaction on a net present value basis. loans.or outofthemoney together). The obligor will almost assuredly drawdown the LC as it approaches credit distress. we explain how CreditMetrics treats derivative instruments that are subject to counterparty default (swaps. In this section. Part I: Risk Measurement Framework . that swap counterparties might be more likely to default in one interest regime than in another). forwards. 4. for instance. the revaluation of swaps in each credit quality state at the risk horizon is much more complicated than that of either bonds or loans. The counterparty undergoes a credit quality change. the obligor owes us money on a net present value basis)..5 Marketdriven instruments 47 funds. and the correlations between credit and market moves (for instance. there is assurance that funds will be available even when other sources of funding may dry up due perhaps to credit quality deterioration. However. This complicates CreditMetrics’ handling of derivatives exposures. we refer to these types of exposures as marketdriven instruments. and commitments. Note that a ﬁnancial letter of credit is distinguished from either performance or a trade letter of credit. Both of these would have smaller risk than ﬁnancial LCs. Such a model would describe both the correlations of swap exposures across a portfolio (capturing. in all the cases where there can be a default. Credit loss occurs when both of the following two conditions are satisﬁed: 1. including the use of the credit spread curves and recovery rates that have been estimated for loans. 4. Performance LCs are typically secured by the income generating ability of a particular project and trade LCs are triggered only infrequently by non credit related events. Throughout this document. it would be necessary to propose an integrated model of credit and market risk. Our goal here is not to provide a fully integrated model.
the revaluation of the swap in any rating category is obtained by subtracting the second (expected default loss) component from the ﬁrst (riskfree value) component. The ﬁrst component is equal to the forward riskfree2 value of the swap cash ﬂows.e.4 for bonds. but also for the random nature of the swap exposure. using the government (i.. The probability of default is obviously an important factor driving this latter component. An enhancement of this procedure might be to account for not only the expected loss due to credit.48 Chapter 4. and then subtracts a penalty (the expected loss) to account for the risk due to the credit quality of the counterparty. However.” In other words.e. 2 Here we mean riskfree from a credit risk perspective. Differing exposure types The market and credit risk calculations for swaps are therefore intimately related. Note that this valuation scheme essentially values the swap as if it were riskfree. the second component varies from one rating category to another. The result would be that two swaps with the same expected losses in each rating state would be distinguished by the amount of uncertainty in their losses. To summarize. A second factor is optionality. We then subtract from this creditriskfree value the amount that we can expect to lose due to a counterparty default. the net present value for the issuer is always negative. Essentially. which is implied in the fact that we have exposure to the counterparty only if the counterparty is inthemoney. By “remaining” we mean all cash ﬂows that occur after the risk horizon (assumed to be one year). The purpose of this exercise is to ﬁll the “value” table analogous to Table 2. Finally. The intuition behind this procedure for calculating the swap value is straightforward. All things remaining equal. 2. the greater the amount exposed to loss during an unfavorable credit event. Since the probability of this default varies by rating category.. creditriskfree) rates for this calculation. credit exposure to the swap counterparty only if the counterparty is outofthemoney) is the feature that makes swaps distinct from bonds. therefore the ﬁrst component is the same for all forward credit rating states. This could be achieved by redeﬁning the expected loss penalty to include a measure of how much the swap is likely to ﬂuctuate in value. the net present value for swaps can be either positive or negative for the counterparties. and also from the volatility of interest rates. The value of this optionality is determined from the amount by which the swap is expected to be inthemoney. We next describe how the swap is reevaluated in each possible rating state at the risk horizon. First we calculate the value assuming that there is no risk whatsoever of the counterparty’s default. there is no optionality involved. This hypothetical value is obtained by ﬁnding the forward value of the swap cash ﬂows by using the government rates rather than the swap rates. CreditMetrics™—Technical Document . since the issuer of debt is always “outofthemoney. the greater the market volatility. we represent the value of the swap as a difference of two components: 1. The second component represents the loss expected on the swap due to a default net of recoveries by the counterparty on the remaining cash ﬂows of the swap. optionality (i. Although the exposure in the case of bonds is also marketdriven.
let us restate mathematically the revaluation of the swap: [4. As a result.P. The calculation of the expected loss. on the other hand. Each of the expected exposure values that enter the average exposure calculation requires a modiﬁed BlackScholes computation to account for the inherent optionality feature. One such source is “On measuring credit exposure. As mentioned before.5 Marketdriven instruments 49 We next detail the calculation of the two components of the swaps value at the risk horizon in each possible credit rating state..Sec. First. All we need do is discount the future swap cash ﬂows occurring between Year 1 through maturity by the forward government zero curve. if the maturity of the swap is ﬁve years. then the fouryear probability of default is required for each of the rating categories AAA through CCC. March 1997. For example. The fouryear default probabilities can then be simply read off from the last (i. Second. the average exposure calculation for swaps is quite complicated and timeconsuming.P. Two assumptions are implicit in this method of generating the longterm default probabilities. Morgan plans to provide a software tool in the near future that enables the user to calculate average and expected exposures.e. the calculation of the riskfree value of the swap in one year is straightforward. We use average exposure in the expected loss expression above to account for the possibility of swap counterparty defaulting at any point in time between the end of the ﬁrst year and maturity. J.P. we assume that there is no autocorrelation in rating movements from one year to another. This is because of the optionality component in swap exposure explained earlier. Part I: Risk Measurement Framework . default) column of this transition matrix. we assume that the transition process is stationary in that the same transition matrix is valid from one year to another. however. For each forward nondefault credit rating. is complicated. the expected loss can be written as: [4. 3 4 The expected exposures are weighted by the appropriate discount factors for this average calculation.1] Value of swap in 1 yearRating R = Riskfree value in 1 year – Expected loss in years 1 through maturityRating R where “R” in the above expression can be any possible credit rating category including default. J.2] Expected lossRating R = Average exposureYear 1 through maturity · Probability of default in years 1 through maturityRating R · (1–Recovery Fraction) The average exposure represents the average3 of several expected exposure values calculated at different forward points over the life of the swap starting from the end of ﬁrst year. This tool will be based on the RiskMetrics market risk methodology. Morgan under the name FourFifteen. a software implementation of which is currently being marketed by J. First.” RiskMetrics™ Monitor. Also. We refer the interested reader to other sources for a more thorough treatment of the expected and average exposure calculations.4 The second term that enters the expected loss calculation is the probability of default for each rating category between Year 1 and the maturity of the swap. These probabilities can be obtained by multiplying the oneyear transition matrix four times to generate the fouryear transition matrix. 4. Morgan.
5) = FV – $6 where FV refers to the forward value.0002 · 61. Let us next consider an example. Given a recovery rate of 50%. Let the risk horizon be one year and the recovery rate in case of default be 0. say one year. This represents an average of the expected exposures between the end of one year and the end of three years (a twoyear time period). for two reasons: CreditMetrics™—Technical Document . and AE refers to the average exposure in one year. 1997. Similarly. the average exposure is a more suitable measure. the value of the swap at the end of risk horizon in the AA rating category is equal to: [4. Therefore we write the expected loss in the defaulted state as: [4.50. To calculate the value in case of a default during the risk interval. The expected exposure at the end of the year is calculated on January 24.5) = FV – $50. we must modify this procedure somewhat.5 we summarize the value of the swap in each possible credit rating states at the risk horizon.24% default likelihood for the CCC rating category. say several years.627 · (1 – 0. This is mainly due to the fact that the average exposure calculation over the life of the swap does not make any sense here. since we know for sure that the swap counterparty has defaulted during the risk interval. 1997. given a 33.4] FV in 1 year – pAA · AE · (1 – R) = FV in 1 year – 0. This is because the swap counterparty can default at any point over the longer risk interval. the average exposure at the end of one year is calculated to be equal to be $61. Now.50 Chapter 4. let us consider what happens in default. it will be more accurate to replace the expected exposure with the average exposure value calculated over the risk interval.5] FV in 1 year – pCCC · AE · (1 – R) = FV in 1 year – 0.3] Expected lossDefault = Expected exposureYear 1 · (1 – Recovery Fraction).721. Assume a threeyear ﬁxed for ﬂoating swap on $10 mm notional beginning January 24. the value in the defaulted state is equal to: [4.304. Next. as compared to the maturity of the swap. In this case. If the risk interval is much longer than this.02% for the AA rating category.627.5) = FV – $10. We do not specify the riskfree component (FV).860 where EE refers to the expected exposure in one year.6] FV in 1 year – EE · (1–R) = FV in 1 year – 101.3344 · 61. This concludes our explanation of the value calculation for swaps at the risk horizon. the corresponding value of the swap in the CCC rating category is equal to: [4. and the expected exposures at these points can be very different from the expected exposure at the risk horizon. 1997 to be equal to $101. therefore.627 · (1 – 0. In Table 4. Differing exposure types The method provided so far enables us to calculate the value of the swap in each of the nondefault credit rating categories.721 · (1–0. The implicit assumption in the above expression is that the risk interval is relatively short. On January 24. given a twoyear default likelihood of 0.
we can more accurately calculate the expected loss component as follows: 1.48 2.24 — 2. Value ($) FV – 1 FV – 6 FV – 46 FV – 148 FV – 797 FV – 3. 3. it sufﬁces to set the riskfree value term to zero and just use the expected loss term in the credit risk calculation. the average exposure represents the average of the expected exposures calculated at several points between Year 1 and maturity.15 0. Weigh each of the expected exposures by a probability factor. the expected loss value far exceeds the riskfree forward value of the swap itself. at least for the lower credit ratings and default.5 Value of swap at the risk horizon in each rating state “FV” represents the riskfree forward value of the swap cash ﬂows in one year.02 0. This factor represents the probability that the counterparty defaults in the year in which the expected exposure is calculated. Yearend rating AAA AA A BBB BB B CCC Default Twoyear default likelihood (%) 0. More importantly. we properly account for the timing of default. Part I: Risk Measurement Framework . Given this. This is especially true when the swap value is near par.5 Marketdriven instruments 51 1.860 We next discuss a reﬁnement to the calculation of the expected loss value in rating states AAA through CCC that produces more accurate expected loss numbers. we now explain it below. oneyear increments between Year 1 and maturity. Also.41 33.59 10.304 FV – 50. Recall that we calculate the expected loss value for rating states AAA through CCC by multiplying the average exposure by the probability of default for the desired rating category.Sec. In this circumstance. The result is a expectedloss calculation which reﬂects reality more accurately than if we were simply to multiply the average exposure by a single default probability. 2. the risk interval is quite small. This is because by thus breaking the expected loss calculation into smaller pieces at different time horizons.00 0. 4. Add these weighted expected exposures after adjusting for the time value of money effect. Table 4. Both the average exposure and the default probability are valid from Year 1 through the maturity of the swap. For the sake of clarity we did not address it earlier.209 FV – 10. and the interest rates are not changing too rapidly. it is conceivable that. say. given that it does not default before then. This calculation is relatively straightforward and involves valuing the future cash ﬂows with the riskfree yield. Calculate the expected exposure in.
2 Extension for forwards and multiple transactions The methodology that we have presented above for swaps can be used in exactly the same manner for forwards. First.5.52 Chapter 4. (Of course. both of which are calculated in exactly the same manner as for swaps. it can be easily extended to the case in which there are several transactions with the same counterparty and netting is enforceable. CreditMetrics™—Technical Document .) Next. Differing exposure types 4. this value comprises riskfree forward value and a expected loss value. the swaps methodology is used to revalue these net cash ﬂows in different rating categories at the risk horizon. all the cash ﬂows from the different transactions conducted by the same counterparty are netted to yield the resulting net cash ﬂows. including forwards. Once again. but can represent a variety of marketdriven instruments. this netting is done according to the particular netting arrangements that are in place with the counterparty. These transactions do not all have to be swaps. Furthermore. The methodology outlined for swaps can be extended to a portfolio of different instrument types with the same counterparty as follows.
53 Model Parameters Part II .
54
CreditMetrics™—Technical Document
55
Overview of Part II
We have seen in the previous section the general overview, scope and type of results of CreditMetrics. Now we will give more detail to the main modeling parameters used in the CreditMetrics calculation: our sources of data, how we use the data to estimate parameters and why we have made some of the modeling choices we did. There is no single step in the methodology that is particularly difﬁcult; there are simply a lot of steps. We devote a chapter to each major parameter and have tried to present each chapter as a topic which can be read on its own. Although we encourage the reader to study all chapters, reading only a particular chapter of interest is also possible. Part II is organized into four chapters providing a detailed description of the major parameters within the CreditMetrics framework for quantifying credit risks. Our intent has been to make this description sufﬁciently detailed so that a practitioner can independently implement this model. This section is organized as follows: • Chapter 5: Overview of credit risk literature. To better place our efforts within the context of prior research in the credit risk quantiﬁcation ﬁeld, we give a brief overview of some of the relevant literature. • Chapter 6: Default and credit quality migration. We present an underlying model of the ﬁrm within which we integrate the process of ﬁrm default and, more generally, credit quality migrations. We argue that default is just a special case of a more general process of credit quality migration. • Chapter 7: Recovery rates. Since changes in value are – naturally – greatest in the state of default, our overall measure of credit risk is sensitive to the estimation of recovery rates. We also model the uncertainty of recovery rates. • Chapter 8: Credit quality correlations. The portfolio view of any risk requires an estimation of – most generally – joint movement. In practice, this often means estimating correlation parameters. CreditMetrics requires the joint likelihood of credit quality movements between obligors. Since the observation of credit events are often rare or of poor quality, it is difﬁcult to further estimate their correlations of credit quality moves. We show that the results of several different data sources corroborate each other and might be used to estimate credit quality correlations.
Part II: Model Parameters
56
CreditMetrics™—Technical Document
57
Chapter 5.
Overview of credit risk literature
One of our explicit goals is to stimulate broad discussion and further research towards a better understanding of quantitative credit risk estimation within a full portfolio context. We have sought to make CreditMetrics as competent as is possible within an objective and workable framework. However, we are certain that it will improve with comments from the broad community of researchers. Extensive previous work has been done towards developing methodologies for estimating different aspects of credit risk. In this chapter, we give a brief survey of the academic literature so that our effort with CreditMetrics can be put in context and so that researchers can more easily compare our approach to others. We group the previous academic research on credit risk estimation within three broad categories: • estimating particular individual parameters such as expected default frequencies or expected recovery rate in the event of default; • estimating volatility of value (often termed unexpected losses) with the assumption of bond market level diversiﬁcation; and • estimating volatility of value within the context of a speciﬁc portfolio that is not perfectly diversiﬁed. Also, there have been several papers on credit pricing, starting with Merton [74], which discuss debt value as a result of ﬁrm risk estimation in an optiontheoretic framework. There is more recent work in this area which has focused on incorporating corporate bond yield spreads in valuation models, see Ginzburg, Maloney & Willner [93], Jarrow, Lando & Turnbull [96] and Das & Tufano [96]. For CreditMetrics, we have chosen to focus on the risk assessment side rather than focus on the pricing side. 5.1 Expected losses
Expected losses are driven by the expected probability of default and the expected recovery rate in default. We cover recovery rate expectations in much more detail in Chapter 7 and so will devote this discussion to the expected default likelihood. The problem of estimating the chance of counterparty default has been so difﬁcult that many systems devote all their efforts to this alone. Certainly, if the underlying estimates of default likelihood are poor, then a risk management system is unlikely to make up for this deﬁciency in its other parts. We will discuss three approaches that are used in practice: • the accounting analytic approach which is the method used by most rating agencies; • statistical methods which encompass quite a few varieties; and • the optiontheoretic approach which is a common academic paradigm for default. We emphasize that CreditMetrics is not another rating service. We assume that exposures input into CreditMetrics will already have been labeled into discrete rating categories as to their credit quality by some outside provider.
Each credit rating label must have a statistical meaning such as a speciﬁc default probability (e.52 1.1 These raw quantitative measures are then tempered by the judgment and experience of an industry specialist.00 0.28 23.g. domestic issuers who are large enough to have at least an S&P or Moody’s rating.1. [87].95 32.00 0. A quantitative credit risk model such as CreditMetrics cannot utilize ratings without additional information. letter (or numerical) rating categories by themselves only give an ordinal ranking of the default likelihoods. S&P and Moody’s.10 27. These studies are indispensable.46 6.26 1.06 0. coupled with an analysis of the quality and stability of the ﬁrm’s earnings and cash ﬂows.00 0.33 22.4).55 23.35 1.S. and it is important to highlight some important points from them: • the evolution and change in the original issue high yield bond market is unique in its history and future high yield bond issuance will be different.42 0. Indeed.. For some users with their own internal rating systems.04 0.20 0. 0.00 0. Altman & Bencivenga [95]. Asquith.10 Table 5.04 0.50 8.07 2.00 0.06 0. and • the deﬁnition of “default” has itself evolved (e. Source: Carty & Lieberman [96a] — Moody’s Investors Service 1 See: http://www.ratings.1 Accounting analytic approach Perhaps the most widely applied approach for estimating ﬁrm speciﬁc credit quality is fundamental analysis with the use of ﬁnancial ratios.75 13.69 9.15 29. 5 0.00 0. have published historical default likelihoods for their letter rating categories.g.73 3.00 0.49 2 0. use of these data must be accompanied by a working knowledge of how they were generated and what they represent.com/criteria/index.. then an estimation algorithm that expresses longterm behavior may be desirable (see Section 6.39 0.00 0.29 0.00 0. An example from Moody’s is shown in Table 5.g. Thus. this will be a necessary ﬁrst step before applying CreditMetrics to their portfolios.09 0. 5.. • most of the default history is tagged to U. However.47 13. it is common that ﬁnancial institutions will maintain their own inhouse credit rating expertise.standardpoor.htm CreditMetrics™—Technical Document . Altman & Nammacher [85]. Carty & Lieberman [96a] and S&P CreditWeek [96]. Overview of credit risk literature As we discuss in Chapter 6.86 3 0.15 0. we would argue that each credit scoring system should be ﬁt with its own transition matrix.79 0.57 0.90 15.70 3.S.53 1.31 0.07 0.05 31. [88].60 1.37 0. We are aware of at least 35 credit rating services worldwide. A good statement of this approach is in Standard and Poor’s Debt Rating Criteria.80 4.49 0.85 0.45% over a oneyear horizon).12 4.58 Chapter 5.00 0. If these systems have limited historical data sets available. Mullins & Wolff [89]. These include Altman [92].18 21.06 2.23 0. Such accounting analytic methods focus on leverage and coverage measures. alphabetic ratings) which label ﬁrms by credit quality.49 17.1.00 0.44 27. Also.37 8.61 1.27 0. agencies. it now typically includes “distressed exchanges”).17 1.08 The two major U.86 36.01 0.03 9.45 0. This approach yields discrete ordinal groups (e. a transition matrix for use by CreditMetrics can be ﬁt to any categorical rating system which has historical data.79 0.31 0.65 0.87 7.09 10.1 Moody’s corporate bond average cumulative default rates (%) Years Aaa Aa1 Aa2 Aa3 A1 A2 A3 Baa1 Baa2 Baa3 Ba1 Ba2 Ba3 B1 B2 B3 1 0.52 12.68 3. There have been many studies of the historical default frequency of corporate publicly rated bonds. This broad description is generally the approach of the major debt rating agencies.88 0.43 18. Altman & Haldeman [92].09 0.84 4 0.21 0.
One can identify three basic approaches to estimating default likelihood: qualitative dependent variable models. which seek to predict future default.1 Expected losses 59 On a more macroeconomic level. This approach has served as an academic paradigm for default risk. In this view. and most notably. The leading commercial exemplar of this approach is KMV. and Tyree & Long [94]. which – with additional processing – can be mapped to default likelihoods.2 Statistical prediction of default likelihood There is a large body of more statistically focused work devoted to building credit quality estimation models. 5.Sec. Default occurs when the value of the ﬁrm falls so low that the ﬁrm’s assets are worth less than its obligations. ﬁrst developed in 1968 and now offered commercially as Zeta Services Inc. Linear discriminant analysis applies a classiﬁcation model to categorize which ﬁrms have defaulted versus which ﬁrms survived. All of these methods can be shown to have some ability to distinguish high from low default likelihoods ﬁrms. and Episcopos. Ingersoll. The application of neural network techniques to credit scoring include Dutta & Shekhar [88]. 5. Pericli & Hu [95]. The popular press reports commercial applications of neural networks to large volume credit decisions such as credit card authorizations. In general. discriminant analysis. manufacturing hours. The best example of this approach is Edward Altman’s Zscores. Then. and subsequently developed by Black. Authors who compare the predictive strength of these diverse techniques include Alici [95]. Altman.1. In this approach. this method yields a continuous numeric value such as the number of standard deviations to the threshold of default. but it is also used as a basis for default risk estimation. Marco & Varetto [93]. which – with additional processing – can be mapped to default likelihoods Part II: Model Parameters .3 Optiontheoretic approach The optiontheoretic approach was proposed by Fisher Black and Myron Scholes in the context of option pricing. a historical sample is compiled of ﬁrms which defaulted with a matched sample of similar ﬁrms that did not default. researchers have found that aggregate default likelihood is correlated with measures of the business and credit cycle. logistic regression. and neural networks. while Jónsson & Fridson [96] examine also corporate proﬁts. Kerling [95]. the statistical estimation approach is applied to identify which variables (and in which combination) can best classify ﬁrms into either group. All of these approaches are strictly quantitative and will at least yield a ranking of anticipated default likelihoods and often can be tuned to yield an estimate of default likelihood. a ﬁrm has a market value which evolves randomly through time as new information about future prospects of the ﬁrm become known. but there do not appear to be commercial application yet of these neural network techniques for large corporate credits. selforganizing feature maps. etc. The academic literature is full of alternative techniques ranging from principal components analysis. 5. probit/logit analysis and hierarchical classiﬁcation models. money supply. This approach yields a continuous numerical score based on a linear function of the relevant ﬁrm variables. Fons [91] correlates aggregate defaults to GDP. For example.1. Cox. Robert Merton.
the hope would be that statistics drawn from observing the index of debt might be applied through analogy to the speciﬁc portfolio.2 Credit quality migration likelihoods for a BBB in one year BBB AAA 0.93% of the time a BBBrated obligor will remain a BBB. several practitioners (see Austin [92].33% A 5. As illustrated in Chart 5. 5. Lando & Turnbull [96] and Das & Tufano [96]). knowing today’s credit rating allows us to estimate from history the possible pattern of behaviors in the coming period. Also. The ﬁrst publication of transition matrices was in 1991 by both Professor Edward Altman of New York University and separately by Lucas & Lonski of Moody’s Investors Service. There have been studies of their predictive power and stationarity (Altman & Kao [91] and [92]). but it may be up(down)graded. Overview of credit risk literature 5. They have since been published regularly (see Moody’s Carty & Lieberman [96a] and Standard & Poor’s Creditweek [15Apr96]) and can be calculated by ﬁrms such as KMV.30% chance that a BBB will downgrade to a BB in one year. Table 5. but there is a 5.4 Migration analysis Chart 5. then chances are the obligor will be BBB in one year’s time. Second.. We have since built upon a broad literature of work which applies migration analysis to credit risk evaluation.1. they have applied methods which are statistically easy by addressing either the special case of correlations all equaling zero (perfectly uncorrelated) or correlations all equaling one (perfectly positively correlated).95% BBB 86. First.2 Unexpected losses The volatility of losses.93% BB 5. In CreditMetrics. Morgan developed transition matrices for our own use as early as 1987.02% AA 0. these tools have been used to both estimate (Crabbe [95]) and even potentially improve Lucas [95b]) holding period returns. Since it is so difﬁcult to explicitly address correlations there have been a number of examples where practitioners take one of two approaches. 86.60 Chapter 5. More speciﬁcally. The institution’s portfolio would be assumed to have the same CreditMetrics™—Technical Document .30% B 1.1. we extend this literature by showing how to calculate the volatility of value due to credit quality changes (i. More recently. the potential magnitude of unexpected losses) rather than just expected losses. commonly termed unexpected losses. Neither of these is realistic. academics have constructed arbitrage free credit pricing models (see Ginzburg. if an obligor is BBB today. Jarrow.12% Default 0. they have taken a middle road and assumed that their speciﬁc portfolio will have the same correlation effects as some index portfolio. and Smith & Lawrence [95]) have used migration analysis to better estimate an accountingbased allowance for loan and lease losses (what we would term expected default losses).18% One of our fundamental techniques is migration analysis. Meyer [95].17% CCC 0. Maloney and Willner [93]. Table 5.2 shows that. The index portfolio can either be the total credit market (“full” diversiﬁcation) or a sector index.e. has proven to be generally much more difﬁcult to estimate than expected losses. Finally.1 Credit migration AAA AA A BBB BBB BB B CCC Understanding the potential range of outcomes that are possible is fundamental to risk assessment. for instance. Thus.
4 3. there are two studies by BenPart II: Model Parameters .2 Unexpected losses 61 correlations and proﬁle of composition as the overall credit markets. For corporate bonds. and • volatility of holding period returns.1 0. Thus.8 1. and • the business cycle sees more ﬁrms default during downturns versus growth phases.3 1. they both suffer from an inability to do meaningful marginal analysis. There would also be no guide to know which speciﬁc names contribute disproportionate risk to the portfolio.3 Volatility of historical default rates by rating category Credit rating Aaa Aa A Baa Ba B Default rate standard deviations (%) Oneyear Tenyear 0. • the volume of high yield bond issuance across years is uneven. A number of academic studies have performed this exercise. 5. Table 5.3.1 Historical default volatility Historical default volatility is available from public studies: see for example Table 5. 5.2. Although these may yield some estimate of general portfolio risk.0 0.6 Source: Carty &Lieberman [96a] — Moody’s Investors Service The problem with trying to understand the volatility of individual exposures in this fashion is that it must be viewed within a portfolio.2.0 0. There are several hypotheses to explain why default rates would be volatile: • defaults are simply random events and the number of ﬁrms in the credit markets is not large enough to smooth random variation.4 4. These techniques would not allow the examination of marginal risk brought by adding some speciﬁc proposed transaction. which is taken from Carty & Lieberman [96a].Sec.9 0. These approaches can be grouped into two categories which we discuss in turn: • historical default volatility.8 5.1 0. it may also be useful to examine the volatilities of total holding period returns.7 0.2 Volatility of holding period returns The volatility of default events is only one component of credit risk. 5. All three hypotheses are likely to have some truth for the corporate credit markets.
CreditMetrics assumes that all migrations might have been realized and each is weighted by the likelihoods of migration which we argue is best estimated using long term data. but the idea is to track a benchmark corporate bond (or index) which has liquidity and observable pricing. Bonds within each credit rating category can be said also to have volatility of value due to daytoday credit spread ﬂuctuations. The resulting estimate of volatility of value is then used to proxy for the volatility of some exposure (or portfolio) under analysis. In this approach.62 Chapter 5. First. but will also sometimes realize a potentially large move due to a credit rating migration. This process of observing volatility should be unbiased over many trials. the last year. portfolio diversiﬁcation is addressed only to the extent that the portfolio under analysis is assuming to be analogous to the credit market universe. Second.2 below. • the data as it has been collected would require a standard deviation estimate over a sample size of less than 30 and so the standard error of the estimate is large. This general approach is sometimes termed the RAROC approach. maturity bucket. Overview of credit risk literature nett. CreditMetrics™—Technical Document . Our approach is probabilistic. However. the benchmark bond will have realized some credit quality migration and the resulting observed volatility will be (relatively) large. the benchmark bond will neither upgraded nor downgraded and the resulting observed volatility will be (relatively) small. Once the historical return volatility is estimated – perhaps grouped by credit rating. Implementations vary. there is the obvious problem of diversiﬁcation differences. Potential problems with this approach arise because of its relative inefﬁciency in estimating infrequent events such as up(down)grades and defaults. Observing some benchmark bond in this fashion over. will yield one of two qualitative results. 2 The credit quality migration and revaluation mechanism in CreditMetrics gives a weight to remote but possible credit quality migrations according to their longterm historical frequency without regard to how a shortterm (perhaps one year) sampling of bond prices would – or would not – have observed theses. there are studies by Asarnow [96] and Asarnow & Marker [95]. the estimation error is potentially high due to the infrequent but meaningful impact of credit quality migrations on value. say. Again. Consider Chart 5. Esser & Roth [93] and Wagner [96]. The RAROC approach seeks to measure these ﬂuctuations. But there are also three practical concerns with this approach: • historical returns are likely to poorly sample returns given credit quality migrations (including defaults) which are lowfrequency but important2. For commercial loans. and industry/sector – some practitioners have applied them to analogous exposures in the credit portfolio. and • the studies listed have commingled all sources of volatility – including interest rate ﬂuctuations – rather than just volatility in value due to credit quality changes. Our approach in CreditMetrics uses long term estimate of migration likelihood rather than observation within some recent sample period and so should avoid this problem.
To do this requires an understanding of credit quality correlations between obligors. bank lenders. For example.3 A portfolio view 63 Chart 5. Modern portfolio theory is commonly applied to market risk.5 95 97. typically set exposure limits against several types of portfolio concentrations. such as industrial sector. Lacking the guidance of a model.5 105 107. They constructed 33 industry indices of default experience as Part II: Model Parameters . primarily in application to equity portfolios. industrial sectors are generally deﬁned by aggregating fourdigit Standard Industrial Classiﬁcation (SIC) codes into 60 or fewer groupings. most institutions still rely on an intuitive assessment as to what level of over concentration to any one area may lead to problems. etc. Although this is likely true for commodity process industries like oil reﬁning and wood/paper manufacture. In contrast. geographical location. we might follow Stevenson & Fadil [95].2 Construction of volatility across credit quality categories Relative frequency All potential migrations A (no sample period migration) BBB AA B. we believe it would be less true for proprietary technology industries like pharmaceuticals and computer software. This implies that the banker is assuming that credit quality correlations are higher within an industry or sector and lower between industries or sectors. This type of analysis was originated by Harry Markowitz. A growing number of major institutions estimate the portfolio effects of credit risk in a Markowitztype framework. if we were interested in modeling the coincidence of just defaults. these groupings tend to be subjective rather than statistical.5 110 5. Thus. However. The volatilities and correlations necessary to calculate portfolio market volatility are generally readily measurable. It is not clear from the data that this is necessarily true. 5. product type. CCC Default BB AAA 90 92.Sec. there has been relatively little academic literature on the problem of measuring diversiﬁcation or overconcentration within a credit portfolio.3 A portfolio view Any analysis of a group of exposures could be called a portfolio analysis.5 100 Bond value 102. and has subsequently gone through considerable development. So. for instance. We use the term here to mean a Markowitztype analysis where the total risk of a portfolio is measured by explicit consideration of the relationships between individual risks and exposure amounts in a variancecovariance framework.
While this approach is ﬁne in concept. Neither of these studies has been realized in a practicable implementation.64 Chapter 5. not just defaulted ﬁrms. This is the model on which we have constructed the equitybased correlation estimation in Chapter 8. An assumption of bivariate normality between ﬁrms’ asset value moves then allows credit quality correlations to be estimated from equity prices series. another approach is to construct indices of. In contrast to these academic suggestions. as illustrated in Chart 3.P. Morgan has talked with KMV for at least four years on this approach to correlation and we are grateful for their input. CreditMetrics™—Technical Document . it suffers from the infrequency of defaults over which to correlate. they estimate the value of a ﬁrm’s debt within the option theoretic framework ﬁrst described in Merton [74]. J. there is a practicable framework which is a commercial offering by KMV Corporation. In brief. Overview of credit risk literature listed in Dun & Bradstreet’s Business Failure Record. Gollinger & Morgan [93] used time series of default likelihoods (ZetaScores™ published by Zeta Services) to estimate default correlations across 42 industry indices. The approach practiced by KMV is to look to equity price series as a starting point to understanding the volatility of a ﬁrm’s underlying (unlevered) asset value moves. Both expected default frequencies (EDFs) and correlations of default expectation are addressed within a consistent – and academically accepted – modeloftheﬁrm. To get around this problem. but default likelihoods of all ﬁrms. These asset values can in turn be mapped ordinally (onetoone) to credit quality measure. We know of two services which publish quantitatively estimated default likelihood statistics across thousands of ﬁrms: KMV Corporation and Zeta Services.3. Asset value moves can be taken to be approximately normally distributed. The correlation between these indices was their industry level estimate default correlation.
1. For instance.1 Default As discussed in the previous chapter. By itself.1 In the academic research. then the default rate will be much lower than if the population broadly includes all extant debt. although senior and subordinated debt to a ﬁrm will encounter what we term “credit distress” at the exact same time. If rates are tabulated for the ﬁrst few years of newly issued. however. this only gives an ordinal ranking of the default likelihoods across the categories. It is the senior rating that we look to as the most indicative of credit distress likelihood. CreditMetrics makes use of an extended deﬁnition that includes also the volatility of up(down)grades. 6. A quantitative framework. Mullins & Wolff [89]. Up to 1989.1 Deﬁning credit distress For our purposes in CreditMetrics. such as CreditMetrics. even the deﬁnition of the default event has evolved over time. credit rating systems typically assign an alphabetic or numeric label to rating categories.65 Chapter 6. Since then. researchers realized that distressed exchanges can play an important role in default statistics. In this chapter we do the following: • detail our modeloftheﬁrm which relates changes in underlying ﬁrm value to the event of credit distress. 6. and • detail an approach to estimate transition probabilities which is sensitive to both the historical tabulation and anticipated longterm behavior. . starting with Asquith. and 1 Rating agencies commonly also include a judgment for differing recovery rates in their subordinated and structured debt rating. we look to the following characteristics when we speak of the likelihood of credit distress: • default rates which have been tabulated weighted by obligors rather than weighted by number of issues or dollars of issuance. Also default rates can be materially different depending upon the population under study. Default and credit quality migration A fundamental source of risk is that the credit quality of an obligor may change over the risk horizon. • generalize this model to incorporate up(down)grades in credit quality. • discuss anticipated longterm behavior of transition matrices. must give meaning to each rating category by linking it with a default probability. As we show here. the anticipated recovery rate for subordinated is lower and thus it is given a lower rating. • discuss the historical tabulation of transition matrices by different providers. it was common to look for only missed interest or principal payments (see Altman [87]). • default rates which have been tabulated broadly upon all obligors rather than just those with recent debt issuance. “Credit quality” is commonly used to refer to only the relative chance of default.
Of course. 6.. There are two ways of using alternative credit rating systems depending upon what historical information is available.4. perhaps a onetoone correspondence could be made to established rating systems based on each credit category’s rating criteria.” CreditMetrics™—Technical Document . 6.2 then the transition matrix which “best replicates” this history can be estimated. Certainly it is the case that equity prices drop precipitously as a ﬁrm moves towards bankruptcy. If we take the default likelihood as given by the credit rating of the ﬁrm. The data which drive this model are the default likelihood and credit rating migration likelihoods for each credit rating. Table 6. which are separate from the likelihood of default). The rating agencies assign lower ratings to subordinated debt in recognition of differences in anticipated recovery rate in default. there are clearly differences in rating – to different debt of the same ﬁrm – between senior and subordinated classes. then tabulating a transition matrix would give ﬁrst direct estimate of the transition likelihoods including default. 2 Moody's terms these aggregated groupings “cohorts” and S&P terms them “static pools.g. We can compactly represent these rating migration probabilities using a transition matrix model (e. However. Likewise. This last point is worth elaborating. • If all that is available are cumulative default histories by rating category. then we can work backwards to the “threshold” in asset value that delimits default. These are included as part of the dataset for CreditMetrics. In the absence of historical information.2 Credit quality migration Credit rating migrations can be thought of as an extension of our model of ﬁrm defaults discussed in Section 3 and illustrated again in Chart 6.2 Fitting probabilities of default with a transition matrix Based on historical default studies from both Moody’s and S&P credit rating systems. • If individual rating histories are available. We say that a ﬁrm has some underlying value – the value of its assets – and changes in this value suggest changes in credit quality.2). Default and credit quality migration • default rates which are tabulated by senior rating categories (subordinated ratings include recovery rate differences. so also do the rating migration probabilities deﬁne thresholds beyond which the ﬁrm would be deemed to up(down)grade from its current credit rating. We utilize credit ratings as an indication of the chance of default and credit rating migration likelihood. just as our ﬁrm default model uses the default likelihood to place a threshold below which a ﬁrm is deemed to be in default.1. It is certainly true that senior debt obligations may be satisﬁed in full during bankruptcy procedures while subordinated debt is paid off only partially. Thus we take the senior credit rating as most indicative of the chance of a ﬁrm encountering credit distress. This is treated more formally in Section 8.66 Chapter 6. In this circumstance we would say that the ﬁrm – and so all its debts – encountered credit distress even though only the subordinated class realized a default.1. we have transition matrices which include historically estimated oneyear default rates. there are many rating agencies beyond S&P and Moody’s.
electric utility as it does for a French bank).g. We assume that all ﬁrms tagged with the “correct” rating label. calls. These probabilities give the likelihood of migrating to any possible rating category (or perhaps default) one period from now given the obligor’s credit rating today.e. Of course. two assumption that we make about transition matrices.3 Thus. there is no reason that transition matrices could not be tabulated more speciﬁcally to reﬂect potential differences in the historical migration likelihoods of industries or countries. One caveat to this reﬁnement might be the greater “noise” introduced by the smaller sample sizes. By this we mean that the rating agencies’ are diligent in consistently applying credit rating standards across industries and countries (i. 3 The securitised form of this structure is called a CreditSensitive Note (CSN) and is discussed in more detail in Das & Tufano [96].1 Model of ﬁrm value and migration BBB BB B CCC Default Firm remains BBB A AA AAA Lower Value of BBB firm at horizon date Higher Many practical events (e. Said another way. a transition matrix is nothing more than a square table of probabilities. we ﬁnd it very convenient to explicitly incorporate awareness of rating migrations into our risk models. 6. They are: 1.Sec.. enforced collateral provisions. however. a “Baa” means the same for a U. This is clearly an advantage since migration volatilities can vary widely between initial credit rating categories. Part II: Model Parameters . For instance. This technique is both powerful and limited. It is powerful in that we can freely model different volatilities of credit quality migration conditioned on the current credit standing. each row in the transition matrix describes a volatility of credit rating changes that is unique to that row’s initial credit rating. spread resets) can be triggered by a rating change..S. 6.3 Historical tabulation We can tabulate historical credit rating migration probabilities by looking at time series of credit ratings over many ﬁrms.3 Historical tabulation 67 In essence. Chart 6. a pricing grid – which predetermines a credit spread schedule given changes in credit rating – can reduce the volatility of value across up(down)grades. These actions can directly affect the realized value within each credit rating category. There are.
it makes sense to eliminate the N. The transition matrix is tabulated upon issuers conditioned on those issuers continuing to be rated at the end of the year.94 90. Here we list three of these sources: Moody’s. Moody’s Baa is “just like” S&P’s BBB. Thus there is no concern with having to adjust for a nolongerrated “rating.g.04 0 Aa 0.24 1.13 0 A 0 0. category and grossup the remaining percentages in some appropriate fashion.61 0. Since S&P describes that they track bankruptcies even after a rating is withdrawn. CreditMetrics™—Technical Document .91 69..91 84. These issuers are predominantly U. the default probabilities are already fully tabulated.63 88.26 4.29 6.” and so we pause to discuss this issue.01 0. but are not shown here.2 Standard & Poor’s transition matrix It happens that the transition matrix published by Standard & Poor’s includes a nolongerrated “rating.64 7. 6.1 Moody’s Investors Service transition matrix 2.40 1.20 Caa Default 0 0.07 0. Each is shown for the major credit rating categories – transition matrices which cover the minor (+/) credit rating are also available.02 0 0.71 5. We believe that there is no systematic reason correlated with credit rat4 KMV is not a credit rating service.48 5.55 2. category in application. etc. each speciﬁc to a particular credit rating service.05 Ba 93. The majority of these withdrawals of a rating occur when a ﬁrm’s only outstanding issue is paid off or its debt issuance program matures.based ﬁrms.02 0 0 Aaa 5.02 0.3.28 0. They quantitatively estimate Expected Default Frequencies (EDF) which are continuous values rather than categorical labels using an option theoretic approach.45 4.46 92.06 Source: Lea Carty of Moody’s Investors Service 6.15 1.4 We advocate maintaining this correspondence even though it is common for practitioners to use shorthand assumptions.12 0.22 4. So there should be no N. and KMV.01 0. Yet our assumption is that CreditMetrics will be applied to obligations with a known maturity. e. Moody’s utilizes a data set of 26 years’ worth of credit rating migrations over the issuers that they cover. Thus.08 0.05 0.26 0. We assume that all ﬁrms tagged with a given rating label will act alike.05 0.42 0.62 Baa 0.S.R. We do this as follows.54 0.3.R.81 24.35 2.76 86.” Table 6.44 5.09 0. S&P.1 Moody’s Investors Service: Oneyear transition matrix Initial Rating Aaa Aa A Baa Ba B Caa Rating at yearend (%) 0 0.08 B 0 0 0. but are including more and more international ﬁrms.16 0. Default and credit quality migration There are several sources of transition matrices.51 0.68 Chapter 6.91 6. By this we mean that the full spectrum of credit migration likelihoods – not just the default likelihood – is similar for each ﬁrm assigned to a particular credit rating.
There are an average of 4.33 90.93 7.84 83.00 4. the sample transition matrix shown in Table 6.03 0 0. Firms that disappeared from the sample were allocated into the rating categories proportionately to the population.81 0.24 B 0 0.17 8.30 80.67 0.12 1. The purpose of this sample is to show how an alternative approach such as EDFs can be utilized to generate a transition matrix.46 11.06 5. resulting in a total of 329.65 2.2 below: Table 6.68 7.18 1.52 86. the rating group based on the EDF of each company for that month was compared against the rating group it was in 12 months hence.3 was constructed from KMV EDFs (expected default frequency) for nonﬁnancial companies in the US using data from January 1990 through September 1995.22 AAA 8. 6.43 1. This gave a single migration. Rating group #8 signiﬁes default.53 6. Each month.14 0. 1996 Both of these tables are included in the CreditMetrics data set.06 0.14 0.48 2. based on its EDF at that date. Initial Rating AAA AA A BBB BB B CCC Rating at yearend (%) 0. EDFs are default probabilities measured on a continuous scale of 0.74 5.24 0.3 Historical tabulation 69 ing stating which would explain rating removals. 6.20 19.02% to 20.73 0.02 0.33 0.2 Standard & Poor’s oneyear transition matrix – adjusted for removal of N.22 A 0 0.79 Source: Standard & Poor’s CreditWeek April 15.95 0. which is treated as a a terminal event for the ﬁrm.64 5. We thus adjust all remaining migration probabilities on a pro rata basis as shown in Table 6. but grouped into discrete “rating” ranges for application in CreditMetrics.30 BBB 0.06 0.01 0.70 0.3.02 0.Sec.09 0.05 5.06 0.0%.12 0.780 companies in the sample each month.79 91.R.26 1.803 migration observations. Part II: Model Parameters . In contrast.86 CCC Default 0 0 0.11 0 AA 0.38 BB 90.27 0.3 KMV Corporation transition matrix Both of the above transition matrices were tabulated by credit rating agencies.07 64.
26 21.. we also have strong expectations about credit rating migrations.94 0.45 6. In addition to what we have historically observed.83 44. Although it would be ﬁne to have some issuers within a portfolio evaluated with one service (i.e.4 Longterm behavior In estimating transition matrices.28 1. the number of grades considered. For instance. The nature and extent of the problems encountered will be a function of the particular rating system. Only subscribers to KMV’s expected default frequency (EDF) data would be users of such a transition matrix and so KMV will be offering it as part of that subscription.63 3.52 44.34 2.58 69. However.22 43.19 22.09 3 (A) 0.24 0. and the amount of historical data available.76 0. Subscribers to KMV’s Expected Default Frequencies utilize a measure of default probability that is on a continues scale rather than discrete groupings offered by a credit rating agency.42 23.86 1. For this reason.41 20. Historical tabulation is worthwhile in its own right.01 7..02 0.00 17. but which does not always follow from straightforward compilation of the historical data.13 Source: KMV Corporation Table 6. it would be inappropriate to mix systems (i.54 22.00 2 (AA) 22. there are a number of desirable properties that one wants a transition to matrix to have.41 20.00 3.20 0.05 0.37 25.47 1.67 0.56 22. In general.e.30 0.26 0.93 3.3 is presented as an example and will not be included in the CreditMetrics data set.01 10. 6. say). it is good practice to impose at least some of the desirable properties on the historical data in the form of estimation constraints.95 24. Default and credit quality migration Table 6. the example KMV transition matrix shown here will not be part of the CreditMetrics data set.77 6 (B) 7 (CCC) 8 (Default) 0.97 6. corporates and industrials by Moody’s. Both KMV and we ourselves advocate that each credit rating (or expected default frequency) be addressed by a transition matrix tailored to that system.69 0. it represents a limited amount of observation with sampling error.08 0. ﬁnancials evaluated by IBCA) and other issuers evaluated by another service (i. By rank order.01 0. as with almost any type of sampling.04 20.66 2.79 1 (AAA) 66.53 53.94 42. over sufﬁcient time we expect that any inconsistencies in rank order across credit ratings will disappear.. S&P ratings applied to Moody’s transition matrix).26 5 (BB) 0.48 0.3 KMV oneyear transition matrices as tabulated from expected default frequencies (EDFs) Initial Rating 1 (AAA) 2 (AA) 3 (A) 4 (BBB) 5 (BB) 6 (B) 7 (CCC) Rating at Yearend (%) 4 (BBB) 2.68 1. we mean a consistent progression in one direction such as default likelihoods always increasing – never then CreditMetrics™—Technical Document .70 Chapter 6.e.39 0.71 2. The following discussion uses S&P ratings as the basis for explaining these issues and how they can be addressed.04 0.10 0.99 7.14 0.80 0.
48 11.29 2. 3. the AAA row in Table 6. 1. For instance.46 25.68 19..64 7 .. Table 6. Table 6... .. We list three potential shortterm sampling error concerns here: • Output cumulative default likelihoods should not violate proper rank order..00 ... Apr..97 4 0.88 40. .16 0. For instance..24 0. 15.. . .02 45.27 0.06% – or 1/1658 – are possible).06 5.14 42. 45.02 0.15 5 ...4 illustrates part of a cumulative default probability table published by Moody’s..48 .10 10 .4 below shows that AAAs have defaulted more often at the 10year horizon than have AAs. ..34 17. . 1996 6. If we ignore for the moment the issue of autocorrelation.65 .. only probabilities in increments of 0.658 ﬁrmyears worth of observation.70 . .91 . For instance.00 0.95 31... the restructuring of the highyield market in the 1980’s).e.... There are other potential problems with historical sampling such as the business cycle and regime shifts (e.97 21..44 1. Table 6..10 15 Source: S&P CreditWeek.01% chance of AAA default. 30. . ... we can always work backwards from a cumulative default table to an implied transition matrix...18 1.06 0.99 14.2 above is supported by 1... if there were truly a 0. 19.00 0. Part II: Model Parameters .12 0..g. 6.. • This lack of resolution may erroneously suggest that some probabilities are identically zero.. ..40 35.89 1.15 0.. 0.78 10.25 0.. .40 1.4 Average cumulative default rates (%) AAA AA A BBB BB B CCC Term 0.20 19.79 1 0..43 0. 1... This is enough to yield a “resolution” of 0.67 1..” Said another way.06% (i.. . But these will not be addressed here.Sec..12 15.4 Longterm behavior 71 decreasing – as we move from high quality ratings to lower quality ratings.1 Replicate historical cumulative default rates The major rating agencies have published tables of cumulative default likelihood over holding periods as long as 20 years – reported in annual increments. . .00 0.27 8.72 6.... then we would have to watch for another 80 years before there would be a 50% chance of tabulating a nonzero AAA default probability.4. 1..12 2.44 3.73 29. .66 0.92 2 0.07 0. then it is generally true that “there exists some annual transition matrix which best replicates (in a least squares sense) this default history.17 4..00 26..63 3 0.. • Limited historical observation yields “granularity” in estimates. 4.40 ...
6 This is an important result.29 0.28 A BBB 0.30 0.60 17.71 1. the most important point is that Table 6. At this point.2.05 0.28 20.5 Imputed transition matrix which best replicates default rates Initial Rating AAA AA A BBB BB B CCC Rating at year end (%) AAA 43.13%.84 3. For instance.56 5.) Table 6.5 Thus.40 41. it is apparently true that defaults over time are closely approximated by a transition matrix model.67 0. CreditMetrics uses a transition matrix to model credit rating migrations not only because it is intuitive but also because it is an extremely powerful statistical tool.17 55.77 8.07 55.30 9.14 0. A transition matrix model is an example of a Markov Process.97 12.12 0. thus the Markov process is a reasonable modeling tool.45 3.17 1.01 0.12 75.58 0. 6 CreditMetrics™—Technical Document .60 2. Again.29 3.60 0. Below we show a transition matrix that has been created using nothing but a least squares ﬁt to the cumulative default rates in Table 6.93 0. The median difference between them is 0.42 90. a transition matrix ﬁt is not as good for cumulative default rates of newly issued debt (as opposed to the total debt population) due to a “seasoning” effect where subinvestment grades have an unusually low default likelihood in the ﬁrst few years. A Markov Process is a statespace model which allows the next progression to be determined only by the current state and not information of previous states.11 AA 1.11 0.16% with a maximum error of 2.20 0.6 are quite close. 5 Empirically.65 6.08 4. This “seasoning” problem has not been apparent for bank facilities.46 Default 0.20 92.07 0.16 0. This behavior – of noncrossing default likelihoods – is a feature that we would expect given very long sampling histories.07 0. This “best ﬁt” Markov process has yielded the side beneﬁt of resolving nonintuitive rank order violations in its resulting cumulative default rates.12 BB 0.95 19.4.78 0.69 6. our problem of AAA’s having a 10 year default rate that was greater than AA’s is now gone.14 4.50 0.04 53.4 and Table 6. we show below in Table 6.4.97 1.04 0.22 2. It demonstrates that the statistical behavior of credit rating migrations can be captured through a transition matrix model. Default and credit quality migration Cumulative default rate tables like this can be ﬁt fairly closely by a single transition matrix.6 the cumulative default rates which result from this transition matrix.25 0.53 17.02 0.5 is a faithful replication of the historically tabulated Table 6.72 Chapter 6.05 32.53 B CCC 0.19 0. (We make no claim that Table 6. we are most interested in showing that: (i) such a matrix can be derived and (ii) that the process of defaults is closely replicated by a Markov process.01 0.15 For comparison to Table 6.
.41 3..95 19..18 0.17 1. 1.04 0.. if we managed a portfolio which was not allowed to invest in subinvestment grade bonds.5 are as shown in Table 6.7. 0. we can ask...69 3 3.6 matrix (%) Term AAA AA A BBB 0....85 .54 2..91 6. “what is the cumulative rate of crossing any given level of credit quality?” For instance.. The cumulative probabilities for crossing the “BB barrier” using the transition matrix in Table 6.39 30.40 10 .15 0..71 13. This table above shows that our imputed transition matrix violates this anticipated longterm behavior.49 18...25 0.. 42.01 0.30 0. .44 0. 14.78 6.25 15 Just as we would expect very longterm historical observation to resolve violations of nonintuitive cumulative default rank order. 28.. .....37 . ........98 39. 1.17 0. (In fact... ..09 6. .01 5 . We can now replay the least squares ﬁt we performed when we produced Table 6.4 Longterm behavior 73 Table 6..66 2 2.. Notice that monotinicity (rank order) is violated for singleAs. ...75 32.14 7 .. In general... Table 6. . 45.74 9.... ..00 11.7 “BB barrier” probabilities calculated from Table 6..27 0.69 30....29 36.09 0. then we might be interested in the likelihood of any credit quality migrations which were to or across the BB rating barrier.2 Monotonicity (noncrossing) barrier likelihoods Cumulative default rates are just a special case of what we term “barrier” likelihoods.25 8. .06 0.08 4.65 1...07 0.05 21..60 ..31 0. we should expect resolution of barrier rank ordering..60 5 ..63 . .6 Resulting cumulative default rates from imputed transition matrix (%) Term AAA AA A BBB BB B CCC 0..57 1 1.17 8. 1.12 3.49 .. 11.46 1.81 3.40 2..33 36..32 4 0.94 .. . . 4. 25..12 12. 2.80 5.. there are six nondefault “barriers” for seven rating categories and our ﬁtting algorithm addressed them all. ..51 5. .66 .17 4.57 25.20 15.43 2 0...46 1.35 47..85 3.53 . . 0. ...14 15.05 15 6..19 .13 19. .76 19..97 27. 2..4.01 0..15 1 0.. 10 .. ..63 3 0.8 shows these same BB barrier probabilities with our new ﬁt..46 . Table 6. 6.54 3.41 10.78 7 ..4 with the added constraint that all possible barrier probabilities must also be in rank order. ..83 20. 7. 1.82 21.93 4 5.71 .) Part II: Model Parameters .95 ...50 6.11 ..47 19. 17. ..79 11. 18.Sec.
26 3.01 0.19 2..91 0. .18 1.25 2.14 0.. The differences in predicted cumulative default rates averages only 0. 16.05 0..73 80.50 16.14 0. we are moving towards a better approximation of the historical transition matrix shown in Table 6. without directly trying.10 B 0.89 7 .03 7.52 8....12 5..36 14.01 21.88 1 1.03 0.63 1.37 12.9 Imputed transition matrix with default rate rank order constraint Rating at year end (%)Initial Rating AAA AA A BBB BB B CCC 58.18 4 4.14 Perhaps the difference between Table 6.02 0.06 4.33 BB 0.6 matrix (%) Term AAA AA A BBB 0.06 0.36 3.29 18.. .84 19.94 19.47 91.48 31.18 0.80 21.04 AAA 39.10) from Standard & Poor’s CreditWeek April 15....13 5. .14 0.5 and Table 6. CreditMetrics™—Technical Document . ...77 3 3..15 10. .5. 10.8 “BB barrier” probabilities calculated from Table 6..29 10..9 is that the weight of probabilities are generally moved towards the upperleft to lowerright diagonal.21 5.34 10 . .02%) between the two ﬁtted transition matrices.18 0.72 0.02 89.91 17.07 0.38 2.74 Chapter 6.82 14..18 54..77 BBB 0. 1996. . Default and credit quality migration Table 6.77 4.01 4.24 32.71 0.97 19.57 0.98 3.44 0.15 4. Also. To represent the rating proﬁle across the bond market..4. we show this new ﬁt of our imputed transition matrix..39 1..83 2.57 5.07 0. we have taken the following data (Table 6. . 6.87 41.73 38.53 55.15 0.06% (median is 0..42 7.82 6.65 CCC Default 0.39 3.45 3. 6. . For comparison with Table 6.09 2.88 25.57 0.62 0.11 5 .2.3 Steady state proﬁle matches debt market proﬁle Another desirable dimension of “ﬁt” for a transition matrix is for it to exhibit a longterm steady state that approximates the observed proﬁle of the overall credit markets.13 15 This reﬁnement was achieved with minimal change in the transition matrix’s ﬁt to the cumulative default rates.. By this we mean that – among those ﬁrms which do not default – there will be some distribution of their credit quality across the available credit rating categories. Table 6..14 AA 1. .07 1. . ..72 2 1..39 9.42 10..55 0....85 A 0.
our ﬁtting algorithm can achieve a closer approximation of the anticipated longterm steady state.4. and Part II: Model Parameters . eventually all ﬁrms will default.0% BBB 231 16.10 Estimate of debt market proﬁle across credit rating categories Count Proportion S&P 1996 AAA 85 6. First.8% BB 87 6.3% A 275 20.2. The transition matrix in Table 6.3% B CCC 13 0. 6. the proﬁle of nondefaulted ﬁrms will converge to some steady state regardless of the ﬁrm’s initial rating. we see that a more balanced proﬁle is achieved. since default is an absorbing state.9% Mathematically. our transition matrix Markov process will have two longterm properties (i.Sec. 2.2 Achieving a closer ﬁt to the longterm steady state proﬁle Frequency 60% Table 6.2% 200 14. Better ratings should never have a higher chance of default. our expectation is that there is a certain rank ordering the likelihood of migrations as follows: 1. we are moving towards a better approximation of the historical transition matrix shown in Table 6. more than 100 periods).4.2 result AAA AA A BBB Credit rating BB BB CCC This additional soft constraint was accomplished with a negligible effect on the matrix’s ability to replicate cumulative default likelihoods – and monotonicity in the barrier was still fully realized.10 50% 40% 30% 20% 10% 0% Section 6.3 result Section 6. Second. without directly trying.4 Monotonicity (smoothly changing) transition likelihoods Though it is certainly not a requirement of a transition matrix..9 shows too strong a tendency to migrate towards singleA.4 Longterm behavior 75 Table 6.5% AA 487 35. 6. Once we add an incentive to ﬁt the anticipated steady state.e. Chart 6. Also. The chance of migration should become less as the migration distance (in rating notches) becomes greater. As the chart below shows.4. since the initial state has geometrically less inﬂuence on future states.
11 1.12 1.25 AA 0. since the chance that a singleB would migrate to a singleA is greater than either a migration to BBB or BB.05 5. as we discuss next.10.27 1. Now we will bring all this together in Table 6. CreditMetrics™—Technical Document .84 0.05 0.34 BB 0. The chance of migrating to a given rating should be greater for more closely adjacent rating categories. 3.06 BBB 0.10 0.60 This transition matrix is meant to be close to the historically tabulated probabilities while being adjusted somewhat to better approximate the longterm behaviors we have discussed in this section.48 6.76 Chapter 6.58 18. 6.47 89.4. there is a “violation” of rule #3.59 1.02 0.29 2.45 7.13 0.36 0.23 1. From a risk estimation standpoint we see that there are now small but nonzero probabilities of default imputed for AAAs and AAs. Up to now we have discussed some of the characteristics of transition matrices and methods of addressing these.02 0.14 10. Also.91 0.25 1.16 7.93 88. we will refer to Table 6.08 0. Any ﬁtting algorithm that addresses smooth transition likelihoods would have to revisit these same probabilities when it includes knowledge of the historically tabulated transition matrix.85 A 0.5 Match historically tabulated transition matrix Standard & Poor’s historically tabulated transition matrix was shown above in Table 6.40 84.74 0.68 2.86 B 0. Default and credit quality migration As an example.05 63.21 5. Since the default likelihoods ascend smoothly there is no violation of #1. The reader can ﬁnd other probabilities in this table which are not monotonic in our deﬁnition.11 Achieving a closer ﬁt to the longterm steady state proﬁle Initial Rating AAA AA A BBB BB B CCC Rating at year end (%) 87.16 4.02 0.06 AAA 10. since singleB has a greater chance of migrating to singleA than does an initial BB or BBB.89 24.97 CCC Default 0.11 to give an estimate of a oneyear transition matrix which is rooted in the historical data and is also sensitive to our expectation of longterm behavior. So we address them both together below.53 8.84 0.29 9. we could add the soft constraint that our ﬁtting algorithm should endeavor to mitigate these nonrank orderings of probabilities as it seeks to replicate the cumulative default likelihoods.07 1.13 39.2.13 0. Table 6.89 2.32 8. As before.63 2. there is one last source of data that we should use in best estimating our transition matrix – an historically tabulated transition matrix.00 3.21 0.32 5.51 74. However. there is a “violation” of #2.03 0.31 12. However.06 0.
or will it be a regrettable but well behaved wrapping up which affects only shareholders but leaves debt holders whole? It is in the remote chance of an outright default that a credit instrument will realize its greatest potential loss. the general ﬁnding that recovery rates are highly uncertain with a distribution that can be modeled is applicable internationally. In this section. (ii) after some reasonable period for information to become available – perhaps a month. (ii) instrument type or use. by seniority level and exposure type. Will it be a catastrophe which leaves no value to recover. their mean and standard deviation. Across a typical portfolio. However. So it can be especially difﬁcult to imagine what the obligor’s position will be in the unlikely event of default. Any worthwhile credit risk model must be able to incorporate recovery rate uncertainty in order to fully capture the volatility of value attributable to credit. a typical portfolio will have a mixture of each. it is in the belief not that the obligor will go bankrupt but that the instrument will outperform. or (iii) after a full settlement has been reached – which can take years. but the important problem of addressing the wide uncertainty of recovery rate experience. Even if these issues are resolved there is the question of whether it is best to estimate values: (i) immediately upon announcement of default. Recovery rates Residual value estimation in the event of default is inherently difﬁcult. Investment grade credits will have relatively more of their volatility attributable to credit spread moves versus subinvestment grade credits. However. we estimate any potential correlation of recoveries across the book. Often there is no market from which to observe objective valuations. and • ﬁtting a full distribution to recovery rate statistics while preserving the required 0% to 100% bounds. But the most striking feature of any historical recovery data is its wide uncertainty. with most of the portfolio risk coming from the subinvestment grades. and if there are market prices available they will necessarily be within a highly illiquid market. However. In this chapter. which will be primarily driven by potential default events.1 Estimating recovery rates There are many practical problems in estimating recovery rates of debt in the event of default. once we contemplate volatility of recoveries. . (iii) credit rating Xyears before default. and (iv) size and/or industry of the obligor. we will discuss not only the estimation of mean expected recovery rate in default. The academic literature in our bibliography focuses primarily on U. most of the credit risk will be attributable to default events.77 Chapter 7. So the magnitude of any recovery rate in default is important to model diligently. 7. This chapter is organized as follows: • estimating recovery rate distributions. defaults post October 1. At the time when a banker makes a loan or an investor buys a bond. We have seen much effort devoted to estimating recovery rates based on: (i) seniority ranking of debt. 1979 – the effective date of the 1978 Bankruptcy Reform Act. we must also address any potential correlation of recoveries across a portfolio.S.
74 $20. Seniority Class Senior Secured Senior Unsecured Senior Subordinated Subordinated Junior Subordinated Carty & Lieberman [96a] Number Average Std. Table 7. Among bank facilities (e. Recovery rates Since there have been academic studies. 7. Table 7.1 Recovery rates of bonds For corporate bonds.g.1 is a partial representation of Table 5 from Moody’s Investors Service. which conclude that the bond market efﬁciently prices future realized liquidation values.89 $22. These studies also note the high average uncertainty of returns and thus the market’s risk pricing efﬁciency.71 177 $34.86 278 $51. This certainly is the value which an active investor would face whether or not he chose to hold his position after the default event.81 226 $32. that report high average holding period returns for debt held between the default announcement and the ultimate bankruptcy resolution.65 $26.08 214 $31. It is likely that there is a selfselection effect here. 85 $57. Dev. 115 $53..38 $25. use of borrowing or security. It is clear from the data that the historical loan recovery rates have been higher than recovery rates for senior bonds. There is a greater chance for security to be requested in the cases where an underlying ﬁrm has questionable hard assets from which to salvage value in the event of default. These studies reﬁne their estimates of recovery rate according to seniority type among bonds. see Swank & Root [95] and Ward & Griepentrog [93]. It is not clear whether this is attributable to differences in relationship.09 $10. Dev.1 Recovery statistics by seniority class Par (face value) is $100. loans. In contrast. commitments. we take comfort in those studies which poll/estimate market valuations about one month after the announcement of default.00.18 9 $17. There is no public study we are aware of that seeks to isolate the effects of different levels of security controlling for the asset quality of the obligor ﬁrm. CreditMetrics™—Technical Document . For bond recoveries we can look primarily to Moody’s 1996 study of recovery rates by seniority class.52 $23. It becomes then a 1 There are two studies. see Eberhart & Sweeney [92]. letter of credit) the studies have viewed these as a separate “seniority” class. the difference between secured versus unsecured senior debt is not statistically signiﬁcant.80 $26.13 $25.34 $22. This study has the largest sample of defaulted bond that we know of.45 196 $38. we have two primary studies of recovery rate which arrive at similar estimates (see Carty & Lieberman [96a] and Altman & Kishore [96]).78 Chapter 7.42 — — — As this table shows.90 Altman & Kishore [96] Number Average Std.1. which shows statistics for defaulted bond prices (1/1/70 through 12/31/1995).1 We look to the following independent studies for use in CreditMetrics.99 221 $47. the subordinated classes are appreciably different from one another in their recovery realizations.
1 Estimating recovery rates 79 practical problem for the risk manager to judge on a bondbybond basis what adjustment is best made to recovery rate estimates for different levels of security. U. 7. 5 77 2. bankruptcy experience.5 1 00 Recovery rate % Source: Asarnow & Edwards [95] A legitimate concern is that all of the studies referenced here are either exclusively based on.1. CreditManager. of 32. 5 37 2. 17 .5 27 2. 5 47 2.79% StDev: 32.Sec.21% mean and 78. Since bankruptcy law and practice differs from jurisdiction to jurisdiction (and even across time within a jurisdiction). or primarily driven by.1 below is reproduced from A&E. it is not clear that these historical estimates of recovery rate will be directly applicable internationally. Chart 7.2 Recovery rate of bank facilities For bank facilities. A&E track 831 commercial and industrial loan defaults plus 89 structured loans while C&L track 58 defaults of loans with Moody’s credit ratings. 2 For this reason.1 Distribution of bank facility recoveries Frequency 120 100 80 60 40 20 0 0 7.2 7. 5 Mean: 65. which is beyond the information reported by A&E.5 2 .7%.70% 2. our software implementation of CreditMetrics. Both studies treat bank facilities as essentially a seniority class of their own – with this being senior to all public bond seniority classes.5 9 . 5 57 2. 5 97 2.S.5 5 . So these two studies are different by no more than 5%.5 4 . Chart 7. will allow recovery rate estimates to be overwritten on the individual exposure level Part II: Model Parameters .5 6 . Moody’s reports a 71% mean and 77% median recovery rate which is within sampling error of Asarnow & Edwards 65.5 3 .5 7 .5 8 . and we have used it to estimate the standard deviation of recovery rates. 5 67 2.21% Median: 78. 5 87 2.79% median recovery rates. we again have two primary studies of recovery rate which arrive at similar estimates see (Asarnow & Edwards [95] and Carty & Lieberman [96b]). 5 5 12 .
2 Distribution of recovery rate Recovery rates are best characterized. We can capture this wide uncertainty and the general shape of the recovery rate distribution – while staying within the bounds of 0% to 100% – by utilizing a beta distribution.5 4.0 0.2 illustrates beta distributions for different seniority classes using some of statistics reported in Table 7. Recovery rates 7. Uniform distributions have a mean of 0. 4.. Loss rates are bounded between 0% and 100% of the amount exposed.0 3. The standard deviations of 25.80 Chapter 7. If we did not know anything about recovery rate.0 1. Beta distributions are ﬂexible as to their shape and can be fully speciﬁed by stating the desired mean and standard deviation. then we would model them as a ﬂat (i. where the shape of the full distribution is required.0 0% 25% 50% Residual value 75% 100% Subordinated Senior subordinated Senior unsecured Junior subordinated This full representation of the distribution is unnecessary for the analytic engine of CreditMetrics. not by the predictability of their mean. It is used later in our simulation framework. that is. CreditMetrics™—Technical Document .29 ( σ = 1 ⁄ 12 ).0 2.1. uniform) distribution between the interval 0 to 1. if we thought that all possible recovery rates were equally likely.45% for senior unsecured bonds and 32.7% for bank facilities are on either side of this and so represent relatively high uncertainties.5 1.5 3. but by their consistently wide uncertainty.5 and a standard deviation of 0. Chart 7.5 0.5 2.e.2 Example beta distributions for seniority classes . Chart 7.
The structure of this chapter is as follows: • First. this would argue for little correlation between different ﬁrms’ rating changes and defaults. we discuss methods to estimate the parameters of the asset value model. we investigate the possibility of modeling joint rating changes directly using historical rating change data. Thus.. One might claim that each ﬁrm is in many ways unique and its changes in credit quality often are driven by events and circumstances speciﬁc to that ﬁrm. which has studied models of credit correlations for a number of years.81 Chapter 8. we discuss evidence from default histories which supports our assertion that credit correlations actually exist. Both of these studies note the practical difﬁculties of estimating default correlations.) to correlate across 42 constructed indices of industry default likelihoods. and present a dataset for this purpose. Inc. it typically achieved by much larger numbers of exposures than for market portfolios. Whereas market risks can be diversiﬁed with a relatively small portfolio or hedged using liquid instruments.e. • Second. we examine several histories of rating changes and defaults in order to establish that such correlations in fact exist. across 33 industry groups.1 Finding evidence of default correlation In this section. simply having many obligors’ names represented within a portfolio does not assure good diversiﬁcation (i. We know of two academic papers which address the problems of estimating correlations within a credit portfolio: Gollinger & Morgan [93] use time series of default likelihoods (ZetaScores™ published by Zeta Services. credit risks are more problematic. as listed in Dun & Bradstreet’s Business Failure Record. . When diversiﬁcation is possible. • Last. Credit quality correlations Central to our view of credit risk estimation is a diligent treatment of the portfolio effect of credit. • Fourth. before moving on to modeling correlations in credit rating changes. we present a model which connects rating changes and defaults to movements in an obligor’s asset value. and Stevenson & Fadil [95] correlate the default experience. The problem of constructing a Markowitztype portfolio aggregation of credit risk has only recently been widely examined. Our portfolio treatment of credit risk was greatly inﬂuenced by various engagements with KMV. • Third. they may all be large banks within one country). For credit portfolios. we discuss the estimation of credit correlations through the observation of bond spread histories. This allows us to model joint rating changes across multiple obligors without relying on historical rating change or bond spread data. it would be desirable for us to ﬁrst ﬁnd evidence of defaults across a large body of companies. 8.
13% 1.33% 1. there are only 25 yearly observations supporting the calculation of σ (and it is reported with signiﬁcant rounding). µ .95% Pr { ρ > X } = 2. we observe default rates which ﬂuctuate. Thus.5% 0. if defaults were uncorrelated. We make this observation more precise below. σ2 N .1% 0.42% 7.1 Inferred default correlations with conﬁdence levels Credit rating category Aa A baa ba B Default rate 0. Credit quality correlations We can do this by examining the default statistics reported by the major rating agencies over many years.4% 4.27% ρ Pr { ρ < X } = 2.05% 0.03% 0. We will use the formula below to compute average default correlation ρ from the data.5% 1.91% 6. On the other hand. the number of names covered by the data. if defaults were perfectly correlated. so the conﬁdence levels around the resulting inferred correlation will be high. We can infer from these ﬁgures that the number of ﬁrmyears supporting the default rate. is in the thousands for all credit rating categories.62% Standard Implied deviation default defaults correlation 0. Since the studies we consider are based on a very large number of observations. That our observations lie somewhere between these two extremes (that is.15% 0. – 1 2 2 σ µ–µ ρ = .83% 2. our approximation formula for ρ is appropriate.3% 1. but not to the extent that they would under perfect correlation) is evidence that some correlation exists. Both Moody’s and S&P publish default rate statistics which could be used to make this type of statistical inference of average default correlations.35% 1. for a full derivation. However.1] where the approximation is for large values of N .45% 4.00% 0.29% 0. In Table 8.8% Lower conﬁdence Upper conﬁdence µ σ 0.69% 1. In the formula.01% 0.47% Source: Moody’s 19701995 1year default rates and volatilities (Carty & Lieberman [96a]) CreditMetrics™—Technical Document .≈ .79% 1. we use data from Tables 3 and 6 from Moody’s most recent default study (see Carty & Lieberman [96a]).2 µ–µ N–1 [8. then we would observe some years where every ﬁrm in the study defaults and others where no ﬁrms default.1% 0.40% 3. see Appendix F.1. µ denotes the average default rate over the years in the study and σ is the standard deviation of the default rates observed from year to year.82 Chapter 8. then we would expect to observe default rates which are very stable from year to year. Table 8.
then between two obligors there are 8•8 or 64 possible joint states whose likelihoods must to be estimated. CCC plus Default). thus the conﬁdence levels for the investment grades will be wider than those calculated. We have done this with a sample of 1. …. For a rating system with seven nondefault categories.234 ﬁrms who have senior unsecured S&P credit ratings reported quarterly for as much as the last 40 quarters. it is possible to avoid having to specify a correlation estimate and an accompanying descriptive model.5% level. In Table 8. our needs go beyond estimations of default correlations.13 million pairwise combinations within our particular sample. 8. is assumed to be constant across all ﬁrms within the credit rating category and constant across time. if the credit rating system recognizes eight states (i. Thus. Since we are interested in tabulating all possible pairwise combinations between ﬁrms. and • the approach is sensitive to the proportion of recession versus growth years which – in the 25year sample – may not be representative of the future. This is a fairly objective indication that default events have statistically signiﬁcant correlations which cannot be ignored in a risk assessment model such as CreditMetrics. 8. The inferred default correlations are all positive and – using the conﬁdence interval technique discussed above – are all statistically greater than zero to at least the 97. In general. In fact.. Part II: Model Parameters . are calculated over a very limited number of observations which lead to wide conﬁdence levels.2 Direct estimation of joint credit moves 83 There are at least four caveats to this approach: • the standard deviations of default rates.e. if a rating series data set offers N observations in a tabulated transition matrix then it will offer on the order of N2 observations of joint migration. AA. we must estimate the joint likelihood of any possible combination of credit quality outcomes.2 we show one of these 28 tabulated results for the case where one ﬁrm starts the period as a BBB and another ﬁrm starts the period with a singleA rating. µ . there will be 28 unique joint likelihood tales. there are over 1. AAA. • the average default rate.2 Direct estimation of joint credit moves Perhaps the most direct way to estimate joint rating change likelihoods is to examine credit ratings time series across many ﬁrms which are synchronized in time with each other. σ .Sec. • the underlying periodic default rates for investment grade categories are not normally distributed. We note that this data set does not include much of the default experience that S&P reports in their more comprehensive studies and stress that we have assembled this data set only to illustrate the principle that joint credit quality migration likelihoods can be estimated directly. With this method.
01 BB 0.584 2.64 0.760 7.105 44.37 0.36 5.00 CCC Default  We emphasize again that this illustration is only to demonstrate a technique for estimating joint credit quality migration likelihoods directly.00 0. However.01 0.507 193 48 15 0 CCC Default 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 This yields our nonparametric estimate of joint credit quality probabilities to be as in Table 8.477 41.02 0.90 0. 8. or rely on distilling the data down to a single parameter – the correlation.14 5.00 B 0.64 78.01 0. This method of estimation has the advantage that it does not make assumptions as to the underlying process.70 5.00 A 0.11 0.436 839 175 55 29 A 0 1.081 2. our own access to the rating agency’s data sets is inadequate to fully estimate a production quality study.03 0. Unfortunately.812 40. it carries the limitation of treating all ﬁrms with a given credit rating as identical.12 BBB 0.01 0.00 0.230 981 BBB 0 54 2.07 0.29 0.523 621. In the following sections. we discuss a method of estimating credit quality correlations which are sensitive to the characteristics of individual ﬁrms.57 0.32 0. it is reasonable to believe that correlations CreditMetrics™—Technical Document .2 Historically tabulated joint credit quality comovements Firm starting in BBB AAA AA A BBB BB B CCC Default Firm starting in A AAA 0 0 0 0 0 0 0 0 AA 0 15 978 12. So two banks would be deemed to have the same relationship as a bank and an oil reﬁner.84 Chapter 8.3 Historically tabulated joint credit quality comovement (%) Firm starting in BBB AAA AA A BBB BB B CCC Default Firm starting in A AAA AA 0.28 0.02 0.02 0. Credit quality correlations Table 8. Because it is intuitive that movements in bond prices reﬂect changes in credit quality.3: Table 8.12 1.3 Estimating credit quality correlations through bond spreads A second way to estimate credit quality correlations using historical data would be to examine price histories of corporate bonds.14 0.04 0.921 532 127 67 BB 0 4 414 5.075 321 76 18 7 B 0 0 224 2.05 0. the joint distribution shape.01 0.00 0.
We are motivated to pursue such an approach by the fact that practical matters (such as the lack of data on joint defaults) make it difﬁcult to estimate such probabilities directly. making the estimation of bond spread correlations impossible in practice. Where bond price histories are available. An extension of this type of model to two or more bonds would allow for the inference of default correlations from the correlation in bond spread moves. Estimate the parameters for the process above. this should be easier than estimating the joint rating change probabilities directly. we present the approach which we introduced in Chapter 3 and which we will use in practice to model joint probabilities of upgrades. While an approach of this type is attractive because it is elegant and consistent with other models of risky assets. and a process which we understand and can observe. This will establish a connection between the events which we ultimately want to describe (rating changes). we propose that a ﬁrm’s asset value be the process which drives its credit rating changes and defaults. This model is essentially the option theoretic model of Merton [74]. and then estimating the correlation in the movements of these spreads. which is discussed further in Kealhofer [95].Sec. it is possible to estimate some type of credit correlation by ﬁrst extracting credit spreads from the bond prices. likelihoods of joint credit quality movements). particularly for low credit quality issues. In this section. it is necessary to adopt a model which links spread movements to credit events. 2. 8.4 Asset value model In this section. and the third indicates whether the bond has defaulted. To arrive at the parameters we require for CreditMetrics (that is. Our approach here then will be indirect. Such an approach has two requirements: adequate data on bond price histories and a model relating bond prices to credit events. but which are not readily observable. downgrades. 8. Part II: Model Parameters . the second is the credit spread. Bond spread data is notoriously scarce. We describe the model which links changes in asset values to credit rating changes and explain how we parameterize 1 2 The evolution of these quantities is generally modeled by diffusion processes with some drift and volatility. A typical approach (see for example Duffee [95] or Nielsen and Ronn [94]) is to assume that the risk free rate and credit spread evolve independently1 and that defaults are linked to the credit spread through some pricing model. its biggest drawback is practical. This is similar to the inference of implied volatilities from observed option premiums. and defaults (all of which will be referred to generically as credit rating changes). If we have been successful in the ﬁrst part. Models of risky bonds typically have three state variables: the ﬁrst is the risk free interest rate. Propose an underlying process which drives credit rating changes.4 Asset value model 85 of bond price moves might allow for estimations of correlations of credit quality moves. It is important to note that such a correlation only describes how spreads tend to move together. It involves two steps: 1. This pricing model allows us to infer the probability of the issuer defaulting from the observed bond spread2.
We may suppose then that there is a speciﬁc level such that if the company’s assets fall below this level in the next year. asset “returns. consider a hypothetical company that is BB rated and whose assets are currently worth $100 million. the assumption is that the asset value in one year determines the credit rating (or default) of the company at that time.5. as in Chart 8. Extending the intuition above. this would be a sufﬁcient model. and parameterized by a mean µ and standard deviation (or volatility) σ. However. We postpone the discussion of parameter estimation to Section 8. we need a slightly more complex framework. We reiterate that we are not yet claiming to know what these thresholds are. Were we only treating value changes due to default. it will be unable to meet its payment obligations and will default. The asset values in the chart which correspond to changes in rating will be referred to as asset value thresholds. since we wish to treat portfolio value changes resulting from changes in credit rating as well. Then the assumption is that there are asset levels such that we can construct a mapping from asset value in one year’s time to rating in one year’s time. Credit quality correlations the asset value model.1. we assert that the percent changes in asset value (that is. Essentially. It is evident that the value of a company’s assets determines its ability to pay its debt holders. we assume there is a series of levels for asset value that will determine a company’s credit rating at the end of the period in question. only that this relationship exists. we only need to model the company’s change in asset value in order to describe its credit rating evolution. Chart 8. Note that this volatility is not the volatility of value of a credit CreditMetrics™—Technical Document . To do this.86 Chapter 8. For example.” which we will denote by R) are normally distributed.1 Credit rating migration driven by underlying BB ﬁrm asset value Rating AAA • • • • • • • • • • • • • • • AA A BBB BB B CCC Default • 0 20 40 60 80 100 120 140 160 180 200 Asset value in one year Assuming we know the asset thresholds for a company.
84 1.06 The connection between asset returns and credit rating may be represented schematically as in Chart 8. from the discussion of asset thresholds above. Continuing with our example of the BB rated obligor.03 0. and so on. we may now establish a connection between the asset thresholds in the chart above and the transition probabilities for our company.67 Probability according to the asset value model Φ(ZAA/σ)−Φ(ZA/σ) 1−Φ(ZAA/σ) BBB BB B 80. and so on. 8.) These probabilities are listed in the third column of Table 8. The integral between adjacent thresholds corresponds to the probability that the obligor assumes the credit rating corresponding to this region. then the obligor goes into default. ZCCC.14 0. if ZDef<R<ZCCC. ZBBB.73 Φ(ZBBB/σ)−Φ(ZBB/σ) Φ(ZB/σ)−Φ(ZXXX/σ) Φ(ZDef/σ) Φ(ZBB/σ)−Φ(ZB/σ) Φ(ZA/σ)−Φ(ZBBB/σ) CCC Φ(ZXXX/σ)−Φ(ZDef/σ) Default 1. 3 This likely will not be the case in practice. we will assume µ=03. It is in fact true that σ does not inﬂuence the ﬁnal result either – and the reader may choose to ignore σ in the expressions to follow – but we retain it for illustrative purposes. but for our purposes here. Since we have assumed that R is normally distributed.4. Given this parameterization of the asset value process. One year transition probabilities for a BB rated obligor Rating AAA AA A Probability from the transition matrix (%)( 0.00 7. we can compute the probability that each of these events occur: [8. then the obligor is downgraded to CCC. Part II: Model Parameters . For ease of exposition.. Pr { CCC } = Pr { Z De ( f < )R < Z CCC } = Φ ( Z CCC ⁄ σ ) – Φ ( Z Def ⁄ σ ) . So for example.4 Asset value model 87 instrument (which is an output of CreditMetrics) but simply the volatility of asset returns for a given name. (Φ denotes the cumulative distribution for the standard normal distribution. where we present the return thresholds superimposed on the distribution of asset returns.4. we read from the transition matrix that the obligor’s oneyear transition probabilities are as in the second column of Table 8. the value of µ will not inﬂuence the result. we know that there exist asset return thresholds ZDef. if ZDef were equal to 70%.2. such that if R< ZDef. On the other hand. this would mean that a 70% (or greater) decrease in the asset value of the obligor would lead to the obligor’s default.2] Pr { Default } = Pr { R < Z Def } = Φ ( Z Def ⁄ σ ) . Table 8. etc.Sec.4.53 8.
04σ −2. we may consider the CCC probability to solve for ZCCC. Note there is no threshold ZAAA.4 Table 8.39σ 1.5. The beneﬁt of the asset value process is only in the consideration of multiple obligors. which lets us solve for ZDef: [8.93σ 2. we only need the transition probabilities to describe the evolution of credit rating changes.06% ) ⋅ σ = – 2.5 Threshold values for asset return for a BBB rated obligor Threshold ZAA ZA Value 3. since any return over 3.30σ .06%. we see that Φ ( Z Def ⁄ σ ) must equal 1.43σ implies an upgrade to AAA. obtaining the values in Table 8.37σ −1.88 Chapter 8. So considering the default probability. CreditMetrics™—Technical Document .43σ 2. and the asset value process is not necessary. the standard deviation of asset returns for this obligor by σ′. Denote this obligor’s asset return by R′. and so on.2 Distribution of asset returns with rating change thresholds Downgrade to B Firm remains BB rated Upgrade to BBB Firm defaults Z CCC Z B Z BBB Z A Z AA Z AAA Asset return over one year Now in order to complete the connection. then the B probability to solve for ZB.3] Z Def = Φ ( 1.30σ ZBBB ZBB ZB ZCCC ZDef Now consider a second obligor. For one obligor. Credit quality correlations Chart 8. –1 where Φ –1 ( p ) gives the level below which a standard normal distributed random variable falls with probability p . we have not added anything to our model. Using this value.23σ −2. A rated. and its asset return thresholds 4 We comment that to this point. we simply observe that the probabilities in the two columns of the Table 1 must be equal.
Any multivariate distribution (including those incorporating fat tails or skewness effects) where the joint movements of asset values can be characterized fully by one correlation parameter would be applicable. We may use the same procedure to calculate the probabilities of each of 5 Technically.Sec.12σ′ 1.72σ′ −3.5 and it only remains to specify the correlation ρ between the two asset returns.74% 0.Σ) is the density function for the bivariate normal distribution with covariance matrix Σ6. we have described the motion of each obligor individually according to its asset value processes.Σ ) ( dr′ ) dr where f(r. To be speciﬁc. Table 8. r′ .05% (the probability that the A rated obligor remains A rated). The transition probabilities and asset return thresholds are listed in Table 8. we assume that the two asset returns are correlated and normally distributed.09% 2.e. Z′CCC. that it is not necessary to use the normal distribution. then we compute: [8.27% 91. We remark. To describe the evolution of the two credit ratings jointly.30σ′ −2.6.51σ′ −2.5] Pr { Z B < R < Z BB .52% 0.05% 5. We then have the covariance matrix for the bivariate normal distribution: [8.r′. ρ=0). 8.. If ρ is not zero.01% 0.6 Transition probabilities and asset return thresholds for A rating Rating AAA AA A BBB BB B CCC Default Probability 0.4] σ ρσσ' Σ = ρσσ'σ' 2 2 This done. If the two asset returns are independent (i. we know how the asset values of the two obligors move together. we assume that the two asset returns are bivariate normally distributed. This is the probability that the asset return for the BB rated obligor falls between ZB and ZBB while at the same time the asset return for the A rated obligor falls between Z′BBB and Z′A. The variables r and r′ in Eq.24σ′ At this point. [8.4 Asset value model 89 by Z′Def. and so on.06% Threshold Z’AA Z’A Z’BBB Z’BB Z’B Z’CCC Z’Def Value 3.26% 0. and can then use the thresholds to see how the two credit ratings move together.98σ′ −1.19σ′ −3.5] represent the values that the two asset returns may take on within the speciﬁed intervals. say we wish to compute the probability that both obligors remain in their current credit rating. 6 Part II: Model Parameters .Z' BBB < R' < Z' A } = ∫Z ∫Z′ B Z BB Z′ A BBB f ( r. then this joint probability is just the product of 80. however.53% (the probability that the BB rated obligor remains BB rated) and 91.
05 0.07 100 Total This table is sufﬁcient to compute the standard deviation of value change for a portfolio containing only issues of these two obligors.02 0. 7 Note that if all pairs of obligors have the same correlation.85 0.61 7. and 28 possible matrices.03 0.01 0.00 0.06%.14 0.75 BB 0.00 0.08 0.00 0. Thus.69 80.11 0. then the joint default probability is the product of the individual default probabilities. it is only necessary to repeat this analysis for each pair of obligors in the portfolio.00 0.00 0.03 0.07 0.00 0.00 1. if the asset returns are perfectly correlated (ρ=1). If the asset returns are independent. In Chart 8. so too does the BB rated obligor. Consider the worst case event for a portfolio containing these two obligors – that both obligors default.00 0.35 1. Credit quality correlations the 64 possible joint rating moves for the two obligors.01 0.02 0. then any time the A rated obligor defaults. On the other hand. Notice that Table 8.24 0.00 0.48 0. Thus.00 0.01 0.00 0.80 0.02 0. there are only 28 possibilities for the ratings of each pair of obligors.04 0. regardless of the size of the portfolio.00 0.00 0.00 0.01 0.02 0.00 0.7 depends only on the ratings of the two obligors and on the correlation between them. and not on the particular obligors themselves. CreditMetrics™—Technical Document .00 0.0006%. To compute the standard deviation for a larger portfolio.7 The effect of the correlation merits further comment.53 8.7 which would be needed is 28.01 0.10 73.13 5. since there are only seven possible ratings for each obligor. the probability that they both default is just the probability that the A rated obligor defaults.20 4.00 0.00 0. Table 8.06 Def 0.00 0.00 0.90 Chapter 8. or 0.67 7. then the maximum number of matrices like Table 8.29 AA 0.01 0.87 1.3.06 A CCC 0.00 0.00 0.00 0.00 0.13 0. we illustrate the effect of asset return correlation on the joint default probability for our two obligors.00 0.01 2.7 Joint rating change probabilities for BB and A rated obligors (%) Rating of ﬁrst company AAA AA A BBB BB B CCC Def Total Rating of second company BBB 0.01 0.00 0.00 0. Note that the totals for each obligor are just that obligor’s transition probabilities.90 91.00 0. 100 times greater than in the uncorrelated case.79 0.04 0.65 7.7. We would then obtain the probabilities in Table 8.01 0.09 0.01 0.26 B AAA 0. or 0.00 0. suppose that ρ=20%.13 0.79 0.18 0.56 0. As an example.
The translation from asset to default correlation lowers the correlation signiﬁcantly.04% 0. and that actual default correlations are not used in any calculations. Chart 8.4 shows how the default correlation is a function of the two obligor's default probabilities.01% 0. we have shown that it is possible to compute p 12 . it is only necessary to specify joint probabilities of rating changes and defaults. However. the probability that obligors 1 and 2 both default.1 require that asset value moves exhibit relatively high correlations. The high “mound” towards the back indicates that junk bond defaults will be far more correlated with each other than will investment grade defaults.0 0.06% 0.0 Correlation We have pointed out before that for pairs of obligors. We see then that even the very small default correlation estimates in Section 8. For an asset correlation ρ A . respectively.8 0.02% 0.03% 0. p1 ( 1 – p1 ) p2 ( 1 – p2 ) where p 1 and p 2 are the probabilities that obligor 1 and obligor 2 default. Asset correlations in the range of 40% to 60% will typically translate into default correlations of 2% to 4%.1 0.4 0.6] p 12 – p 1 p 2 ρ D = .3 0..05% 0. Part II: Model Parameters . An asset correlation of 30% was assumed and default probabilities range from 1bp to above 10%.5 0.9 1.00% 0.7 0. many people are accustomed to thinking in terms of default correlations.6 0. 8. and so we touch brieﬂy on them here. The default correlation between these two obligors can then be written as [8.4 Asset value model 91 Chart 8.Sec.2 0.3 Probability of joint defaults as a function of asset return correlation Joint default probability 0.
but this just means that its asset return thresholds are greater than those of the other ﬁrm.00% Default correlation 0.00% Default likelihood Obligor #2 Before moving on to estimation of parameters. The simplest is just to use some ﬁxed value across all obligor pairs in the portfolio. the only change is that we use the sixmonth transition probabilities to calibrate the asset return thresholds.100 0. that is. that in a risk model we are ignoring asset volatility.175 0. for exam CreditMetrics™—Technical Document .00% 1.5 Estimating asset correlations The user can pursue different alternatives to estimate ﬁrm asset correlations. This precludes the user having to estimate a large number (4.000 10. This may seem counterintuitive.175 0. One last comment is that it is a simple matter to adjust for different time horizons. but where the asset volatility for one obligor is ten times greater than the other. to perform this analysis for a sixmonth time horizon.125 0.050 0.200 0. The consequence of this is that we may consider standardized asset returns. In the end.025 0.01% 0.5] above does not depend on either of the volatilities σ or σ'. which is the focus of the next section.4 Translation of equity correlation to default correlation Default correlation 0.075 0. The only parameter to estimate then is the correlation between asset returns.10% 1. As an example. asset returns adjusted to have mean zero and standard deviation one. For example. Credit quality correlations Chart 8.00% Default likelihood Obligor #1 0.075 0.025 0 10.050 0.10% 0.92 Chapter 8.100 0. the only parameters which affect the risk of the portfolio are the transition probabilities for each obligor and the correlations between asset returns. we make one important observation: Equation [8.01%0.950 for a 100obligor portfolio) of individual correlations. However.125 0. One obligor does have a more volatile asset process.200 0.150 0. the ability to detail risk due to overconcentration in a particular industry. We know that the credit risk is the same to either obligor. 8. consider two obligors which have the same rating (and therefore the same transition probabilities).150 0. but essentially all of the volatility we need to model is captured by the transition probabilities for each obligor. while still providing reasonable portfolio risk measures.
8 For more speciﬁc correlations. it is more accurate than using a ﬁxed correlation. there are two steps: • First. we present an example to describe the methods by which the user speciﬁes the weightings for individual obligors and arrives at individual obligor correlations. a company might be mapped as 80% Germany and 20% United States. make this approach untenable. the scarcity of data for many obligors. hence a 20%to35% range. we could produce correlations for any pair of obligors which a user might request. and is based on much more readily available data than credit spreads or actual joint rating changes. 8. 24% in German ﬁnance. 9 Part II: Model Parameters .1. One fundamental – and typically very observable – source of ﬁrmspeciﬁc correlation information are equity returns. we obtain the correlations between obligors. we also report the volatility for each of these indices. Below. Here. for mapping individual obligors to the indices. Using these weights and the countryindustry correlations from above. to produce individual obligor correlations.5 Estimating asset correlations 93 ple. We will see that the volatilities of the indices are necessary. The result is that we obtain the correlation. A typical average asset correlation across a portfolio may be in the range of 20% to 35%. resulting in 56% participation in the German chemicals industry. and 70% chemicals and 30% ﬁnance. we discuss the data we provide and the methodology which goes into its construction. we utilize industry indices in particular countries to construct a matrix of correlations between these industries. we use the correlation between equity returns as a proxy for the correlation of asset returns. Thus. However. for example. is lost. For example. we map individual obligors by industry participation. The last subsection is a generalization of this example. For reasons which will become clear below. Our own research suggests that it is easier to construct higher correlation portfolios versus lower correlation portfolios. In the best of all possible worlds.5. of the German chemical industry with the United States insurance industry.Sec. Recall from Section 8.4 that volatilities do not ﬁgure into the model for joint rating changes. In the following subsection. 14% in American chemicals. 8 Based on conversations with Patrick H. Therefore. In Section 8. and 6% in American ﬁnance. however. we present our own interpretation of this type of underlying ﬁrm asset correlation estimation. Part of his research inferred average asset correlations of corporate & industrial loan portfolios within midsized US banks to be in the range 20%to25%. there are independent data providers that can provide models which are independent of – but can be consistently used in – CreditMetrics. While this method has the drawback of overlooking the differences between equity and asset correlations. we resort to a methodology which relies on correlations within a set of indices and a mapping scheme to build the obligorbyobligor correlations from the index correlations. as well as the impossibility of storing a correlation matrix of the size that would be necessary. McAllister in 1994 when he was an Economist at the Board of Governors of the Federal Reserve System.9 • Next.
and then eliminating groups which appear redundant.8.5. reasoning that the two indices essentially explain the same movements in the market. Credit quality correlations 8.94 Chapter 8. In Table 8.9. and so consolidate these two groups into one. we provide the user a matrix of correlations between industries in various countries. the broad country index used is the MSCI index. along with the family of industry speciﬁc indices we use for each country.1 Data As mentioned above. we ﬁnd that the correlation between the Health Care and Pharmaceuticals indices is over 98%. we list the industries for which we provide indices in one or more of the countries. We choose these industry groups by beginning with the major groups used by Standard & Poor for the United States. insufﬁcient industry index data was available and we utilize only the data for the broad country index.8 Countries and respective index families Country Australia Austria Belgium Canada Finland France Germany Greece Hong Kong Indonesia Italy Japan Korea Malaysia Index family ASX Mexico New Zealand Norway Philippines Poland Portugal Singapore South Africa Spain Sweden Switzerland Thailand United Kingdom United States Country Index family Mexican SE Toronto SE Helsinki SE SBF CDAX Athens SE Hang Seng Milan SE Topix Korea SE KLSE Oslo SE Philippine SE AllSingapore Stockholm SE SPI SET FTSEA S&P In Table 8. Table 8. For instance. For countries where no index family appears. In this section. CreditMetrics™—Technical Document . we list the countries for which we provide data. For each country. we discuss the data and the methods by which we construct this matrix.
the user may then choose to proxy the French chemical index with a combination of the MSCI France index and the MSCI worldwide chemicals index. where there is no countryindustry index.9 Industry groupings with codes Grouping General country index Automobiles Banking & ﬁnance Broadcasting & media Chemicals Construction & building materials Electronics Energy Entertainment Food Health care & pharmaceuticals Hotels GNRL AUTO BFIN BMED CHEM CSTR ELCS ENRG ENMT FOOD HCAR HOTE Code Insurance Machinery Manufacturing Metals Mining Oil & gas – reﬁning & marketing Paper & forest products Publishing Technology Telecommunications Textiles Transportation Utilities Grouping Code INSU MACH MANU MMIN OGAS PAPR PUBL TECH TCOM TXTL TRAN UTIL Because the industry coverage in each country is not uniform. In the end. 19 worldwide industry indices. realizing that it may at times be more feasible to describe a company by a regional index rather than a set of country indices. Finally. Part II: Model Parameters . For the speciﬁc index titles used in each case.10. In a case such French chemicals. we select the indices for which at least three years of data are available. leaving us with 152 countryindustry indices. 28 country indices. The available countryindustry pairs are presented in Table 8. 8.5 Estimating asset correlations 95 Table 8. we also provide data on MSCI worldwide industry indices. and 6 regional indices. refer to Appendix I.Sec. we provide data on six MSCI regional indices.
Credit quality correlations Table 8.10 Countryindustry index availability MACH CHEM HOTE FOOD GNRL ENRG AUTO OGAS TECH TRAN TXTL PUBL CSTR ELCS PAPR UTIL BFIN INSU Total Country Australia Austria Belgium Canada Finland France Germany Greece Hong Kong Indonesia Italy Japan Korea Malaysia Mexico New Zealand Norway Philippines Poland Portugal Singapore South Africa Spain Sweden Switzerland Thailand United Kingdom United States MSCI worldwide Total X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X TCOM MANU BMED ENMT MMIN HCAR X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X 20 X X X X X X 12 X X X X X X 13 X X X X X 7 X X X 28 X X 5 X X X 6 X X X X 8 X X 2 X X X X 10 X X X X 6 X X X X 6 X X X X 12 X X X 6 X 1 X X X X 13 X X 4 X X X X 11 X X 3 X X 2 X X X 3 X X X X 7 X X X X 10 X X 4 10 1 1 15 5 6 11 3 3 1 6 16 11 3 4 1 3 3 1 1 3 3 1 6 5 17 17 24 19 199 CreditMetrics™—Technical Document .96 Chapter 8.
Sec.9] 1 COV ( k. σk σl We provide these correlations to the user. Thus. if we denote the tth week’s return on the kth index by R t ( k ) . and the weekly standard deviation of return by [8. The motivation for this is that we are interested in computing correlations which are valid over the longer horizons for which CreditMetrics will be used. 8. for all pairs of indices. whereas the statistics in RiskMetrics vary more from day to day. where T is 190 in our case.7] (k) R 1 = T ∑ Rt t=1 T (k) . The steps of this computation are as follows: Part II: Model Parameters . In addition. (l) and the correlation of weekly returns by [8. 8. and capture shorter term behavior. l ) = T–1 ∑ ( Rt t=1 T (k) –R (k) ) ( Rt (l) – R ). we compute the average weekly return on this index by [8.8] 1 T–1 σk = ∑ ( Rt t=1 T (k) –R (k) 2 ) . l = . As mentioned above. Note that our computations of volatilities and correlations differ from the standard volatility computations in RiskMetrics in that we weight all of the returns in each time series equally. Note also that the correlations we compute are based on historical weekly returns. and compute the mean and standard deviation of each return series. we consider the last 190 weekly returns.5 Estimating asset correlations 97 For each of the indices. The statistics here tend to be more stable over time.10] COV ( k. It is therefore an assumption of the model that the weekly correlations which we provide are accurate reﬂections of the quarterly or yearly asset moves which drive the CreditMetrics model. and discuss their use in the next section.2 Obligor correlations – example Now that we have described how to calculate correlations between countryindustry pairs. and reﬂect longer term trends. we compute the covariance of weekly returns by [8. l ) ρ k. we provide the user with the standard deviations (volatilities). it only remains to illustrate how to apply these to obtain correlations between individual obligors.5..
Assign weights to each obligor according to its participation in countries and industries.00 For the ﬁrm ABC.16 0. we consider two independent standard normal random variables. chemical index and ABC’s r ﬁrm speciﬁc standardized returns. and the idiosyncratic risk will be greater. Chemicals 1.25% Correlations Germany Insurance 0.S. The remainder is explained by ABC’s ﬁrm speciﬁc movements.34 1. and the idiosyncratic portion of the risk to these companies is small. prices for companies with less market capitalization will move more independently of the indices.S. prices for companies with large market capitalization will track the indices closely.: Chemicals Germany: Insurance Germany: Banking Volatility 2.11.03% 2.08 Germany Banking 0. on the other hand. Express the standardized returns for each obligor as a weighted sum of the returns on the indices and a companyspeciﬁc component.34 U. ABC and XYZ. and that its equity returns are explained 90% by returns on the United States chemicals index and 10% by companyspeciﬁc movements. Use the weights along with the index correlations to compute the correlations between obligors. or idiosyncratic. We assume that these companyspeciﬁc movements are independent of the movements of the indices.09% 1. and specify how much of the obligor’s equity movements are not explained by the relevant indices. CreditMetrics™—Technical Document . Credit quality correlations 1.16 1. we are describing this obligor’s ﬁrmspeciﬁc.00 0. The volatilities listed are for weekly returns. risk.98 Chapter 8.S. Generally.11] r ( ABC ) = w1 r ( USCm ) ˆ + w2 r ( ABC ) . which represent the standardized returns of the U. Table 8. We then write ABC’s standardized returns as [8. To apply these weights and describe the standardized returns for the individual obligors. respectively.11 Volatilities and correlations for countryindustry pairs Index U. we need the volatilities and correlations of the relevant indices. Suppose we wish to compute the correlation between two obligors. Assume that XYZ participates 75% in German insurance and 25% in German banking and ﬁnance and that 20% of the movements in XYZ’s equity are companyspeciﬁc. chemicals index is 90% of the ﬁrm’s total volatility. and also independent of the companyspeciﬁc movements for all other companies. We present these in the Table 8. the volatility explained by the U. We will explain each of the steps above through an example. 3. 2. Assume that we decide that ABC participates only in the United States chemicals industry. r ( USCm ) and ( ABC ) ˆ .S.00 0.08 0. By specifying the amount of an obligor’s equity price movements are not explained by the relevant indices. Thus.
11 The above illustrates the method for computing correlations between pairs of obligors.13] 0. we may compute the correlation between ABC and XYZ by: [8.5 Estimating asset correlations 99 We know that 90% of ABC’s volatility is explained by the index.16] r ( ABC ) = 0. 1 For XYZ. Since the idiowhere r syncratic returns are independent of all the other returns.74 .74r ( DeIn ) + 0.25 ⋅ σ DeBa 0. 2 2 2 2 We then scale the weights so that the total volatility of the index portion of XYZ’s standardized returns is 80%.8 ⋅ . but generalized to handle obligors with participations in more industries and countries.90r ( USCm ) ˆ ( ABC ) . ˆ ˆ ( ABC ) and r ( XYZ ) are the idiosyncratic returns for the two ﬁrms.15] and [8.74 ⋅ ρ ( USC m.75 ⋅ 0.15 . and thus we know that w 1 = 0. In the next subsection.14] 0.25 ⋅ ρ ( D eIn. the weightings on standardized index returns which allow us to describe standardized ﬁrm returns. D eBa ) ⋅ σ DeIn ⋅ σ DeBa = 0.25 ⋅ σ DeBa + 2 ⋅ 0.9 . 2 Finally. D eIn ) + 0.44 .15 ⋅ ρ ( USC m. and suggests a more general framework.8 ⋅ . We also know that the total volatility must be one (since the returns are standardized).15r ˆ + 0.75 ⋅ σ DeIn + 0. D eBa ) = 0. that is. Thus. we know that the weight on the idiosyncratic return must be At this point. We ﬁrst ﬁgure the volatility of the index movements for XYZ. + 0.90 ⋅ 0. ˆ σ and the weight on the German banking index is [8. and thus w 2 = 1 – w 2 = 0.8 = 0. by [8.= 0.44r ( DeBa ) r ( XYZ ) = 0. XYZ ) = 0. we present the same methods. ˆ σ 1 – 0.12] ˆ σ = 0.= 0.90 ⋅ 0. Recall that for our example we describe the returns for ABC and XYZ by: [8. the weight on the German insurance index is [8.6 . that is. we have what we will refer to as each ﬁrm’s standard weights. we proceed in a similar vein.6r ( XYZ ) . in order that the total standardized return of XYZ have variance one. 8. Part II: Model Parameters .75 ⋅ σ DeIn 0.Sec.17] ρ ( ABC.017 . the volatility of an index formed by 75% German insurance and 25% German banking.
19] w3 σ3 w1 σ1 w2 σ2 ˆ ˆ ˆ w 1 = α ⋅ . ˆ ˆ ˆ σ σ σ 1–α . First.2 σ 1 σ 2 + 2w 2 w 3 ρ 2 . w 2 . or 3 to 1. and we wish to compute the equity correlations between these ﬁrms. Since the weightings are on both the indices and the idiosyncratic components.3 σ 1 σ 3 ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ 2 2 2 2 2 2 Scale the weights on each index such that the indices represent only α of the volatility of the ﬁrm’s standardized returns. C . consider a ﬁrm with industry participations of ˆ w 1 .100 Chapter 8.. [8. we provide generalizations of the methods above for computing standard weights and for calculating correlations from these weights.. Now suppose we have n different ﬁrms with standard weightings on m indices. where the indices account for α of the movements of the ﬁrm’s equity. Credit quality correlations Note that the index volatilities do not actually enter into the correlation calculations. Let the correlation matrix for the indices be denoted by C. but since the industry with 75% participation (insurance) is more volatile than the other industry (banking).18] σ = w 1 σ 1 + w 2 σ 2 + w 3 σ 3 + 2w 1 w 2 ρ 1 . to compute standard weights. and constructed as below: CreditMetrics™—Technical Document . w 2 = α ⋅ . where industry participation is split 75% and 25%. 8. we need to create a correlation matrix. which covers both of these. but do play a role when we convert industry participations to standard weights.5. This matrix will be m+n by m+n. the standard weight on insurance is actually more than three times greater than the standard weight on banking. This allows us to account for cases like our example.3 Obligor correlations – generalization To complete our treatment of obligor correlations. ˆ ˆ We compute the ﬁrm’s standard weights in the following steps: Compute the volatility of the weighted index for the ﬁrm. that is. and w 3 = α ⋅ .. and w 3 . The scaling is as below: [8.3 σ 2 σ 3 + 2w 1 w 3 ρ 1 . 2 Compute the weight on the idiosyncratic returns by taking The generalization to the case of four or more indices should be clear.
. in the kth column of W.08 0 1 0. 0 … … … … … … … … … … 0 . .. For the example in the previous subsection (where m = 3 and n = 2 ). reﬂecting that each ﬁrm’s idiosyncratic component has correlation one with itself and is independent of the other ﬁrms’ idiosyncratic components. the ﬁrst m entries will give the ﬁrst ﬁrm’s weights on the indices. . . .60 [8. 0 1 0 … . and the remaining entries will be zero. 0 …… …… …… …… 0 . . . .20] We then create a m+n by n weight matrix W.5 Estimating asset correlations 101 m columns n columns C 0 . . the upper left of C is the m by m matrix C. … .15 . 1 Thus. .08 0 0 0. and the remainder consists of only zeros. . . Part II: Model Parameters n rows . . and each row represents weights on indices and idiosyncratic components. . 0… … 1 0 . For our example.16 0.90 0 0 0. reﬂecting that there is no correlation between the idiosyncratic components and the indices. . .16 C = 0. . . .74 0. .34 0 0. we would have 1 0. . .. . the lower right is the n by n identity matrix. 8. the matrix W would be given by 0. . . .34 1 0 0 0 1 0 0 0 0 0 0 . 0 .Sec. the m+n+k entry will give the ﬁrm’s idiosyncratic weight. . . . 0 m rows 0 .21] W = The n by n matrix giving the correlations between all of the ﬁrms is then given by W′ ⋅ C ⋅ W . Thus. 0 0. . representing the correlations between indices. where each column represents a different ﬁrm. . .44 0 0 0. 0 1 [8.
102 Chapter 8. Credit quality correlations CreditMetrics™—Technical Document .
103 Applications Part III .
104 CreditMetrics™—Technical Document .
on the other. we have detailed an analytic approach to compute the mean and standard deviation of portfolio value change. describe the Monte Carlo approach to this distribution. setting credit risk limits. we are able to describe in much more detail the distribution of portfolio value changes. Both issues – alternative measures of risk and computations for a larger portfolio – point us to a central theme of this section: simulation. • Chapter 10: Simulation. We choose a portfolio of 20 instruments of varying maturities and rating and specify the asset correlations between their issuers. Implementation of a simulation approach involves a tradeoff. and discuss how to produce percentile levels as well as marginal statistics. We focus on computing advanced (beyond the mean and standard deviation) risk estimates for larger portfolios. and assessing economic capital. We then utilize the simulation approach of the previous section to estimate certain risk statistics and interpret these results in the context of the portfolio. Part III is composed of four chapters which describe the methods and discuss the outputs of the CreditMetrics methodology for larger portfolios. • Chapter 11: Portfolio example. and discussed the inputs to these calculations. • Chapter 12: Application of model outputs. We will continue to discuss this tradeoff as we go. The chapters dealing with simulation focus on computing advanced (beyond the mean and standard deviation) risk estimates. We consider how the analysis in Chapter 11 might lead to risk management actions such as prioritizing risk reduction. By this we mean the generation of future portfolio scenarios according to the models already discussed. This section is organized as follows: • Chapter 9: Analytic portfolio calculation. presented calculations for one. In this section we discuss approaches to computing risk measures other than standard deviation and apply the CreditMetrics methodology to a larger portfolio. Part III: Applications . We extend the methods discussed in Chapter 3 for computing the standard deviation and marginal standard deviation to a large (more than two instruments) portfolio. On the one hand. We address the assumptions necessary to specify the portfolio distribution completely.and twoasset portfolios. we introduce noise into what has been an exact solution for the risk estimates.105 Overview of Part III To this point.
106 CreditMetrics™—Technical Document .
30 2.22 1.1 Transition probabilities (%) Rating AAA AA A BBB BB B CCC Default Firm 1 Transition probability (%) Firm 2 0. We take the ﬁrst two of these bonds to be issued by the BBB and A rated ﬁrms of Chapter 3 and the third to be a twoyear bond paying a 10% coupon and issued by a CCC rated ﬁrm.09 2.05 5.93 5. and the Firm 3 issue an amount of 1mm. Denote by V 1 .01 0.86 19.24 64.107 Chapter 9.33 5. We present transition probabilities for the three ﬁrms in Table 9. 2. We refrained from extending this computation to larger portfolios.30 1. and 3.00 0. and is detailed in Appendix A. Analytic portfolio calculation In Chapter 3.1 Threeasset portfolio Our example is a portfolio consisting of three assets.27 91.18 0. Table 9. We will refer to the ﬁrms respectively as Firms 1. we discussed the computation of the standard deviation of value change for a portfolio of two instruments. In this chapter.26 0. and revaluations in Table 9. the values at the end of the risk horizon of the three respective issues. all annual coupon bonds.95 86.74 0.79 . and V 3 . 9.12 0. we illustrate this point for a threeasset portfolio. stating that the standard deviation of value for larger portfolios involves no different calculations than the standard deviation for twoasset portfolios.52 0. Suppose that the Firm 1 issue has a notional amount of 4mm. V 2 .1 below. The generalization of these calculations to portfolios of arbitrary size is straightforward.2.06 Firm 3 0.02 0.38 11. and discuss as well the calculation of marginal standard deviations for this portfolio.17 0. the Firm 2 issue an amount of 2mm.22 0.
113 Firm 3 1.161 1.132 2. We may also compute the variance of value for each of the three assets.302 4.157 1. 2 2 σ ( V 2 ) . 2 2 2 Note that since the standard deviations are in units of ($mm).014 .023 2.3] σ p = σ ( V 1 ) + σ ( V 2 ) + σ ( V 3 ) + 2 ⋅ COV ( V 1.044 .126 2.4] σ ( V 1 + V 2 ) = σ ( V 1 ) + 2 ⋅ COV ( V 1. we could use the standard formula [9. obtaining [9.375 4.108 Chapter 9.38mm . the units for σ 2 ( V 1 ) . the standard deviation of value for the portfolio.142 1.162 1. giving a portfolio mean of µ p = $7. Analytic portfolio calculation Table 9.28mm .774 1. and σ ( V 3 ) are ($mm)2.125 4.2] σ ( V 1 ) = 0. and µ 3 = $0.346 2. we may compute the mean value for each issue: [9. 2 2 2 2 – σ (V1) – σ (V2) – σ (V3) CreditMetrics™—Technical Document 2 2 2 .001 .161 1. Now to compute σ p .081 3.1] µ 1 = $4. 2 2 2 we may express σ p by [9. µ 2 = $2. V 2 ) + σ ( V 2 ).5] σ p = σ (V1 + V2) + σ (V1 + V3) + σ (V2 + V3) . noting that [9. and σ ( V 3 ) = 0. V 2 ) .137 1. V 3 ) This would require the calculation of the various covariance terms.130 2. 2 2 2 2 + 2 ⋅ COV ( V 1.97mm .924 3. σ ( V 2 ) = 0.056 0551 BBB BB B CCC Default Utilizing the methods of Chapter 2 and the information in the tables above. Alternatively.063 2.12mm .368 Firm 2 2.346 4.2 Instrument values in future ratings ($mm) Value of issue ($mm) Future rating AAA AA A Firm 1 4.028 1. V 3 ) + 2 ⋅ COV ( V 2.
The calculation of portfolio variance in terms of the variance of twoasset subportfolios may seem unusual to those accustomed to the standard covariance approach.51 6. Similarly.40 6.3.3 differ from those in Table 3. In order to compute the variance for the portfolio containing only these assets.95 5.49 6. [9.49 6.70 5.26 AA 6.43 6. and so in principle.37 6. and apply Eq. we then compute σ 2 ( V 1 + V 2 ) = 0.48 6. Table 9.42 6.05 5. since [9.95 4. we present these values in Table 9.05 5.14 6.3.24 BB 6.37 5.2 since the notional amounts of the issues in these two cases are different. the BBB and A rated bonds.48 6.19 6.43 6. Thus.08 5. we specify that the asset correlations between the ﬁrst and third and between the second and third obligors are also 30%. compute the standard deviations of each.12 3.41 4.37 4. 2 2 2 2 we have COV ( V 1.2. we have described all of the portfolio calculations.26 A 6.90 Default 5.37 3. it only remains to identify each twoasset subportfolio. Along with these probabilities we need the values of this twoasset portfolio in each of the 64 joint rating states.15 CCC 6. σ ( V 1 ) ) and the standard deviations of twoasset subportfolios (e.. This allows us to then compute correlations between the asset values using Part III: Applications .47 4. [3.305mm .2 and Table 9. we utilize the joint transition probabilities in Table 3.43 6.46 6.0125 and COV ( V 2. we apply Eq. We present the twoasset case again as a review.15 Applying Eq.48 4.33 5.99 5.12 6.Sec. V 3 ) = 0.40 6. we obtain COV ( V 1.39 5.10 4.40 5.3.48 6.51 6.25 BBB 6. and thus σ p = $0.6] σ (V1 + V2) – σ (V1) – σ (V2) COV ( V 1.50 6. We remark that we have all of the information necessary to compute the covariances and correlations between our three assets. V 2 ) = . Thus.04 5.g.6] to obtain 2 σ p = 0. [9.018 .21 6.5]. Note that the values in Table 9. σ ( V 1 + V 2 ) ).11 5.47 6. The standard deviation for twoasset portfolios was covered in Chapter 3. In a similar fashion.2 and the values in Table 9.15 6.48 4. V 3 ) = 0.3 Values of a twoasset portfolio in future ratings ($mm) New rating for Firm 1 (currently BBB) AAA AA A BBB BB B CCC Default New rating for Firm 2 (currently A) AAA 6.44 6. with an assumed asset correlation of 30%. Consider the ﬁrst pair of assets.43 6. which are an output of the asset value model of the previous chapter.37 6.06 5.21 6.1] to the probabilities in Table 3.41 6. Finally.14 5. and then create analogs to Table 3. to complete our computation of σ p .0015 .50 6.083 and σ ( V 2 + V 3 ) = 0.21 6.33 6.093 .50 6.0030 .86 5.051 . This allows us to compute 2 2 σ ( V 1 + V 3 ) = 0. V 2 ) = 0.46 4.1 Threeasset portfolio 109 The above formula has the attractive feature of expressing the portfolio standard deviation in terms of the standard deviations of single assets (e.19 B 6.g. 9.
225mm . we might also wish to characterize the riskiness of this instrument independently of its size. However.1% in our example) as the percent portfolio standard deviation. V 2 ) = .2 Marginal standard deviation As deﬁned in Section 3. we need CreditMetrics™—Technical Document . We refer to this ﬁgure. The risk measures produced in this section may strike the reader as a bit small. Thus. 9.051 . Consider the Firm 1 issue in our portfolio above. Alternatively. Thus.. To this end. Note that if we consider the Firm 1 issue alone. The marginal standard deviation of the Firm 1 issue is ˆ then the difference between the absolute portfolio standard deviation and this ﬁgure. If we remove the Firm 1 issue. However.2% . making the new portfolio standard deviation σ p = $0. or ˆ σ p – σ p = $0. The difference between marginal and standalone statistics gives us an idea of the effect of diversiﬁcation on the portfolio. σ (V1) × σ (V2) 2 2 We then have CORR ( V 1. as the percent marginal standard deviation of this issue.4% . since we now are able to compute the standard deviation for a portfolio of arbitrary size. [9. These notions of absolute and percent measures will be used for other portfolio statistics. We refer to σp as the absolute measure of the portfolio standard deviation. Analytic portfolio calculation [9.46mm. with the percent statistic always representing the absolute statistic as a fraction of the mean portfolio value. we have seen that the marginal impact of the Firm 1 issue is only $0. we may express this risk in percentage terms.080mm . V 3 ) = 50.1] yields the same value for σ p as we computed above. we may express the marginal standard deviation as a percentage of µ 1 . and thus that we beneﬁt from the fact that this issue is not in fact perfectly correlated with the others. particularly in light of the riskiness of the CCC rated issue in our example. This might be explained by the fact that the size of this issue is quite small in comparison with the other assets in the portfolio. it may be that to adequately describe the riskiness of the portfolio.3.080mm if we liquidate the Firm 1 issue While this is a measure of the absolute risk contributed by the Firm 1 issue. V 2 ) = 40.080mm.9% in this case. its standard deviation of value is $0. the calculation of marginal standard deviations is clear. then the new 2 ˆ portfolio variance is given by σ p = σ 2 ( V 2 + V 3 ) = 0. We present the details of this in Appendix A. CORR ( V 1. It is a simple matter then to check that the standard formula Eq. To extend this calculation to larger portfolios is straightforward. We have seen that the standard deviation for the entire portfolio is $0. V 3 ) = 45.110 Chapter 9. If this asset were perfectly correlated with the other assets in the portfolio.1% . we see that we can reduce the total portfolio standard deviation by $0. 1. since we have only considered the standard deviation to this point. its marginal impact on the portfolio standard deviation would be exactly this amount.7] COV ( V 1. the mean value of the Firm 1 issue. V 2 ) CORR ( V 1. the marginal standard deviation for a given instrument in a portfolio is the difference between the standard deviation for the entire portfolio and the standard deviation for the portfolio not including the instrument in question.117mm. and CORR ( V 2. we thus refer to σp/µp (which is equal to 4.
which is the subject of the following two chapters. 9. In order to obtain this higher order information.2 Marginal standard deviation 111 more detailed information about the portfolio distribution. Part III: Applications .Sec. it will be necessary to perform a simulation based analysis.
Analytic portfolio calculation CreditMetrics™—Technical Document .112 Chapter 9.
For each scenario. we limit the available of statistics that can be estimated. Precision. Establish asset return thresholds for the obligors in the portfolio. Generate scenarios of asset returns according to the normal distribution. no error in the risk estimates. Particularly for smaller portfolios. 2. number 1 above is no longer true. we present in this chapter a simulation approach known as “Monte Carlo. We may then choose to report any number of descriptive statistics for this distribution. Summarize results. One is that for large portfolios. we have discussed methods to compute the standard deviation of portfolio value. No random noise is introduced in the calculations and. 2. This step gives us a large number of possible future portfolio values. yet we have also stressed that this may not be a meaningful measure of the credit risk of the portfolio. we will discuss how to generate scenarios of future credit ratings for the obligors in our portfolio. Given the value scenarios generated in the previous steps. and CCC rated ﬁrms. We will continue to consider the example portfolio of the previous chapter: three twoyear par bonds issued by BBB. However. 10. Throughout this document. For our purposes. it has also two principal disadvantages. The other is that by restricting ourselves to analytical approaches. and thus can be computed more quickly. 2. we revalue the portfolio to reﬂect the new credit ratings. 3. Generate scenarios. and $1mm. 3. The steps to scenario generation are as follows: 1. Speed. Value portfolio. The notional values of these bonds are $4mm. Simulation Our methodology up to this point has focused on analytic estimates of risk. therefore. we have an estimate for the distribution of portfolio values.4.” The three sections of this chapter treat the three steps to a Monte Carlo simulation: 1. A.113 Chapter 10. that is. Map the asset return scenarios to credit rating scenarios. To provide a methodology that better describes the distribution of portfolio values. estimates which are computed directly from formulas implied by the models we assume. Each scenario corresponds to a possible “state of the world” at the end of our risk horizon. . We will rely heavily on the asset value model discussed in Section 8. the direct calculations require fewer operations.1 Scenario generation In this section. the “state of the world” is just the credit rating of each of the obligors in our portfolio. $2mm. This analytical approach has two advantages: 1.
63 2.51 2.85 Table 10.02 0.4.21 the asset return thresholds for the three ﬁrms. and specify the correlations for each pair of ﬁrms2. we state that the asset returns in for each ﬁrm are normally distributed.53 1. 1 Recall the comment at the end of Chapter 8 that asset return volatility does not affect the joint probabilities of rating changes.86 2.22 0. 2 CreditMetrics™—Technical Document . a return less than ZBB (but greater than ZB) corresponds to a rating of BB.86 2.11 1.75 2. For this reason.12 1.09 2.33 5.02 0.2 Recall that the thresholds are labeled such that a return falling just below a given thresholds corresponds to the rating in the threshold’s subscript.78 1.114 Chapter 10.18 2. the assumption is that the joint distribution of the asset returns of any collection of ﬁrms is multivariate normal.38 11. Technically. That is.18 Firm 2 0.24 Firm 3 2. Transition probabilities (%) Transition Probability (%) Rating AAA AA A BBB BB B CCC Default Firm 1 0. In order to describe how the asset values of the three ﬁrms move jointly.30 2.72 3.98 1. we restate the transition probabilities for the three issues. For our example.1 below. we assume the correlations in Table 10.27 91.54 2.22 1.74 1.49 2.24 64.52 0.17 0.74 0. which are obtained using the methods of Section 8.30 1. Asset return thresholds ZAA ZA ZBBB ZBB ZB ZCCC ZDef Threshold Firm 1 3.91 Firm 2 3.00 0.86 19.05 5.1 We then present in Table 10.12 0.79 Table 10.06 Firm 3 0.95 86.26 0.93 5.30 2. Simulation In Table 10. we may consider standardized asset returns.01 0.3. and report the thresholds for these.19 3.
6874 0.4690 0.1532 0.5. the three numbers represent the standardized asset return for each of the three ﬁrms.5191 0.0249 Firm 1 0.4 Scenario 1 2 3 4 5 6 7 8 9 10 Scenarios for standardized asset returns 0.0 Generating scenarios for the asset returns of our three obligors is a simple matter of generating correlated.49 from Table 10. Continuing this process. singular value decomposition. Strang [88].1 Firm 2 0. etc. we list ten scenarios which might be produced by such a procedure.2832 1. for example.1 0. we may ﬁll in Table 10.5639 0.7769 2. This corresponds to a new rating of BB.0 0.2 Firm 3 0.0606 0. it is only necessary to assign ratings to the asset return scenarios.4642 Firm 2 0.8750 2.0 0.2 1.1 Scenario generation 115 Table 10.0646. In Table 10.9276 0.1252 0.1778 1. the return is –2. which falls between ZB (–2.1060 0.1060.2) and ZBB (–1.3533 Firm 3 To fully specify our scenarios.3 Correlation matrix for example portfolio Firm 1 Firm 2 Firm 3 Firm 1 1. which completes the process of scenario generation PartI III: Applications .1510 0.4.6994 1.3 0. normally distributed variates. For example.9479 1. which falls between ZBB and ZBBB for this name.Sec.1202 0. – for discussions of which see. Table 10.6342 2.1631 0. 10.7068 1.5570 1.0646 0. There are a number of methods for doing this – Cholesky factorization. For Firm 2. corresponding to a new rating of BBB.8480 0.7759 1.6454 0.6503 1.4.2996 2. The standardized return for Firm 1 is –2. consider scenario 2 of Table 10.2) for this name.18 from Table 10. In each scenario.3 1.
0646 0.1631 0.7759 1.1.0249 Firm 1 0. we generate a random recovery rate according to a beta distribution3 with these parameters4.8480 0. We discussed in Chapter 7 that recovery rates are not deterministic quantities but rather display a large amount of variation.6874 0. the scenarios do not correspond precisely to the transition probabilities in Table 10.5570 1. the new rating maps directly to a new value.2832 1.1202 0.2 Portfolio valuation For nondefault scenarios.0606 0. This is the topic of Appendix B. in our BBB rated senior unsecured issue.5191 0. As we generate more scenarios.3533 Firm 3 BBB BB BBB BBB BBB BBB BBB BBB A BBB Firm 1 New Rating A BBB A A A A A A AA A Firm 2 CCC CCC A Default CCC Default Default Default B CCC Firm 3 Table 10. For each default scenario. For default scenarios. the situation is slightly different. while the probability that this occurs is just 20%.9479 1. Note that we assume here that the recovery rate for a given obligor is independent of the value of all other instruments in the portfolio.6454 0. Firm 3 defaults.6994 1.5 Notice that for this small number of trials. For example. in four of the ten scenarios.8750 2. 4 CreditMetrics™—Technical Document . 10. These recovery rates then allow us to obtain the value in each default scenario.6.9276 0. For each scenario and each issue. we obtain a portfolio value for each scenario.4690 0. To model this variation. the recovery mean is 53% and the recovery standard deviation is 33%. but it is important to quantify how large we can expect the ﬂuctuations to be. 3 Recall that the beta distribution only produces numbers between zero and one. these ﬂuctuations become less prominent.7769 2. This variation of value in the case of default is a signiﬁcant contributor to risk. refer back to Chapter 4. Simulation Mapping return scenarios to rating scenarios Asset Return Scenario 1 2 3 4 5 6 7 8 9 10 0.1252 0. To recall the speciﬁcs of valuation. The results for the ﬁrst ten scenarios for our example are presented in Table 10.116 Chapter 10. (For example. we obtain the mean and standard deviation of recovery rate for each issue in our portfolio according to the issue’s seniority.6503 1. this step is no different here than in the previous chapters.1510 0.1060 0.5639 0.1778 1.4642 Firm 2 0.) These random ﬂuctuations are the source of the lack of precision in Monte Carlo estimation.2996 2.1532 0. In the end. so that we are assured of obtaining meaningful recovery rates.6342 2.7068 1.
this is not the case – the values of the Firm 3 issue in the default scenarios are different – since recovery rates are themselves uncertain.056 1. PartI III: Applications .302 4. the value is the same in scenarios with the same (nondefault) credit rating.126 2.302 4.484 7.613 7.126 2. but for an example in which we consider a larger portfolio and a larger number of scenarios.056 1. each default scenario requires an independently generated recovery rate.137 1.126 Value Firm 2 Firm 3 1. This plot is presented in Chart 10. Thus.161 0.063 2.081 4. we ﬁrst examine a plot of the ten scenarios for our example.1.697 6.302 4.Sec.182 6.657 1.126 2.302 4. so as to obtain more signiﬁcant results. In order to gain some intuition about the distribution of values.484 Note that for a given issue. 10.754 0.1. we would expect this plot to become more smooth.130 2.484 7.3 Summarizing the results At this point.056 Portfolio 7. we will examine a number of descriptive statistics for the scenarios we have created. and approach something like the histogram we will see in Chart 11.126 2. 10. The ﬁnal task is then to synthesize this information into meaningful risk estimates. In this section.302 4.126 2.126 2. For defaults.589 7.126 2.579 7.269 0.056 0.3 Summarizing the results 117 Table 10.6 Valuation of portfolio scenarios ($mm) Rating Scenario 1 2 3 4 5 6 7 8 9 10 Firm 1 BBB BB BBB BBB BBB BBB BBB BBB A BBB Firm 2 A BBB A A A A A A AA A CCC CCC A Default CCC Default Default Default B CCC Firm 3 Firm 1 4. we will examine the same statistics.151 1.346 4.302 4.085 7. For a larger number of scenarios.302 2. In the section to follow.200 7. we have created a number of possible future portfolio values.302 4.
This imprecision is due to simulation noise. if we wish to compute the marginal tenth percentile of the third asset in our portfolio (the CCC rated bond).05 0. For our scenarios.8 7.10 0. we take CreditMetrics™—Technical Document ..24mm and σ p = 1 N–1 ∑ (V i=1 N (i) – µ ) = $0.V(2).20 0. the mean and standard deviation may not be the best measures of risk in that. Thus. we have considered only statistics which describe the portfolio distribution.25 0. In general. we cannot infer percentile levels from the standard deviation.00 Frequency plot of portfolio scenarios 6.37mm 2 where N is the number of scenarios (in our case.4 7. to compute the tenth percentile given our scenarios. this level is between $6. To this end.5 7.0 7. we will describe marginal statistics.1 Frequency 0. we choose a level (x) at which one of the ten scenarios is less than x and the other nine scenarios are greater than x. Estimates of percentile levels are straightforward.4 6.58mm and $6.6 6.15 0. For example. This concept may be generalized. but we will see in the next chapter that as we consider more scenarios.70mm. The ﬁrst statistics we examine are those which we are able to compute analytically: the mean and standard deviation of future portfolio value. As we have mentioned before. the marginal statistic for a particular asset is the difference between that statistic for the entire portfolio and that statistic for the portfolio not including the asset in question.118 Chapter 10. Let V(1).1] 1 µ p = N ∑ i=1 N V (i) = $7. Then we may compute the sample mean (µ) and standard deviation (σ) of the scenarios as follows: [10. we begin to see the heavy downside tail typical of credit portfolio distributions. We are thus motivated to perform simulations in order to capture more information about the distribution of values.. our estimates of percentiles become more precise. indicate the portfolio value in the respective scenarios. We have discussed marginal standard deviations previously. We would also like to consider individual assets and to ascertain how much risk each asset contributes to the portfolio. N=10). percentile) discussed above. since the distribution of values is not normal.7 Even for small number of scenarios.2 7.. Simulation Chart 10. To this point.30 0.V(3). and we may compute a marginal analog of any of the statistics (standard deviation.
and third assets. second.2] θ 10 ( V 1 + V 2 + V 3 ) – θ 10 ( V 1 + V 2 ) where V1. how close we expect our estimates of various portfolio statistics to be to their true value. given the number of scenarios which are generated. Quantifying the precision of simulation based statistics is the subject of Appendix B. any set of scenarios may not produce a sample mean or sample 5th percentile which is equal to the true mean or 5th percentile for the portfolio. Thus. it is important to quantify. This marginal ﬁgure may be interpreted as the amount by which we could decrease the risk on our portfolio by removing the CCC rated bond.29mm. while that for just the ﬁrst two assets is $6. and thus the marginal standard deviation for the third asset is $0. the statistics obtained through Monte Carlo simulation are subject to ﬂuctuations. For the scenarios above. the tenth percentile for the entire portfolio is $6. As we have mentioned a number of times. and θ10 represents the tenth percentile of the values in question.35mm.3 Summarizing the results 119 [10. V2. PartI III: Applications . and generate enough scenarios to achieve this.Sec. and V3 represent the future values of the ﬁrst. respectively. a reasonable way to choose the number of scenarios to be generated is to specify some desired level of precision for a particular statistic. In fact. 10.64mm.
120 Chapter 10. Simulation CreditMetrics™—Technical Document .
Portfolio example In this chapter.000 1.087.000 2.000 Maturity (years) 3 4 3 4 3 4 2 8 2 2 4 5 3 2 2 4 6 5 3 5 Market value 7.263.1.000.523 2.263.000 1.315.386. we examine a more realistic example portfolio and discuss the results of a simulationbased analysis of this portfolio. The bonds are listed in Table 11. the credit rating determines the distribution of future credit rating.000.020.821. For the portfolio.000 1.000 10.000.432 1. 11.121 Chapter 11.246 2.127.046 1.000 8.000.000.181.831 1.603 3.000. however.000.784 Recall that for each asset. Example portfolio Asset 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Credit rating AAA AA A BBB BB B CCC A BB A A A B B B B BBB BBB BBB AA Principal amount 7.000.000.181.1 The example portfolio In this chapter.000 1.071 5.000 3.000. we consider a portfolio of 20 corporate bonds (each with a different issuer) of varying rating and maturity.000 1.000.000.1.000 1. For this example.000. we must also specify the asset correlations in order to describe the distribution of future ratings and values.000 5.000 1.118.000.178 6.2. and thus also the distribution of future value.181. . The risk estimates are no different than those in the previous chapter.527.000 1.000 2.000.720 10.246 1.841 3.000.000 3.483.049 1. but should take on more meaning here in the context of a larger portfolio. The total market value of the portfolio is $68mm.000 5.322 705.000.000. Table 11.611 1.268 1.628 14. we assume the correlations in Table 11.229.177.523 1.000 1.120.641 1.409 1.000 1.189.154.000 600.000.
45 0.15 0.2 0.15 0.15 0.15 0.1 0.15 0.1 0.15 0.2 0. Portfolio example Table 11.1 0.1 0.1 0.15 0.1 0.15 0.1 0.35 0.2 0.2 0.15 0.15 0.1 0.15 0.15 0.2 Simulation results Using the methodology of the previous chapter.2 0.1 0.2 0.35 0.15 0.1 0.2 0.1 0.15 0.3.2 0.15 0.35 1 0.1 0.2 0.35 0.2 0.35 0. Note the axes on each chart carefully.1 0.1 0.1 0.1 0.2 0.35 0.15 0.1 0.2 0.25 0.1 0.45 0.35 0.25 18 0.2 0.15 0.2 0.2 0.2 0.2 0.1 0.35 1 0.1 0.2 0.35 0.15 0.2 0.2 0.1 0.65 20 0.1 0.2 0.15 0.25 1 0.1 9 0.1 0.25 17 0.15 0.45 1 0.35 0.2 0.000 portfolio scenarios.1 0. For each scenario.15 0.1 0.1 0.2 0.1 11 0.1 0.1 5 0.1 0.1 0.15 0. while the correlations between these groups are lower.15 0.2 0.15 0.35 0.15 0.1 0.45 0.15 0.1 0.35 0.45 1 0.2 0.1 0.15 0.2 0.2 0.35 0.2 0.45 0.35 0.1 0.15 0.45 0.1 0.25 0.35 0.1 0.15 0.1 0.1 0.2 0.2 Asset correlations for example portfolio 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 1 0. In Charts 11.2 0.1 0. in the shaded areas of the table) within which the asset correlations are relatively high.15 0.15 0.1 0.2 0.15 0.2 0.1 0.2 0.1 0.1 0.2 0.2 0.2 0.1 0.2 0.2 0.2 0.15 0.35 0.1 through 11. is ten times smaller in the second chart than in the ﬁrst.35 0.1 0.1 0.15 0.2 0.000 possible future occurrences in one year’s time of the credit ratings for each of our issues. the second moves a bit further into the left tail of the distribution.45 0.15 0.15 0.1 0.15 1 0.2 0.1 0.1 0.2 0.1 0.1 0. and the third shows the distribution of the most extreme 5% of all cases.2 0. This might be the case for a portfolio containing issues from ﬁrms in ﬁve different industries.35 1 0.15 0.2 0.45 0.15 0.65 1 Observe that there are ﬁve groups of issuers (those for assets 14.2 0.15 0.45 0.15 0.1 0.1 0.45 0.25 0.1 0.35 1 0.15 0.15 0.1 0.1 0.1 0.122 Chapter 11.2 0.45 1 0. and 1920.2 0.1 0.2 0.1 0.45 1 0.45 0.2 0.1 0. 1115.35 0.1 0.2 0.2 0.1 0. 20.15 0.1 2 0.35 0.25 0.15 0.15 0.1 0.2 0.2 0.45 1 0.2 0.2 0.2 0.1 0. and twenty times smaller in the third chart than in the second.15 0.1 0.45 0.1 0.2 0.2 0.1 7 0.15 0.55 0.15 0.1 0.15 0.1 13 0.1 0.15 0.2 0.15 0.1 0.15 0.15 0.1 0.45 0.1 0.15 0. 1618.55 1 0.15 0.2 1 0. The vertical axis.1 10 0.1 0.1 3 0.1 0.1 0.1 0.45 0.45 0.1 0.1 0.2 0.1 0.15 0.35 0.15 0. which represents relative frequency.1 0.2 0.25 0.1 0.45 0.2 0.1 0.1 0.2 1 0.35 0.15 0.1 0.45 0.1 16 0.15 0.1 0.2 0.2 0.1 12 0.1 8 0. the correlations between ﬁrms in a given industry are high.1 0.15 0.15 0.1 0.1 0.45 0.1 0.1 0.55 0.1 0.2 0.2 0.45 0.2 0.25 0.2 0.55 0.1 0.2 0. 11.15 0.35 0.55 1 0.2 0.1 0.1 0.1 0.1 0.2 0. CreditMetrics™—Technical Document .1 0.45 0.15 0.15 0. we present histograms of the portfolio value scenarios.2 0.35 0.2 0.15 0.2 0. we then obtain a portfolio value for one year into the future.45 0.55 0.2 0.1 0.1 0.45 0.15 0.35 0.15 0.15 0.15 0.15 0.1 14 0.2 0.15 0.1 0.1 0.2 0.45 1 0.25 0.1 0. we generate 20.2 0.15 0.35 0.1 0.15 0.1 0.35 0. 610.15 0.2 0.1 15 0.1 6 0.35 1 0.2 0.35 0.45 1 0.15 0.15 0. while correlations across industries are lower.2 0.1 0.15 0.15 0.45 0.1 0.15 0.45 0.1 0.2 0.45 0.15 0.2 0.1 0.1 0. The ﬁrst chart illustrates the distribution of the most common scenarios.15 0.25 19 0.25 0. that is.15 0.45 0.1 0.1 0.2 0.2 0.2 0.2 0.2 0.1 4 0.15 0.
Sec.2 1.0 0.6 67.8 68.2 Simulation results 123 Chart 11.2 67.0 See Cht. 11.2 0 59 65 66.8 67.3 Portfolio value ($mm) Part III: Applications .4 66.1 Histogram of future portfolio values – upper 85% of scenarios Relative frequency 10 9 8 7 6 5 4 3 2 1 0 65 66.3 67.6 0. 11.2 Histogram of future portfolio values – scenarios between 95th and 65th percentiles Relative frequency 1.8 0.4 0.5 See Cht.9 67. 11.2 Portfolio value ($mm) Chart 11.
Regardless of the particulars of the shape of the value distribution. The second observation is the odd bimodal structure of the distribution.005 0 56 58 60 62 64 65 Portfolio value ($mm) We may make several interesting observations of these charts.8mm in Table 11. • Standard deviation of portfolio value (σ) = $1. For our case. in well over half of the scenarios.136. since the loss distribution is not normal. the distribution of portfolio value is driven primarily by the number of issues which default. one feature persists: the heavy downward skew.020 0. by far the most common occurrence (almost 9% of all scenarios.025 0. Portfolio example Chart 11. The two other humps further to the left in the distribution represent scenarios with two and three defaults. the ﬁrst two statistics we present are the mean and standard deviation of the portfolio value.3 Histogram of future portfolio values – lower 5% of scenarios Relative frequency 0. while for smaller ones. Further. and the portfolio appreciates. Our example distribution is no different.077. these humps become even more smoothed out.124 Chapter 11.888. we cannot infer conﬁdence levels from these parameters. This is due to the fact that default events produce much more signiﬁcant value changes than any other rating migrations. respectively.2mm) represents scenarios in which one issue defaults. For larger portfolios.030 0. the humps are generally more prominent.284.010 0. We can however estimate percentiles directly from our scenarios. displaying a large probability of a marginal increase in value along with a small probability of a more signiﬁcant drop in value. CreditMetrics™—Technical Document . Thus. As in the previous chapter. As we have mentioned before. there are no signiﬁcant credit events.015 0.1) is that none of the issuers undergoes a rating change. The second hump in the distribution (the one between $67mm and $67. First. exhibited by the spike near $67. we have: • Mean portfolio value (µ) = $67. the mean and standard deviation may not be the best measures of risk in that.
and examine how these conﬁdence bands evolve as we increase the number of scenarios which we consider. we also give the percentiles which we would have estimated had we utilized the sample mean and standard deviation.33σ 64. By contrast. if we wish to compute the 5th percentile (the level below which we estimate that 5% of portfolio values fall).36 µ−3. because the portfolio distribution is not normal. we present various percentiles of our scenarios of future portfolio values.15 µ 67.93 67.80 64. On the other hand. For the 20.42 µ−1.64 µ−2.4.Sec.3 below.000 scenarios in ascending order and take the 1000th of these sorted scenarios (that is.84 57.06mm. below the mean.58σ 64. Part III: Applications . For comparison and in order to illustrate the nonnormality of the portfolio distribution.9 standard deviations.97 Normal distribution Portfolio value Formula ($mm) µ+1.06 µ−2.65σ 69.5% of the time (or one year in forty). where in the analytic approach. Table 11.09σ 63. we are only able to compute two statistics. the normal assumption leads to a more pessimistic forecast: there is a 50% chance that the portfolio is less valuable than the mean value of $67.69mm. respectively. and assumed that the distribution was normal.3 Percentiles of future portfolio values ($mm) Actual scenarios Portfolio value ($mm) 67. a much more optimistic risk estimate. we would have estimated that this percentile would correspond to only a drop to $65. 11.3 Assessing precision 125 For example.96σ 65. $64. 11. (Our assumption is then that since 5% of the simulated changes in value were less than $5.) Here we see the advantage of the simulation approach. In Table 11.97 62.28 µ−1.5% 0.000 scenarios in our example.65σ 65. in that we can estimate arbitrary percentile levels. if we examine the median value change (the 50% level).1% Using the scenarios.98 63. and thus to a greater than 50% chance that the portfolio value will exceed its mean.3 Assessing precision In this section. we estimate that 2. our portfolio in one year will drop in value to $63. If we had used a normal assumption. This is further evidence that it is best not to use the standard deviation to infer percentile levels for a credit portfolio.85 61.97mm or less. we have the results shown in Table 11. the scenarios point to a higher mean.28mm.98mm) as our estimate. we utilize the methods of Appendix B to give conﬁdence bands around our estimated statistics.77 Percentile 95% 50% 5% 2. we sort our 20. there is a 5% chance that the actual portfolio value change will be less than this level. Another interesting observation is that the 5th and 1st percentiles of the scenarios are 2 and 2.5% 1% 0.
11 Estimate 67.000 chance of shortfall For both the mean and standard deviation.126 Chapter 11.26 56. we may examine the evolution of our conﬁdence bands for each estimate as we consider more and more scenarios.1 percentile2 1 2 Lower bound 67. Portfolio example Table 11.4 67.85 61.000 Number of scenarios 15.02 62.000 CreditMetrics™—Technical Document .94 62.08 58.14 64.4 Portfolio value statistics with 90% conﬁdence levels ($mm) Statistic Mean portfolio value Standard deviation 5th percentile 1st percentile 0. the conﬁdence bands are reasonably tight. We present this information for the six statistics above in the following charts. we see that the true loss level could well be at least 10% greater than our estimate. we would be best off generating more scenarios.000 20.30 1.5 67. For the more extreme percentiles.66 61.4 Evolution of conﬁdence bands for portfolio mean ($mm) Mean portfolio value ($mm) 67.27 1.84 57.28 1. and for the 5th and 1st percentiles.73 1 in 200 chance of shortfall 1 in 1. and we feel assured of making decisions based on our estimates of these quantities.97 62.10 64. With regard to the question of how many scenarios we need to obtain precise estimates. Chart 11.3 67.000 10.2 67.1 67.97 Upper bound 67.0 Estimate Confidence bands 0 5.5 percentile1 0. If we desire estimates for these levels.17 65.98 62.
3 Assessing precision 127 Chart 11.000 Number of scenarios 15.00 Estimate 0.5 Evolution of conﬁdence bands for standard deviation ($mm) Standard deviation 1.75 1.000 Number of scenarios 15. 11.25 1.000 20.000 10.Sec.000 Part III: Applications .75 0 5.000 Chart 11.000 10.6 Evolution of conﬁdence bands for 5th percentile ($mm) Portfolio value ($mm) 66 64 Estimate 62 Confidence bands 60 58 0 5.000 20.50 Confidence bands 1.
000 10.7 Evolution of conﬁdence bands for 1st percentile ($mm) Portfolio value ($mm) 64 62 Estimate Confidence bands 60 58 56 0 5.000 Number of scenarios 15.5 percentile ($mm) Portfolio value ($mm) 64 Confidence bands 62 Estimate 60 58 56 0 5.8 Evolution of conﬁdence bands for 0.000 20.000 Number of scenarios 15.000 20. Portfolio example Chart 11.000 Chart 11.128 Chapter 11.000 10.000 CreditMetrics™—Technical Document .
For the most extreme percentile level. As an example.000 10. These four statistics are presented for each of the 20 assets in Table 11. Last. This is due to the fact that on average only one in one thousand scenarios produces a value which truly inﬂuences our estimate. For each asset in the portfolio. First.5. Third. which is just the standalone standard deviation expressed as a percentage of the mean value for the given asset. 11. that is the standard deviation of value for the asset computed without regard for the other instruments in the portfolio. the marginal risk of a given asset is the difference between the risk for the entire portfolio and the risk of the portfolio without the given asset.000 15. Part III: Applications .000 20. we will compute four numbers. Recall that for any risk measure. we express this ﬁgure in percent terms. we compute marginal statistics. Second. if we had been most concerned with the 5th percentile. the impact of the given asset on the total portfolio standard deviation.000 scenarios. we compute each asset’s standalone standard deviation of value. we compute the standalone percent standard deviation.1 percentile ($mm) Portfolio value ($mm) 62 60 58 56 54 Confidence bands 52 50 Estimate 0 5. In fact.4 Marginal risk measures To examine the contribution of each individual asset to the risk of the portfolio. we might have been satisﬁed with the precision of our estimate after only 5000 trials.Sec. note that the estimates and conﬁdence bands do not change frequently.000 Number of scenarios It is interesting to note here that few of the plots change beyond about 10. This suggests that to meaningfully improve our estimate will require a large number of additional scenarios.9 Evolution of conﬁdence bands for 0. let us consider the standard deviation. 11. and could have stopped our calculations then. we compute each asset’s marginal standard deviation. giving the percent marginal standard deviation.4 Marginal risk measures 129 Chart 11. we could have obtained similar estimates and similar conﬁdence bands with only half the effort.
39 37.99 15.892 3.5 Standard deviation of value change Asset 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Credit rating AAA AA A BBB BB B CCC A BB A A A B B B B BBB BBB BBB AA Standalone Absolute ($) Percent 4.900 18.48 35.043 3.531 0. We see in general that for the higher rated assets. as in Chart 11.17 17.051 2. Points in the upper left of the chart represent assets which are risky in percent terms.01 The difference between the standalone and marginal risk for a given asset is an indication of the effect of diversiﬁcation.25 105.06 1.16 9. This is in line with our intuition that a much larger portfolio is required to diversify the effects of riskier credit instruments.06 63.01 389 0.21 167.69 120 0.10 represents points with the same absolute risk.079 6.905 0.84 255.251 12.16 Marginal Absolute ($) Percent 239 0.06 2.63 162. The curve in Chart 11.007 0.190 3.624 0.511 15.607 8.934 0.130 Chapter 11.56 40.068 0.97 5.39 380.01 693 0.775 0. while points which fall below have less.949 1.000 8.085 1.837 0.30 43.232 0. An interesting way to visualize these outputs is to plot the percent marginal standard deviations against the market value of each asset.27 89.827 4.53 2.523 1.12 270.10 4.16 1.207 1. but whose exposure sizes are small.37 99.916 0.25 16. points which fall above the curve have greater absolute risk. the percent risk multiplied by the market value) gives the absolute marginal risk.360 1.73 107.13 306.684 3.720 12. Note that the product of the two coordinates (that is.98 73.21 69.680 22.046 1. there is a greater reduction from the standalone to marginal risk than for the lower rated assets.18 25.77 28.314 15.141 7. while points in the lower right represent large exposures which have relatively small chances of undergoing credit losses.72 322.67 197.40 610.64 44.152 1.664 2.104 0.06 2.00 114 0.10. Portfolio example Table 11. CreditMetrics™—Technical Document .
11. the CCC rated issue.5% •• • • • • •• • • Asset 16 • • • 5 Asset 9 0. we may identify with the aid of the curve the ﬁve greatest contributors to portfolio risk. Asset 9 has a reasonably secure BB rating. With this. Finally. we conclude the chapter. Some of these “culprits” are obvious: Asset 7 is the CCC rated issue. The reader should now an understanding of the various descriptors of the future portfolio distribution which can be used to assess risk.0% 0 • • • Asset 18 • 15 10 Credit exposure ($mm) Based on the discussion above. For instance.Sec.5% • • Asset 7 Asset 15 5. as well as how the use of a risk measure should inﬂuence the decision on precisely which measure to use. and has a much larger likelihood of default. but has a 55% correlation with Asset 18. On the other hand. the other “culprits” seem to owe their riskiness as much to their correlation with other instruments as to their individual characteristics. but is a rather large exposure.0% 7.0% 2.10 Marginal risk versus current value for example portfolio Marginal standard deviation 10. the appearance of Asset 15 as the riskiest in absolute terms seems to be due as much to its 45% correlation with two other B issues as to its own B rating. whereas Asset 18 is BBB rated. but has a correlation of 35% with Asset 7.4 Marginal risk measures 131 Chart 11. and discuss what policy implications the assessment of credit risk might have. while Asset 16 is rated B. Part III: Applications . we step away from the technical. In the following chapter.
132 Chapter 11. Portfolio example CreditMetrics™—Technical Document .
• to measure and compare credit risks so that an institution can best apportion scarce risktaking resources by limiting overconcentrations. There are at least two features of risk which are worth reducing. 1 Researchers interested in valuation and pricing models may refer to the following: Das & Tufano [96]. Jarrow & Turnbull [95]. Madan & Unal [96]. is a policy issue. Hurley & Johnson [96]. we will make reference to Chart 12. and Sarig & Warga [89]. Fridson & Gao [96]. Neilsen & Ronn [96]. Note that we do not address the issue of credit pricing. For this discussion. Application of model outputs The measures of credit risk outlined in the preceding sections can have a variety of applications. or by allocating capital to functions which have proven to take risk most effectively. • reevaluate obligors having the highest percentage level of risk (the upper left corner of the chart) arguing that these are the most likely to contribute to portfolio losses.1. Moore & Roenfeldt [90].1 Prioritizing risk reduction actions The primary purpose of any risk management system is to direct actions. Although credit risk can be an important input into a credit pricing decision. These additional factors are nontrivial and so we have chosen to focus this current version on the already challenging task of risk estimation. The bottom line is that in order to optimize the return we receive for the risk we take. we will highlight just a few: • to set priorities for actions to reduce the portfolio risk. Merton [74]. which is exactly like Chart 11. Thus approaches include: • reevaluate obligors having the largest absolute size (the lower right corner of the chart) arguing that a single default among these would have the greatest impact.10. and this is the contribution of CreditMetrics. The objective of all of the above is to utilize risktaking capacity more efﬁciently. Foss [95].133 Chapter 12. and (ii) statistical risk level. Skinner [94]. Whether this is achieved by setting limits and insisting on being adequately compensated for risk. Tejima & Van Deventer [93]. and • to estimate economic capital required to support risktaking. but the tradeoff between them is judgmental: (i) absolute exposure size. but for a hypothetical portfolio with a very large number of exposures. . it is necessary to measure the risk we take.1 12. Eberhart. But there are many actions that may be taken towards addressing risk – so they must be prioritized. Shimko. Other research on historical credit price levels and relationships includes: Altman & Haldeman [92]. and Sorensen & Bollier [94]. we believe that there are signiﬁcant other determinants for pricing which are beyond the scope of CreditMetrics.
9 1 Absolute exposure size (by obligor) Like Chart 11.” whose large exposures were created when their credit ratings were better. It is this type of obligor which we advocate addressing ﬁrst. Unfortunately. but who now have much higher percentage risk due to recent downgrades. there is another issue to consider when considering which exposures should be addressed: whether the returns on the exposures in question adequately compensate their risk.1 does not completely describe the portfolio in question. the quality of a credit can change over time and a large exposure may have its credit rating downgraded (i.1 0. its point will move straight up in this chart).4 0. These are the parties which contribute the greatest volatility to the portfolio. The portfolio will then have a large exposure with also a relatively large absolute level of risk.Low impact •• • 8 •• • • • • • 7 •• • •• 6 ••• • •• • • • • • • •• • High risk and large size 5 4 3 2 1 0 •• • • • •• • • • • •• • • • ••• • • •• • ••• • • •••• • • • • • • •••• • • •••• •• • • • • •• • • •• •• •••• •• • • • • ••• • ••••• • •• • ••• • • ••• • ••••••• •• • • • • High absolute standard deviation .8 0.3 0. these are often “fallen angels.1 Risk versus size of exposures within a typical credit portfolio Marginal standard deviation % (by obligor) 10 • 9 • • • • • • Highest standard deviation % . Thus. as it does not address the issue of returns.2 0. Obligors with high percentage risk – and presumably high anticipated return – can be tolerated if they are small in size. we advocate the last one. it can be assumed that assets will be priced according to their risk on a standalone basis. Chart 12. In general.10.low percentage risk •••••• •• •• ••• •• • •• • •• • ••• • •• • • •• • • • • • •• ••• •• • • •• • •• ••• • • • ••• • • •• • • •• • • •• • • •••• • • • • •• ••• • • •• • • • • • • • • ••••••••• • ••••••• ••••• • • • • •••••• • ••• ••• • ••••• •• • • ••• • •• • • • •• •••• •• ••• •• ••• • •• • • • • • • ••• • ••• • ••• • ••• • • •• • • •• • • ••• •• • • • •••• •• •• • • •• • • • •• • •• •• • • •• • • • • • • • • • • • •• • • • •• • •• • • • • ••• •• • ••• ••• • • • • •• • • •••• ••••••••••••••••• •••••• •• •••••••• •••••••••• • •••• • • • • • • • ••• ••• • •••• ••• •• • ••• • ••• • • •• • •• • •• • • ••• •• •• •• ••• •• • • • • • •• • • •• • • •• •••• ••••• •• • • ••••• ••••• ••• •• • •• • • ••• ••• ••••• ••• •••• •••••••• • •••• • •• • • •• ••• •• ••• • • • • • •• • • •• • •••• • • •••• • •• •••• •• • ••• • • • •• ••••• • •• • • ••••••••• • • •• •• •• • ••••• •• •••••• ••• • •••• ••••••••••••• •• • •••••••••••• •• •• •••• • • • •• • •• • • • • • ••• •• • • • • • • • • • •• ••• •• •• • •• ••• •••• • • • •• • • • ••• • •• •• •• • •• • •• • • • • • •• • • • • • • 0 0. This is where the power of a portfolio analysis becomes evident. Application of model outputs • reevaluate obligors contributing the largest absolute amount of risk (points towards the upper right corner of the chart) arguing that these are the single largest contributors to portfolio risk.e. in a CAPM (capital asset pricing model) framework. setting as the highest priority to address those obligors which are both relatively high percentage risk and relatively large exposure. What this CreditMetrics™—Technical Document . however. Although all three approaches are perfectly valid.5 0. according to their correlation with a broad universe of assets. In practice.. or otherwise.6 0. this chart illustrates a risk versus size proﬁle for a credit portfolio. Large exposures are typically allowed only if they have relatively small percentage risk levels.134 Chapter 12.7 0. Chart 12.
because of the structure of the two portfolios. but also the identiﬁcation of how each asset makes its contribution.000 • • • Asset 18 • $15. When the risk of an asset is due largely to concentrations particular to the portfolio. credit rating. We see then not only the importance of evaluating the contribution of each asset to the risk of the portfolio.000 Market value Part III: Applications . and what policy to employ with regard to the limits. Chart 12.000.2 Possible risk limits for an example portfolio Marginal standard deviation 10% 9% 8% 7% 6% 5% 4% 3% 2% 1% 0% $0 • • •• • • • • •• Asset 7 Asset 15 • Asset 16 • • • Asset 9 $5. 12. However. and price. Consequently. Two managers identify a risky asset in their portfolios. the managers reduce their concentration. 12. and thus their risk. This might be the case if two managers are heavily concentrated in two different industries. the next step beyond using risk statistics for prioritization is to use them for limit setting. we discuss three aspects a user might consider with regard to using CreditMetrics for limit purposes: what type of limit to set. without reducing their expected proﬁts. but with two additional barriers included. the risk of both portfolios will be reduced without the expected return on either being affected. In this section. and yet yield the same returns in either case. which the reader might recognize as exactly the same as Chart 11. and are expected to yield equivalent returns. It turns out that the two assets are of the same maturity.2.000. as well as what type of policy to take with regard to the limits.10.2.Sec. which risk measure to use for the limits. what type of risk measure to use for limits. are management decisions. By swapping similar risky assets. an opportunity could well exist to restructure the portfolio in such a way as to reduce its risk without altering its proﬁtability.000. Of course.000 $10.1 Types of credit risk limits This section’s discussion will make reference to Chart 12. we can imagine the following situation. if the managers swap these assets.2 Credit risk limits 135 means is that a given asset may contribute differently to the risk of distinct portfolios.2 Credit risk limits In terms of policy rigor. as in the example above. 12.
If we measured risk in absolute terms. This would correspond to a limit like the vertical bar in Chart 12. • Set limits based on exposure size. say. above a given size. it is necessary next to choose the speciﬁc risk measure to be used. this would cap the total risk of the portfolio at a certain amount above the current risk. regardless of this counterparty’s credit standing. These statistics allow the user to examine an exposure with regard to its effect on the actual portfolio.2 Choice of risk measure Given a choice of what type of limit to implement.2. Application of model outputs We might consider each of the three possibilities mentioned in the previous section as candidates for credit risk limits. This would correspond to a limit like the curve in Chart 12.1 and 12. are certainly appropriate. and do not depend on the probability that the counterparty actually defaults.2. in other words.2. whether to use standard deviation. Such a limit would prevent the addition to the portfolio of any exposure which increased the portfolio risk by more than a given amount. it is perfectly intuitive to require that exposures which pose a greater chance of decreases in value due to credit be smaller. in that they reﬂect the amount we are willing to lose conditioned on a given counterparty’s defaulting. whether to use a marginal or standalone statistic. regardless of credit quality. since these have the greatest impact on the total portfolio risk. there are two choices to make: ﬁrst. The arguments for using marginal statistics have been made before. B or higher. that is.2. The limits proposed in this section are meant to supplement. Thus. limiting absolute risk is consistent with the natural tendencies of portfolio managers. in both cases. Essentially. a limit restricting the portfolio to contain only exposures rated. We see the natural tendency to structure portfolios in this way in both Charts 12. the risk proﬁles tend to align themselves with the curve rather than with either the vertical or horizontal line. this would correspond exactly to a limit on credit quality. or another statistic. this limit would be slightly different in that it would also restrict exposures that are more correlated to the portfolio. • Set limits based on absolute risk. since these contribute more to portfolio risk.136 Chapter 12. and second. Since we measure risk in marginal terms. Such limits may be thought of as conditional. it is most sensible to set limits in terms of absolute (rather than percent) risk. these conditional limits. takCreditMetrics™—Technical Document . setting limits based on absolute risk would take the qualitative intuition that currently drives decisions and make it quantitative. In the previous section. Such a limit would restrict the portfolio to have no exposures. We treat each in turn: • Set limits based on percentage risk. while allowing those with less chance of depreciating to be greater.2. Moreover. Limits based on the notion that there is a maximum amount of exposure we desire to a given counterparty. In effect. we argued that it is best to address exposures with the highest level of absolute risk ﬁrst. percentile level. but not replace. This would correspond to a limit like the horizontal line in Chart 12. 12. By the same token. It is worth mentioning here that the risk limits we have discussed are not meant to replace existing limits to individual names.
Part III: Applications . we would expect our losses to be $12mm. our measure is subject to the random errors inherent in Monte Carlo approaches. and particularly for prioritization. then the interpretation would be that in the worst 1 percent of all possible cases. this statistic is deﬁned as the expected loss given that the loss is more extreme than the 1st percentile level. the easiest statistic to compute is the standard deviation. if we say that given a loss of over $3mm occurs. Thus. it does not have quite as concrete an interpretation as a percentile level. and must resort to simulations. as a measure of credit risk. certain circumstances suggest the use of absolute risk measures for limits. As to what statistic to use. we might consider using the expected excession of a percentile level. we know that this is precisely the level below which we can expect losses only one percent of the time. We have also discussed the use of percentile levels at some length. however. suppose a portfolio contains a large percentage of a bond issue of a given name. it has a number of deﬁciencies. requires a simulation approach. it is important in this case to know the standalone riskiness of the position. Another statistic which is often mentioned for characterizing risk is average shortfall. This is a very reasonable characterization of risk. but like percentile level and average shortfall.Sec. Along the same lines. On the other hand. and discuss the applicability of each to credit risk limits. the standard deviation is a “twosided” measure. measuring the portfolio value’s likely ﬂuctuations to the upside or downside of the mean. There is a price for this precision. the position should be considered risky because of the liquidity implications of holding a large portion of the issue. What is most important is that the risk estimates give us an idea of the relative riskiness of the various exposures in our portfolio. this makes the standard deviation somewhat misleading. If this statistic were $12mm. or as the expected loss given that losses exceed a given level. we expect that loss to be $6mm. While this does give some intuition about a portfolio’s riskiness. The advantages of this statistic are that it is easy to deﬁne and has a very concrete meaning. since distributions of credit portfolios are mostly nonnormal. 12. As always. It is reasonable to claim that the standard deviation does this. For the 1st percentile level. we still do not have any notion of how likely a $6mm loss is. When choosing a risk statistic. However. For limits. First. Thus. as we cannot derive such a measure analytically. we describe four statistics below. This statistic is deﬁned as the expected loss given that a loss occurs. marginal statistics provide a better picture of the true concentration risk with respect to a given counterparty. Even if the name has a very low correlation with the remainder of the portfolio (meaning that the bond has low marginal risk). it is important to keep in mind its application.2 Credit risk limits 137 ing into accounts the effects of correlation and diversiﬁcation. it would be sensible to sacriﬁce the intuition we obtain from percentile levels or expected excessions if using the standard deviation provides us with signiﬁcant improvements in computational speed. Thus. In addition. for the purpose of prioritization or limit setting. there is no way to infer concrete information about the distribution from just the standard deviation. Since we are essentially concerned with only the downside. For instance. For instance. When we state the ﬁrst percentile level of a portfolio. it is not absolutely necessary that we be able to infer great amounts of information about the portfolio distribution from the risk statistics that we use. Thus.
Thus. but rather are seeking to understand the risk of the entire portfolio with regard to what this risk implies about the stability of our organization. The total risk of the portfolio might guide the limitsetting process. Excessions of this type are essentially uncontrollable.3 Economic capital assessment For the purposes of prioritization and limit setting. The common thread is that exposures which exceed the limits are permitted. additional authorization to increase exposure size.138 Chapter 12. For instance. an excession of the limit might require more indepth reporting. we examine a different application of credit risk measures. plus even an overall credit portfolio limit. We are no longer trying to compare different exposures and decide which contribute most to the riskiness of the portfolio. one might implement both types of limits – an informational limit at some low level of risk or exposure and a hard limit at a higher level. insurers. CreditMetrics™—Technical Document . There can be many levels of limits which we classify according to the severity of action taking in the case the limit is exceeded.2. the ﬁnancial sector limit should not be greater than the sum of limits to industries underneath it such as banks. or even supplemental covenant protection or collateral. this will not be the case. Application of model outputs 12. but trigger other actions which are not normally necessary. and each exposure we add to the portfolio satisﬁes the limits. It is not uncommon to set limits at different levels of aggregation since different levels of oversight may occur at higher and higher levels. or instrument type. industry. Alternatively. etc. geographical region. but subsequently exceed the limits due to a change in market rates or to a credit rating downgrade. the subjects of the ﬁrst two sections. that of assessing the capital which a ﬁrm puts at risk by holding a credit portfolio. The assumption for both types of limits above is that the limits are in place before the exposures. brokers. plus industry limits. although a portfolio manager might seek to reduce the risk in these cases by curtailing additional exposure.3 Policy issues The fundamental point of a limit is that it triggers action. or hedging with a credit derivative. which would preclude any further exposure to an individual name. one might set hard limits. However. For informational limits. And these limits might even be based on two different risk measures – a marginal measure at one level and an absolute measure at the other. reducing existing exposure. In this section. In practice. It should always be the case that a limit will be less than or equal to the sum of limits one level lower in the hierarchy. These exposures satisfy the risk limits when they are added to the portfolio. 12. there might be limits on individual names. we examined risk measures in order to evaluate and manage individual exposures. This will be true whether limits are set according to exposures (which can be aggregated by simply summing them) or according to risk (which can be aggregated only after accounting for diversiﬁcation). plus sector limits. for the aforementioned fallen angels. but it was the relative riskiness of individual exposures which most concerned us.
it is difﬁcult to consider an expected shortfall of $6mm as a usage of capital since we do not know how likely such a loss actually is. This ﬁts nicely with our discussion of capital above. then the interpretation would be 2 As quoted in Bernstein [96]. For limits we could argue that the standard deviation was an adequate statistic in that it could capture the relative risks of various instruments.3 Economic capital assessment 139 To consider risk in this way. 12. which might be thought of as risktaking capability. the distinct use of risk measures here make the decision different. a percentile level seems quite appropriate. we look at risk in terms of capital. For example. These issues should become clear as we consider the risk statistics below. since eventually the portfolio risk will surpass the “comfort level. To measure or assess the economic capital utilized by an asset portfolio. each additional exposure utilizes some of a scarce resource. Yet this statistic is practical to compute and for this reason alone may be the logical choice. Recall that if this statistic were $12mm at the ﬁrst percentile level. there was a ten percent chance for such a depreciation to occur in the next year as to cause organizationwide insolvency. If this level ever reaches the point at which such a loss will prevent us from meeting obligations. of what statistic to use to describe this distribution. If it is our desire to be 99% certain of meeting our ﬁnancial obligations in the next year. we were concerned with individual exposures and relative measures. the manager cannot add new exposures indiscriminately. This involves a choice. As an indicator of economic capital. however. we examine capital from a risk management informational view. but rather than considering the standard regulator or accounting view of capital. it is difﬁcult to argue that a standard deviation represents a good measure of capital since we are unable to attach a concrete interpretation to this statistic. we could deﬁne economic capital as the level of losses on our portfolio which we are 99% certain (or in the words of Jacob Bernoulli. then it is taking risk by holding assets that are volatile. The choice is in some ways similar to the choice of risk statistic for limits which we discussed in the previous section. we may consider average shortfall as a potential statistic. On the other hand. or as the economic capital which we are allocating to our asset portfolio. In this case. then we may think of the 1st percentile level as the risk we are taking. This risktaking capability is not unlimited. Using for example the 1st percentile level. Yet just as in the case of limits. we are interested in a portfolio measure and have more need for a more concrete meaning for our risk estimate. As with limits. The general idea is that if a ﬁrm’s liabilities are constant. the expected excession of a percentile level does seem worth consideration. For limits. then he would likely seek to decrease the risk of the asset portfolio. to the extent that the asset volatility could result in such a drop in asset value that the ﬁrm is unable to meet its liability obligations. as there is a level beyond which no manager would feel comfortable. then we will have surpassed the maximum amount of economic capital we are willing to utilize. for economic capital. as economic capital. Part III: Applications . For a portfolio with a more reasonable level of risk.Sec. if a manager found that given his asset portfolio. then. or alternately. “morally certain”2) that we will not experience in the next year.” Thus. however. we may utilize the distribution of future portfolio values which we describe elsewhere in this document.
assuring that if individual or industry level exposures are within the limits. the CreditMetrics methodology gives the user a variety of options to use for measuring economic capital which may in turn lead to further uses of CreditMetrics. How to choose which exposures to treat could then be guided by the discussions in Section 12. A third use is performance evaluation. An assessment of economic capital may guide the user to actions which will alter the characteristics of his portfolio. it will be necessary to take actions on one or more exposures. for offbalancesheet exposures. FRB of Chicago. leading to an incentive structure which encourages these managers to take on lower rated exposures in order to boost performance. CreditMetrics™—Technical Document . On the other hand.. on the transaction's exposure type (e. We brieﬂy touch on three applications of an economic capital measure: exposure reduction. with particular concern paid to the uneconomic incentives created by the regulatory regime and the inability of regulatory capital adequacy ratios to accurately portray actual bank risk levels. if the use of economic capital is too high. we would expect our losses to be $12mm. and performance evaluation. where each transaction's capital requirement depends on a broad categorization (rather than the credit quality) of the obligor. In response to these concerns. So like the percentile level above. The weaknesses of this riskbased structure – such as its onesizeﬁtsall risk weight for all corporate loans and its inability to distinguish diversiﬁed and undiversiﬁed portfolios – are increasingly apparent to regulators and market participants. Application of model outputs that in the worst 1 percent of all possible cases. Adding a measure of economic capital utilization allows for a more comprehensive measure of performance. drawn loans versus undrawn commitments). and thus seems a very appropriate measure. Examining performance in this way retains the incentive to seek high returns.4 Summary In summary. Board of Governors of the Federal Reserve System. limit setting. possibly by prohibiting additional exposure.g. For example. For a portfolio of positions not considered to be trading positions. one might wish to use the measure of economic capital in order to aid the limitsetting process. before the 32nd Annual Conference on Bank Structure and Competition. but penalizes for taking undue risks to obtain these returns. 1996.1. it can be seen which managers make the most efﬁcient use of the ﬁrm’s economic capital.140 Chapter 12. or else by reducing existing exposures by unwinding a position or hedging with a credit derivative. the BIS riskbased capital accord of 1988 requires capital that is a simple summation of the capital required on each of the portfolio's individual transactions. on whether the transaction's maturity is under oneyear or over one year.3 12. May 2. All of the above measures of economic capital differ fundamentally from the capital measures mandated for bank regulation by the Bank for International Settlements (BIS). The traditional practice has been to evaluate portfolio managers based on return. bank regulators are increasingly looking for insights in internal credit risk models that generate expected losses and a probability distribution of unexpected losses. this seems to coincide with our notion of economic capital. then the level of capital utilization will be at an acceptable level. when managers’ returns are paired with such a risk measure. 3 See Remarks by Alan Greenspan. and.
4 Summary 141 By examining rates of return on economic capital and setting targets for these returns. And just as it is possible to allocate any other type of capital.Sec. 12. but within transparent. proﬁts are maximized. Identifying portfolios or businesses that achieve higher returns on economic capital essentially tells a manager which areas are providing the most value to the ﬁrm. By focusing capital on the most efﬁcient parts of a ﬁrm or portfolio. responsible risk guidelines. or more risktaking ability. a manager or ﬁrm goes a step beyond the traditional practice of requiring one rate of return on its most creditworthy assets and a higher rate on more speculative ones. areas where the return on risk is higher may be allocated more economic capital. the new approach is to consider a hurdle rate of return on risk. Part III: Applications . which is more clear and more uniform than the traditional practice.
Application of model outputs CreditMetrics™—Technical Document .142 Chapter 12.
143 Appendices .
144 CreditMetrics™—Technical Document .
and indices in several different capacities. Appendices . Used to: (i) combine the uncertainty of spread and exposure risk and (ii) for the derivation of risk across mutually exclusive outcomes. A generalization of the methods presented in Chapter 9 to compute the standard deviation for a portfolio of arbitrary size. We have chosen to address each of them in detail here in an appendix so that we may give them the depth they deserve without cluttering the main body of this Technical Document. These appendices include: Appendix A: Analytic standard deviation calculation. Appendix I: Indices used for asset correlations. Appendix H: Model inputs. Techniques to assess the precision of portfolio statistics obtained through simulation. Used for both: the value variance of a position across Nstates and the covariance between positions across Nstates. Appendix D: Derivation of risk across mutually exclusive outcomes. Appendix B: Precision of simulationbased estimates. Used to link correlation between ﬁrms’ value to their default correlations. Describes the CreditMetrics data ﬁles and required inputs. Contains this information in tabular format. Appendix C: Derivation of the product of N random variables. Contains this information in tabular format. Used as alternative method to estimate default correlations which corroborates our equity correlation approach. Appendix E: Derivation of the correlation of two binomials. Appendix F: Inferring default correlations from default volatilities. data. Appendix G: International bankruptcy code summary.145 Appendices In CreditMetrics we use certain general statistical formulas.
146 CreditMetrics™—Technical Document .
.. Alternatively. [A. This makes the computation of the portfolio standard deviation straightforward. express the portfolio standard deviation in terms of the standard deviations of subportfolios containing two assets: [A. σ ( V 2 ) . we then identify all pairs of assets among the n assets in the portfolio1 and compute the variances for each of these pairs using the methods in Chapter 3. Analytic standard deviation calculation In Chapter 9. we present this generalization in detail.. and stated that the generalization of this calculation to a portfolio of arbitrary size was straightforward..1] 2 σp = ∑ σ (V ) + 2 ⋅ ∑ ∑ 2 i i=1 i = 1j = i+1 n n–1 n COV ( V i.µ2. The value of the portfolio at the end of the forecast horizon is just V1+V2+. . 2 2 2 and using this fact. [A. The calculation of these individual means and variances is detailed in Chapter 2. let these values’ means be µ1. σ ( V n ) .. V2..˙ V j ) . we may relate the covariance terms to the variances of pairs of assets.+µn. We begin by computing the variances of each individual asset. To compute the portfolio standard deviation (σp).3] n–1 2 σp = ∑ ∑ n σ (Vi + V j) – (n – 2) ⋅ 2 ∑ σ ( V ).. and the mean value is µp=µ1+µ2+. we apply Eq..... Denote the value of these assets at the end of the horizon by V1... we may use the standard formula: [A..147 Appendix A. CreditMetrics™—Technical Document .µn and their variances be 2 2 2 σ ( V 1 ) . V j ) + σ ( V j ). we see that the portfolio standard deviation depends only on the variances for pairs of assets and the variances of individual assets.+Vn. Vn. Consider a portfolio of n assets. In this appendix. 2 i i=1 n i = 1j = i+1 As in Chapter 9.3]. we presented the calculation of the standard deviation for an example three asset portfolio.2] σ ( V i + V j ) = σ ( V i ) + 2 ⋅ COV ( V i. 1 There will be n ⋅ ( n – 1 ) ⁄ 2 pairs. ﬁnally..
148 CreditMetrics™—Technical Document .
and examine how much ﬂuctuation there is in these estimates.V(2).. In order to extrapolate to an estimate for the standard error of σ20000. We devote one subsection each to the treatment of the sample mean.. This procedure is commonly referred to as “jackkniﬁng.. then we might divide these scenarios into ﬁfty separate groups of 400.V[3]. Precision of simulationbased estimates In Chapter 10. The simplest approach here is to break the full set of scenarios into several subsets.000 portfolio scenarios. for example V[2] is the second smallest value). and thus discover how conﬁdent we may be of the risk estimates we compute.. Motivated by the fact that the estimates we compute are sums over a large number 2 Since the probability that a normally distributed random variable falls within one standard deviation of its mean is 68%. we will use V(1). we discuss how we may quantify the sizes of these errors.. our approach to assessing precision was the same..65 ⋅ σ n ⁄ n ) and µ n + 1. sample standard deviation.σ(50). obtaining ﬁfty different estimates σ(1).149 Appendix B. we assume that the same scaling holds as with the sample mean. Note that these bands will tighten as n increases. and take s ⁄ 50 . we presented a methodology to compute portfolio statistics using Monte Carlo simulation and mentioned that statistics which are estimated in this way are subject to random errors. B. let µn denote the sample mean and σn the sample standard deviation of the ﬁrst n scenarios. which we denote by s.2 Sample standard deviation Our conﬁdence in the estimate σn is more difﬁcult to quantify since the distribution of the estimate is less well approximated by a normal distribution.65 ⋅ s ⁄ 50 ) and µ 2000 + ( 1.V[2]. but in practice suggest that the user experiment with various numbers in order to get a feel for the sensitivity of the conﬁdence estimates to this choice..1 Sample mean Quantifying the error about our estimate of the mean portfolio value is straightforward. B. and the standard deviation of the estimate is much harder to estimate. σ(2).65 ⋅ s ⁄ 50 ) 3.. compute the sample standard deviation for each subset. For example..V(3). For large n. Throughout this section. we may say that we are 68%2 conﬁdent that the true mean portfolio value lies between µ n – σ n ⁄ n and µ n + σ n ⁄ n and 90% conﬁdent the true mean lies between µ n – ( 1. The sample standard deviation of these estimates. In this appendix.V(N) to indicate the portfolio values across scenarios and V[1]. This methodology is somewhat sensitive to the choice of how many separate groups to divide the sample into. We could then compute the sample standard deviation within each group. 3 Appendices .V[N] to indicate the same values sorted into ascending order (so that.. is then an estimate for the standard error of σ400. if we have generated 20. Thus. after generating n scenarios. Then we can say that we are approximately 90% conﬁdent that the true value of our portfolio standard deviation lies between µ 2000 – ( 1. and sample percentile levels. µn will be approximately normally distributed with standard deviation σ n ⁄ n .. We choose 50 here.” For the sample mean and standard deviation.65 ⋅ σ n ⁄ n .. Further.
1 and 50+6. This implies that θ5 is at least as large as the 43rd smallest of our portfolio values. and thus are much more concerned with the precision of our estimates. The rest of the analysis then focused on computing the standard errors for the estimates.3 Sample percentile levels As an example. Moreover. say we are trying to estimate the 5th percentile level.) On the other hand. and thus [B. At this point we have characterized N5.65 ⋅ 6. we have argued that the event [B. further.9. B. there is a 90% chance that N5 falls between 50 – 1.9 = 38. In other words. the assessment of precision for estimates of these two statistics is somewhat redundant. since we can gain a large amount of information from its distribution.2 .9=43. we treat estimates of percentile levels. We assert that it is not necessary to know N5 exactly. using 1000 scenarios. (Recall that in our notation. and let N5 be the number of these scenarios which fall below θ5. however. For this many trials. it is reasonable to approximate the distribution of N5 by the normal. this scenario is denoted by V[43]. then it must be true that θ5 is no larger than the 57th smallest of the portfolio values (that is. we have no way of computing percentile levels directly. in some sense. Thus.1] 43 ≤ N 5 ≤ 57 is exactly the same as the event [B. Further. In the next section. and state CreditMetrics™—Technical Document . we estimate the 5th percentile portfolio value by the 50th smallest scenario. Now consider 1000 independent scenarios.65 ⋅ 6. Now since these two events are the same. Clearly. as it is possible to obtain exact values in both cases. then at least 43 of our scenarios are less than θ5. while the standard deviation is 1000 ⋅ 5% ⋅ ( 100% – 5% ) = 6. Note that N5 follows the binomial distribution.3] Pr { V [ 45 ] < θ5 < V [ 57 ] } = Pr { 43 ≤ N 5 ≤ 57 } = 68% and so we have a conﬁdence bound for our estimate of θ5. we estimate that there is a 68% chance that N5 will be between 506. Estimates are not just sums over the scenarios. they must have the same probability. This may not seem particularly useful.150 Appendix B. since N5 is not actually observable. To recap. if N5 is less than or equal to 57.2] V [ 43 ] < θ5 < V [ 57 ] . since we do not actually know the level θ5 (this is what we are trying to estimate). and let θ5 be the true value of this level. the expected value of N5 is 1000 ⋅ 5% = 50 . Observe that if N5 is greater than or equal to 43.9=56. Each scenario which we generate then (by deﬁnition) has a 5% chance of producing a portfolio value less than θ5. Precision of simulationbased estimates of independent trials. we have no way of knowing how many of our scenarios fell below θ5.6 and 50 + 1. for which neither of these points applies.9 = 61. and a slightly higher chance that N5 will be between 43 and 57. and thus we cannot expect the distributions of the estimates to be normal. V[57]).9 . Thus. we approximated the distributions of the estimates as normal.
then α=1. s = and N ⋅ p ⋅ (1 – p) [B. this will be the case as long as the expected number of scenarios falling below the desired percentile (that is. Precision of simulationbased estimates 151 that we are 68% conﬁdent that the true percentile lies somewhere between the 43rd and 57th smallest scenarios.65. (That is. while if u is not a whole number.4] upper bound: u = N ⋅ p + α ⋅ N ⋅ p ⋅ ( 1 – p ) where α depends on the level of conﬁdence which we desire. see DeGroot [86]. we round them downwards. We characterize this number via the following: lower bound: mean: l = N ⋅ p – α ⋅ N ⋅ p ⋅ (1 – p) m = N ⋅ p.Appendix B. we may take the same approach as in this section. and then proceed to infer conﬁdence intervals on the estimated percentile. In general. if we desire 90%. p. if we desire 68% conﬁdence. in that we will obtain conﬁdence bands on the number of scenarios falling below the threshold. . we round upwards. but characterize the distribution precisely rather than using the approximation. Note that the only assumption we make in this analysis is that the binomial distribution is well approximated by the normal. we ﬁrst consider the number of scenarios that fall below the true value of this percentile. In general. N·p) is at least 20 or so. then α=1. We then estimate our percentile by V[m] and state with our desired level of conﬁdence that the true percentile lies between V[l] and V[u].) If either l or m are not whole numbers. The result will be similar. In cases where this approximation is not accurate. etc. For further discussion of these methods. 563. if we wish to estimate the pth percentile using N scenarios.
152 CreditMetrics™—Technical Document .
the result is: Appendices .153 Appendix C. are independent and standardized but can otherwise have any desired shape: normal.6] ∏ i N Φ i = ∏µ i N i where all Φ i ∼ µ i + σ i ⋅ Z i and all Z i are standardized (0.3] X ⋅ Y = µxµy + µxσy ZY + µyσx Z X + σx Z X σy ZY Since the expected value of Z is zero. binomial. [C. highly skewed. the expectation of this product is simply the product of its expectations: E [C. 2 – 1 . we can we can extend the volatility estimation for the product of arbitrarily many independent events.1] X ∼ µ x + σ x ⋅ Z x Y ∼ µ y + σ y ⋅ Z y ( where ∼ denotes distributed as ) where all distributions. Z. we will multiply out x · y. σX ⋅ Y = (µx µy + µx σy + µy σx + σx σy ) – (µx µy ) 2 2 2 2 2 2 2 2 2 2 2 [C. First.4] E( X ⋅ Y ) 2 = µx µy 2 2 (Since: E(Z) = 0) + 2 2 µx σy E( X ⋅ Y ) = 2 2 2 µx µy + 2 2 µy σx + 2 2 σx σy (Since: E(Z) = 1 ) 2 Now σX·Y is only a matter of algebra.1) N The variance of the product of N distributions will in general have.5] 2 = µx σy + µy σx + σx σy σX ⋅ Y = 2 2 2 2 2 2 2 2 2 2 2 2 µx σy + µy σx + σx σy By induction. the E( )’s simplify greatly. For the case of the product of three distributions. terms. E( X ⋅ Y ) = µxµy 2 [C. Let X and Y be any independent and uncorrelated distributions deﬁned as follows: [C.2] σX ⋅ Y = E( X ⋅ Y ) – E( X ⋅ Y ) 2 2 2 2 (Textbook formula) First. Derivation of the product of N random variables First we examine in detail the volatility of the product of two random variables. [C. etc.
i 1 k CreditMetrics™—Technical Document .154 Appendix C. 3. … j 1 ≤ j < … < j ≤ N . Derivation of the product of N random variables [C. j and m denote sets whose elements comprise the product sums: VAR [C. k ≤ N } . N } ˙ and s ( N . In this notation. 2. …. the pattern continues and can be denoted as follows for N distributions. k ) ∑ ∏ k N σ2 ⋅ j m = S( N ) – j ∏ 2 µ m – ∏µ i N 2 i where the sets S ( N ) = { 1. k ) = { j .8] ∏ i N Φ i = k = 1 J ∈ s ( N.7] + µ2 µ2 σ2 + µ2 σ2 σ2 x y z x y z VAR ( Φ X ⋅ Φ Y ⋅ Φ Z ) = + µ 2 σ 2 µ 2 + σ 2 µ 2 σ 2 + σ 2 σ 2 σ 2 x y z x y z x y z 2 2 2 2 2 2 + σ x µ y µz + σ x σ y µz In general.
If the two are the same distribution. For completeness. Deﬁnitions: 1 = p 1 + p 2 + p ω and Φ T ( x ) = p 1 Φ 1 ( x ) + p 2 Φ 2 ( x ) + p ω Φ ω ( x ). Derivation of risk across mutually exclusive outcomes Imagine that there were two alternative outcomes (subscripts 1 and 2) that might occur in the event of default with probabilities of p1 and p2 which sum to the total probability of default. . then the correlation is simply 1. [D.0. µi and σi. Each of these three cases has some distribution of losses denoted.0 Substitution made above. The problem of multiplying a random variable by itself was addressed in the prior appendix note (see Appendix C). with statistics.2] Variance of Total Loss 2 σT = = ∫ ( x – µ ) Φ ( x ) dx = ( x – 2xµ + µ ) ( p Φ ( x ) + p Φ ( x ) + p Φ ( x ) ) d x ∫ p ∫ x Φ ( x) + p ∫ x Φ ( x) + p ∫ x Φ ( x) 2 T T 2 T 2 T 1 1 2 2 ω ω 2 2 1 1 2 2 ω 2 ω These simplify Note that this equals µ T see above + Note that this sums to 1 p1 ( µ2 + σ2 ) + p2 ( µ2 + σ2 ) + pω ( µ2 + σ2 ) 1 1 2 2 ω ω 2 = – 2µ T 2 + µT = p1 ( µ1 + σ1 ) + p2 ( µ2 + σ2 ) + pω ( µω + σω ) – µT 2 2 2 2 2 2 2 The above derivation requires a substitution for an integral that merits further discussion.1] Expected Total Loss ∫ xΦ ( x ) dx = x( p Φ ( x) + p Φ ( x) + p ∫ T 1 1 2 2 ω Φω ( x ) ) dx = p1 µ1 + p2 µ2 + pω µω [D.155 Appendix D. µT = [D. Φi(x).3] Mean of Product of Two Random Variables µ ( i ⋅ j ) = µ i µ j + ρσ i σ j = = 2 µi + 2 2 σi i ∫ x Φ ( x ) dx Appendices – 2µ T ( p1 µ1 + p2 µ2 + pω µϖ ) µT ( p1 + p2 + pω ) 2 See prior appendix note. Since i = j and ρ = 1. subscript ω is the case of no default.
156 Appendix D. Derivation of risk across mutually exclusive outcomes For completeness. But these are both zero since there will be no losses in the case of no default. Thus the overall total mean and standard deviation of losses in this process simpliﬁes as follows: [D. we have included terms describing the losses in the case of no default: µω and σω.4] σT = ∑ i=1 S pi ( µi + σi ) – µT 2 2 2 where µ T = ∑pµ i=1 S i i CreditMetrics™—Technical Document .
(that is µx times µy). there will be exactly four possible outcomes. We will label this joint probability as α. y = n ∑ W ( x – µ )( y – µ ) i i x i y i=1 The expected probabilities. cov x. Derivation of the correlation of two binomials The traditional textbook formula for covariance is shown below. are termed µx and µy respectively. p’s. see below. Appendices . As shown below. but here the probability weights Wi will equal the likelihood of each possible outcome. but we will leave them as variables to allow for any degree of possible correlation. otherwise it is negative. If the joint default probability. defaults will have value 1 and nondefaults will have value 0. α. yields an intuitive result for our covariance. of the two binomials. x and y. The probability weights Wi are easily calculated for the case of independence. [E. Normally all the n observations would be equally weighted (1/n).157 Appendix E. y = + W 2(0 – µx )(1 – µy) + W 3(1 – µx )(0 – µy) + W 4(0 – µx )(0 – µy) The difﬁcult problem in deﬁning the probability weights W’s is knowing the correlated joint probability of default (cell #1 above). then the covariance is positive. is greater than the independent probability.1] Obligor Y Default 1: X& Y default No Default Default Obligor X No Default 2: Only Y defaults 3: Only X defaults 1: X& Y default 2: Only Y defaults 4: Neither defaults 3: Only X defaults 4: Neither defaults W 1(1 – µx )(1 – µy) cov x. For the joint occurrence of two binomials. Multiplying and simplifying the resulting formula. We can simply list them explicitly.
this correlation ρxy is the resulting correlation of the joint binomials4. α.4] ( max ( 0. This formula for ρxy implies that there are bounds on ρxy since α is at least max(0. µ y ) – µ x µ y ) ≤ ρ x. y σ x σ y cov x. the above deﬁnition of α and ρ is identical the formula for the mean of the product of two correlated random variables as shown above (see Appendix A). The σ’s here are the usual binomial standard deviations. y and = (α – µxµy) ⁄ σxσy Interestingly. y = W 1(1 – µx )(1 – µy) +W 2 ( 0 – µ x ) ( 1 – µ y ) +W ( 1 – µ ) ( 0 – µ ) 3 x y +W 4 ( 0 – µ x ) ( 0 – µ y ) [ α ](1 – µx )(1 – µy) +[µ – α](0 – µ )(1 – µ ) y x y = +[µ – α](1 – µ )(0 – µ ) x x y + [ 1 – µ x – µ y + α ] ( 0 – µ x ) ( 0 – µ y ) α – αµ y – αµ x + αµ x µ y 2 – µ µ + µ µ + αµ – αµ µ x x y x y x y = 2 – µ µ + µ µ + αµ – αµ µ x y x y y x y 2 2 +µ x µ y – µ x µ y – µ x µ y + αµ x µ y = α – µxµy Now that we have derived the covariance as a function of the joint default probability. Importantly. µ x + µ y – 1 ) – µ x µ y ) ( min ( µ x. y σ x σ y α = µ x µ u + ρ x. µ ( 1 – µ ) .2] cov x. CreditMetrics™—Technical Document .158 Appendix E. Again. y = ρ x. y σ x σ y ρ x. we can redeﬁne α in terms of the correlation of our two binomials. µy).3] thus so α – µ x µ y = ρ x. see Lucas [95a]. Thus: [E. µx+µy1) and at most min (µx. It does not represent some underlying ﬁrmasset correlation that (via a bivariate normal assumption) might lead to correlated binomials. y ≤ σ σ σ σ x y x y 4 Other researchers have used this same binomial correlation. we can start with a textbook formula for the covariance: [E. Derivation of the correlation of two binomials [E.
159 Appendix F. within a single credit rating category). and deﬁne this as follows [F. D = ∑X i N i .1] 1 if company i defaults Xi = 0 otherwise µ CrRt σ( Xi) = 1 = µ ( X i ) = N ∑X i N i µ CrRt ( 1 – µ CrRt ) Let D represent the number of defaults. let X 1 be a random variable which is either 1 or 0 according to each ﬁrm’s default event realization with mean default rate. deﬁned as follows: [F. µ ( X 1 ) . Inferring default correlations from default volatilities For N ﬁrms in a grouping with identical default rate (i.. thus: Appendices . 2 2 2 Across many ﬁrms we can observe the volatility of defaults. σ CrRt = VAR ( D ⁄ N ) . we are interested in the average correlation. So the variance of D is as follows: VAR ( D ) = ∑∑ρ ∑∑ρ ∑∑ρ i j i N j N i N j N N N ij σ ( X i )σ ( X j ) = ij σ ( X i ) 2 Since all i and j have the same default rate. σ ( X 1 ) .2] = ij ( µ CrRt – µ CrRt ) = ( µ CrRt – µ CrRt ) N + 2 ∑∑ρ i j≤i N N ij Rather than each ρ ij . 2 [F.3] ρ CrRt = ∑∑ρ i j≤i N N ij ⁄ (N – N) 2 and so we can now deﬁne [F. and binomial default standard deviation.e.4] VAR ( D ) = ( µ CrRt – µ CrRt ) [ N + ( N – N )ρ CrRt ] . ρ CrRt .
3886% = N–1 8.= 1.42% 2 Ba = . 500 – 1 2 2 σ CrRt 1. Inferring default correlations from default volatilities [F.6] ρ CrRt The estimate of 8. see Carty & Lieberman [96a].160 Appendix F.42% Ba – 1.∴ ρ CrRt N σ CrRt N .42%).= .4% Ba 8.5000 ﬁrmyears above stems from Moody’s reporting of 120 ﬁrms being rated Ba one calendar year prior to default (8. 500 . CreditMetrics™—Technical Document .4% Ba = .= 1. = 2 N N 1 + ( N – 1 )ρ CrRt 2 = ( µ CrRt – µ CrRt ) ⋅ . – 1 2 µ CrRt – µ CrRt 1.500 ≅ 120/1. – 1 2 µ CrRt – µ CrRt = N–1 2 This can be applied with good result in a simpliﬁed form if N is “large”: 2 σ CrRt 1.4002% 2 2 µ CrRt – µ CrRt 1.42% Ba – 1. – 1 N .42% Ba 2 [F.5] VAR ( D ) D 2 σ CrRt = VAR .
The list in order of increasing eligibility. this will affect the likely recovery rate distributions.Negotiated Settlement dicaire) (Reglement Amiable) where a court appointed conciliator attempts a settlement with creditors and Judicial Arrangement (Redressement Judiciare). and no stay in a voluntary arrangement until a proposal is approved. Only debtors ﬁle. (Kaisha Koseiho). Compulsory winding up. Manage. Repossession even close to bankruptcy ﬁling is permitted. A creditor with a ﬁxed or ﬂoating charge can appoint an administrative receiver to realize the security and pay the creditor. International bankruptcy code summary The practical result of the seniority standing of debt will vary across countries according to local bankruptcy law. can be subject on the petition that is ﬁled.Third party is appointed exporate Arrangement (Kaisha cept in composition and corSeiri) and Reorganization porate arrangement.for only by debtor. Germany Liquidation (KonkursordComposition (Vergleich or Receiver appointed to mannung) can be requested by Zwangvergleich) can be ﬁled age ﬁrm. Trustee may be appointed to oversee management in some reorganizations at the discretion of the court. Secured creditors can recover their claims even after a bankruptcy ﬁling. and untary winding up. and White [93].Stay on all creditors in dation. Lo Pucki & Triantis [94].attempts to collect debt priority in any attempts to coluntary (creditors ﬁle). Appendices .Stay on all creditors in zations while trustee is apreorganization.(management ﬁles) or invol.161 Appendix G.Debtor is removed from contive Receivership (usually trol except in members’ volends in sale of business).1 Summary of international bankruptcy codes Country United States Japan Forms of Liquidation Forms of Reorganization Management Control in Bankruptcy Automatic Stay Rights of Secured Creditors Chapter 7: Can be voluntary Chapter 11: Can be voluntary Trustee appointed in Chapter Automatic stay on any Secured creditors get highest (management ﬁles) or invol. Secured creditors stayed in bankruptcy. Only unsecured creditors are stayed. Debtor remains in judicial arrangement. pointed for liquidation. Canada Liquidation proceedings much like Chapter 7 in the United States Firms can ﬁle for automatic stay under the Companies Creditors Arrangement Act or the Bankruptcy and Insolvency Act. Italy Bankruptcy (Fallimento) United Members’ voluntary windKingdom ing up. but stayed after ﬁling. Debtor is removed from con. Creditors’ voluntary winding up. Cor.7.Stay on all creditors. on unsecured only in liquidation. No stay for secured creditors. Composition (Wagiho). The latter is less costly and a broader set of ﬁrms are eligible to ﬁle. control otherwise but submits to court appointed administrator’s decisions in a judicial arrangement. Stay on all creditors in administration. Secured creditor may prevent administration order by appointing his own receiver. Secured creditors have to give 10 days notice to debtor of intent to repossess collateral. Of course. Administra. ment required to ﬁle as soon as it learns it is insolvent. Preventive Composition (Concordato Preventino) Debtor loses control in liqui. untary (creditors ﬁle). trol over the ﬁrm. However. Table G. Secured creditors may lose status if court determines the security is necessary for continuation of the business. All creditors are stayed except in court supervised liquidation and composition where only unsecured creditors are stayed. or if the securing asset is sold as part of settlement. Administration. once ﬁling takes place. Voluntary Arrangement. creditors or debtor. trol in Chapter 11. Secured creditors follow administrative claims in priority. Secured Creditors have highest priority and greater voting rights in renegotiation. Management stays in con. lect payment are also stayed unless court or trustee approves Court Supervised Liquidation (Hasan) and Special Liquidation (Tokubetsu Seisan). France Liquidation (Liquidation Ju. through composition allowed only if enough value exists to pay secured creditors in full and 40% of unsecured creditor claims. The following summary table is reproduced from Rajan & Zingales [95] – who in turn reference Keiser [94]. Firm is in control in reorgani. Major differences will apply to secured versus unsecured debt.
International bankruptcy code summary CreditMetrics™—Technical Document .162 Appendix G.
1 Common CreditMetrics data format characteristics In general.2 Country/industry index volatilities and correlations This ﬁle is named indxvcor.0257 0. Every CreditMetrics data ﬁle begins with a header.0000 0.5528 0.0000 0.6377 ASX Banks & Finance Index (. and carriage returns/line feeds as row delimiters. Here.4436 ASX Media Index (.CIAU) 0. we brieﬂy list what is provided and the format in which it is available.0 Date 02/15/1997 DataType CountryIndustryVolCorrs IndexName Volatility MSCI Australia Index (. the cell should contain the keyword NULL. The data represent the weekly volatilities and correlations discussed in Chapter 8.com/ is a data set of all the elements described in this technical document and necessary to implement the CreditMetrics methodology. If the value is unavailable or not applicable.3525 Appendices . H.7343 0. • yield curves.ABII) 0.6911 0. CreditMetrics data ﬁles are text (ASCII) ﬁles which use tab characters (ASCII code 9) as column delimiters.0 02/15/1997 CountryIndustryVolCorrs The header is followed by a row of column headers. H. for example: CDFVersion Date DataType v1.4360 1. • spread curves. CreditMetrics data ﬁles include: • country/industry index volatilities and correlations.4580 0.0219 0. Cells in the data rows must contain data.AMEI) 0.163 Appendix H.6840 0.0171 1. Model inputs Available for free download from the Internet http://jpmorgan.6911 0.cdf. and • transition matrices.4360 0.0000 0. followed by the data. CDFVersion v1.6840 1.
cdf.019365 0.01118 0. GBP. Morgan.015811 0.P.05707 H. they will have data for a one year risk horizon.0 02/15/1997 YieldCurves CompoundingFrequency 1 1 H.5 Transition matrices This ﬁle is named trnsprb. CHF. Allowable currencies are the standard threeletter ISO currency codes (e.3 Yield curves Maturity 1.022361 H. A yield curve is deﬁned by currency .0 Spread 0.007071 0.0 02/15/1997 SpreadCurves Rating Currency Aaa CHF Aaa CHF Aaa CHF Aaa CHF Aaa CHF Aaa CHF AssetType BOND BOND BOND BOND BOND BOND CompoundingFrequency 1 1 1 1 1 1 Maturity 5. Their contact number is (1800) 828 .g. Allowable currencies are the standard 3letter ISO currency codes (e. However. DEM.055 0. Lehman Brothers. A team of evaluators reviews the contributed information on a daily bvasis to ensure accuracy and consistency. CS First Boston.cdf. rating. A spread curve is deﬁned by a combination of rating system..P. DEM. and MDI.0 10. S&P major rating transition matrix.0 2. COMMITMENT. and a yield curve (a yield curve being deﬁned as a combination of currency and asset type). Allowable asset types are BOND. This ﬁle is named sprdcrv. RECEIVABLE.00866 0. Morgan Stanley. This contains transition probabilities for both Moody’s major and modiﬁed ratings. Initial data is available only for USD and BOND CDFVersion Date DataType RatingSystem Moody8 Moody8 Moody8 Moody8 Moody8 Moody8 v1. Bridge credit spread data is derived through a compilation of information provided by major dealers including Citibank. Model inputs This ﬁle is named yldcrv. JPY. Goldman Sachs. Salomon Brothers and J. LOAN. CreditMetrics™—Technical Document . GBP. USD).0 YieldToMaturity 0.0 3. the format supports other horizons. JPY. USD).8010. and J.cdf.4 Spread curves Bridge will be the initial data provider for credit spreads.g. Morgan derived matrices estimating longterm ratings behavior. CDFVersion Date DataType Currency CHF CHF v1.0 15.164 Appendix H.0 2. CHF. Liberty Brokerage.0 20. Initially.
Table H.880784 0. Each rating has an agency (Moody's.Appendix H.2 Description Must be unique. etcetera) Stock price times number of shares outstanding Proportion of sales assigned to speciﬁed countries and industries.6 Data Input Requirements to the Software Implementation of CreditMetrics Required inputs for each issuer Data Type Issuer name Credit Rating/Agency Market Capitalization Country & Industry Issuerspeciﬁc risk Table H.0 02/15/97 TransitionProbabilities FromRank ToRank FromRating 0 0 Aaa 0 1 Aaa 0 2 Aaa ToRating Aaa Aa1 Aa2 HorizonInMonths 12 12 12 Probability 0. CreditMetrics only utilizes the numerical FromRating and ToRating columns.029015 H. Volatility of issuer asset returns not explained by industry/country group(s). Long term rating that applies to the issuer's senior unsecured debt regardless of the particular seniority class(es) listed as its exposure. Model inputs 165 The FromRating and ToRating columns of descriptive rating labels are included for readability.050303 0. S&P.1 Required inputs for each exposure type Property Issuer Name Portfolio Currency Asset type Par value Maturity Seniority class Recovery rate Recovery rate std Fixed or ﬂoating Coupon or spread Coupon frequency Current line Current drawdown Expected drawdown Duration Expected exposure Average exposure Forward value Bond x x x x x x x x x x x x Loan x x x x x x x x x x x Commitment x x x x x x x x x x x x x x x MDI x x x x Receivable x x x x x x x x x x x x x . CDFVersion Date DataType RatingSystem Moody18 Moody18 Moody18 v1.
166 Appendix I. Food & Health Index Toronto SE Insurance Index Toronto SE Food Stores Index Toronto SE Electrical & Electronics Index Toronto SE Metals Mines Index Toronto SE Integrated Oils Index Toronto SE Biotechnology & Pharmaceuticals Index Toronto SE Publishing & Printing Index Toronto SE Transportation Index MSCI Germany Index CDAX Automobiles Index CDAX Investment Company Index CDAX Chemicals Index CDAX Construction Index CDAX Insurance Index CDAX Machinery Index CDAX Paper Index CDAX Textiles Index CDAX Transport Index CDAX Utilities Index MSCI Greece Index Athens SE Banks Index Athens SE Insurance Index MSCI Finland Index Helsinki SE Banks & Finance Index Helsinki SE Metal Index Helsinki SE Forest & Wood Index Helsinki SE Insurance & Investment Index Austria Belgium Canada Germany Greece Finland Appendices . Australia Indices used for asset correlations Asset Category General Banking and ﬁnance Broadcasting and media Construction and building materials Chemicals Energy Food Insurance Paper and forest products Transportation General General General Automobiles Banking and ﬁnance Broadcasting and media Construction and building materials Chemicals Hotels Insurance Food Electronics Metals mining Energy Health care and pharmaceuticals Publishing Transportation General Automobiles Banking and ﬁnance Chemicals Construction and building materials Insurance Machinery Paper and forest products Textiles Transportation Utilities General Banking and ﬁnance Insurance General Banking and ﬁnance Metals mining Paper and forest products Insurance Index MSCI Australia Index ASX Banks & Finance Index ASX Media Index ASX Building Materials Index ASX Chemicals Index ASX Energy Index ASX Food & Household Goods Index ASX Insurance Index ASX Paper & Packaging Index ASX Transport Index MSCI Austria Index MSCI Belgium Index MSCI Canada Index Toronto SE Automobiles & Parts Index Toronto SE Financial Services Index Toronto SE Broadcasting Index Toronto SE Cement & Concrete Index Toronto SE Chemicals Index Toronto SE Lodging.
Indices used for asset correlations 167 France Asset Category General Automobiles Banking and ﬁnance Construction and building materials Energy Food General Banking and ﬁnance Utilities General General Chemicals Banking and ﬁnance Food Paper and forest products Metals mining General Banking and ﬁnance Broadcasting and media Construction and building materials Chemicals Electronics Food Insurance Machinery Metals mining Health care and pharmaceuticals Paper and forest products Energy Oil and gas .reﬁning and marketing Textiles Transportation General Banking and ﬁnance Construction and building materials Chemicals Food Insurance Machinery Metals mining Paper and forest products Textiles Transportation General Banking and ﬁnance Metals mining Index MSCI France Index SBF Automotive Index SBF Finance Index SBF Construction Index SBF Energy Index SBF Food Index MSCI Hong Kong Index Hang Seng Finance Index Hang Seng Utilities Index MSCI Indonesia Index MSCI Italy Index Milan SE Chemical Current Index Milan SE Financial Current Index Milan SE Food & Groceries Current Index Milan SE Paper & Print Current Index Milan SE Mine & Metal Current Index MSCI Japan Index Topix Banking Index Topix Communications Index Topix Construction Index Topix Chemical Index Topix Electrical Appliances Index Topix Foods Index Topix Insurances Index Topix Machinery Index Topix Mining Index Topix Pharmaceuticals Index Topix Pulp and Paper Index Topix Electric Power and Gas Index Topix Oil and Coal Products Index Topix Textile Products Index Topix Transportation Equipment Index MSCI Korea Index Korea SE Finance Major Index Korea SE Construction Major Index Korea SE Chemical Company Major Index Korea SE Food & Beverage Major Index Korea SE Insurance Major Index Korea SE Fabricated Metal & Machinery Major Index Korea SE Mining Major Index Korea SE Paper Product Major Index Korea SE Textile & Wear Major Index Korea SE Transport & Storage Major Index MSCI Malaysia Index KLSE Financial Index KLSE Mining Index Hong Kong Indonesia Italy Japan Korea Malaysia .Appendix I.
168 Appendix I.reﬁning and marketing General General General Hotels Banking and ﬁnance General General Banking and ﬁnance Construction and building materials Chemicals Paper and forest products General Banking and ﬁnance Construction and building materials Chemicals Electronics General Banking and ﬁnance Chemicals Electronics Technology Construction and building materials Energy Food Health care and pharmaceuticals Insurance Hotels Machinery Metals mining Paper and forest products Publishing Textiles Transportation Index MSCI Mexico Index Mexican SE Commercial & Transport Index Mexican SE Mining Index Mexican SE Construction Index MSCI New Zealand Index MSCI Norway Index Oslo SE Bank Index Oslo SE Insurance Index MSCI Philippines Index Philippine SE Mining Index Philippine SE Oil Index MSCI Poland Index MSCI Portugal Index MSCI Singapore Index AllSingapore Hotel Index AllSingapore Finance Index MSCI Spain Index MSCI Sweden Index Stockholm SE Banking Sector Index Stockholm SE Real Estate & Construction Index Stockholm SE Pharmaceutical & Chemical Index Stockholm SE Forest Industry Sector Index MSCI Switzerland Index SPI Banks Cum Dividend Index SPI Building Cum Dividend Index SPI Chemical Cum Dividend Index SPI Electronic Cum Dividend Index MSCI Thailand Index SET Finance Index SET Chemicals & Plastics Index SET Electrical Components Index SET Electrical Products &Computers Index SET Building & Furnishing Materials Index SET Energy Index SET Food & Beverages Index SET Health Care Services Index SET Insurance Index SET Hotel & Travel Index SET Machinery & Equipment Index SET Mining Index SET Pulp & Paper Index SET Printing & Publishing Index SET Textile Index SET Transportation Index New Zealand Norway Philippines Poland Portugal Singapore Spain Sweden Switzerland Thailand CreditMetrics™—Technical Document . Indices used for asset correlations Mexico Asset Category General Transportation Metals mining Construction and building materials General General Banking and ﬁnance Insurance General Metals mining Oil and gas .
reﬁning and marketing Paper and forest products Publishing Technology Telecommunications Textiles Transportation Utilities General Banking and ﬁnance Metals mining Index MSCI United Kingdom Index FTSEA 350 Banks Retail Index FTSEA 350 Media Index FTSEA 350 Building Materials & Merchants Index FTSEA 350 Chemicals Index FTSEA 350 Electronic & Electrical Equipment Index FTSEA 350 Food Producers Index FTSEA 350 Health Care Index FTSEA 350 Insurance Index FTSEA 350 Leisure & Hotels Index FTSEA 350 Extractive Industries Index FTSEA 350 Gas Distribution Index FTSEA 350 Oil Integrated Index FTSEA 350 Paper. Indices used for asset correlations 169 United Kingdom Asset Category General Banking and ﬁnance Broadcasting and media Construction and building materials Chemicals Electronics Food Health care and pharmaceuticals Insurance Hotels Metals mining Oil and gas .reﬁning and marketing Energy Paper and forest products Telecommunications Textiles Transportation General Automobiles Banking and ﬁnance Broadcasting and media Construction and building materials Chemicals Electronics Energy Entertainment Food Health care and pharmaceuticals Insurance Hotels Machinery Manufacturing Metals mining Oil and gas . Packaging & Printing Index FTSEA 350 Telecommunications Index FTSEA 350 Textiles & Apparel Index FTSEA 350 Transport Index MSCI United States Of America Index S&P Automobiles Index S&P Financial Index S&P Broadcasting (Television. Radio & Cable) S&P Building Materials Index S&P Chemicals Index S&P Electronics (Instrumentation) S&P Energy Index S&P Entertainment Index S&P Foods Index S&P Health Care Index S&P Insurance Composite Index S&P LodgingHotels Index S&P Machinery (Diversiﬁed) S&P Manufacturing (Diversiﬁed) S&P Metals Mining Index S&P Oil & Gas (Reﬁning & Marketing) S&P Paper & Forest Products Index S&P Publishing Index S&P Technology Index S&P Telecommunications (Long Distance) S&P Textiles (Apparel) S&P Transport Index S&P Utilities Index MSCI South Africa (Gross Dividends Reinvested) Johannesburg SE Financial Index Johannesburg SE Mining Holding Index United States South Africa .Appendix I.
170 Appendix I. Indices used for asset correlations MSCI Worldwide Asset Category Automobiles Banking and ﬁnance Broadcasting and media Construction and building materials Chemicals Electronics Energy Entertainment Food Health care and pharmaceuticals Insurance Hotels Machinery Metals mining Paper and forest products Telecommunications Textiles Transportation Utilities EMF Latin America Europe 14 Nordic Countries North America Paciﬁc Paciﬁc ex Japan Index Automobiles Price Index Banking Price Index Broadcasting & Pubs Price Index Construction & Housing (US$) Price Index Chemicals Price Index Electronic Comps/Inst. Price Index Energy Sources Price Index Recreation & Other Goods Price Index Food & Household Products Price Index Health & Personal Care Price Index Insurance Price Index Leisure & Tourism Price Index Machinery & Engineering Price Index Metals Nonferrous Price Index Forest Products/Paper Price Index Recreation & Telecommunications Price Index Textiles & Apparel Price Index Transport Shipping Price Index Utilities Electric & Gas Price Index MSCI Regional CreditMetrics™—Technical Document .
171 Reference .
172 CreditMetrics™—Technical Document .
. (See page 42). for completeness. or as the expected loss given that losses exceed a given level. a downgrade will suggest another downgrade soon). concentration risk.) autocorrelation (serial correlation).e. (See page 49..) credit exposure. by industry or location) obligors. But note that mean reversion does not technically necessitate negative autocorrelation.) average shortfall. (See page 58.. they predict a “quiet” period. with an evaluation of the level and stability of earnings and cash ﬂows. It is common to consider this reserve as the buffer for expected losses and some riskbased economic capital as the buffer for unexpected losses. see also Altman & Kao [92]. this includes revolving facilities. accounting analytic.g.) allowance for loan and lease losses. The use of ﬁnancial ratios and fundamental analysis to estimate ﬁrm speciﬁc credit quality examining items such as leverage and coverage measures. Portfolio risk resulting from increased exposure to one obligor or groups of correlated (e. The expected loss given that a loss occurs. (See page 6. commitment. A legally binding obligation (subject usually both to conditions precedent and to continuing conditions) to make available loans or other ﬁnancial accommodation for a speciﬁed period. But. When time series observations have a nonzero correlation over time. (See page 32.e. An accounting reserve set aside to equate expected (mean) losses from credit defaults. • Agency credit ratings typically exhibit move persistence behavior and are positively autocorrelated during downgrades (i. note that upgrades do not better predict future upgrades – we ﬁnd. The amount of exposure relevant to our credit analysis is the timebucketed average exposure in each forward period across the life of the transaction across all – probability weighted – market rate paths. Two empirical examples of autocorrelation are: • Interest rates exhibit mean reversion behavior and are often negatively autocorrelated (i. Credit exposure arising from marketdriven instruments will have an everchanging marktomarket exposure amount. Even during publicly known credit distress. The amount subject to changes in value upon a change in credit quality through either a market based revaluation in the event of an up(down)grade or the application of a recovery fraction in the event of default.173 Glossary of terms This glossary deﬁnes important terms in CreditMetrics.) . an up move one day will suggest a down move the next). (See page 60.) average exposure. a commit can be legally binding if drawndown before it is formally withdraw for cause. (See page 137.
These may be of all manner of seniority. A linear statistical measure of the comovement between two random variables. (See page 42.) credit distress. More speciﬁcally. They clearly deﬁne what is meant by “signiﬁcant” deterioration in the obligor’s credit quality. (See page 47. (See page 35.0. A ﬁrm can have many types of credit obligations outstanding.g. pronounced “rho”) will range from +1. the company is at risk if the bank fails just as much as the bank is at risk if the counterparty fails (although for the opposite movement in exchange or interest rates). The partner in a credit facility or transaction in which each side takes broadly comparable credit risk to the other. credit scoring.Y ( Xi – X )(Y i – Y ) COV X. Covenants are most effective when they are speciﬁc measures that state the acceptable limits for change in the obligor’s ﬁnancial and overall condition. swap positions often move in/outofthemoney and the relative credit risk changes accordingly.) ρ X . When the same two agree on an atthemoney forward exchange contract or swap. credit scoring refers to the estimation of the relative likelihood of default of an individual ﬁrm.) CreditMetrics™—Technical Document .). Y i= 1 = . etc. In our terminology. When a bank lends a company money. Aa. Cross default provisions are common: allowing acceleration of debt repayment. Aaa. (See page 65. Financial covenants are more explicit (and therefore more desirable) than a “material adverse change” clause. Generically. usually expressed in alphabetic terms (e. (See page 57.. it is the obligor that encounters credit distress carrying all of his obligations with him even though some of these may not realize a true default (i.= σ X ⋅ σY N N 2 2 ( Xi – X ) ⋅ (Yi – Y ) ∑ N ∑ ∑ i= 1 i= 1 counterparty. After inception. A correlation (Greek letter “ρ”.) covenants.174 correlation. The terms under which a credit facility will be monitored. The amount subject to either changes in value upon credit quality up(down)grade or loss in the event of default.) credit quality. Generally meant to refer to an obligor’s relative chance of default. (See page 43.e. this is a reference to the application of linear discriminant analysis to combine ﬁnancial rations to quantitatively predict the relative chance of default. it is not uncommon for different obligations to realize different recovery rates including perhaps 100% recovery – zero loss. In bankruptcy proceedings. security and instrument type.0 to 1. Observing “clumps” of ﬁrms defaulting together by industry or geographically is an example of positive correlation of default events. A.) credit exposure.. the borrower (not Counterparty) has no meaningful credit risk to the bank. some may have zero loss). CreditMetrics makes use of an extended deﬁnition that includes also the volatility of up(down)grades.
(See page 65. and/or lengthened maturity. it only states the amount at risk. (e.g. the amount it would cost to replace a transaction today should a counterparty default. Refer to Asquith. since the worst assumption is that the borrower draws the full amount and then goes bankrupt.S. Note that this leaves open the possibilities that: • Subordinated debt might default without impairing senior debt value.) dirty price. delayed sinking funds.) duration. exchange offers on traded bonds may be either registered with the SEC or unregistered if they meet requirements under Section 3(a)(9) of the Securities Act of 1933. otherwise. CreditMetrics estimates dirty prices since the coupon is paid in nondefault states but assumed not paid in default. options and derivatives) a proxy for exposure is estimated given the volatility of underlying market rates/prices. swaps. The likelihood that an obligor or counterparty will encounter credit distress within a given time period. (See page 65. For historical estimation of default probabilities. They may have a lower coupon. letters. debt holders may be effectively forced to accepted securities in exchange for their debt claim – such securities being of a lower value than the nominal present value of their original claim. then the current exposure would be the net replacement cost. a 6% annual coupon bond trading at par would have a dirty price of $106 just prior to coupon payment. the expected loss given that the loss is more extreme than that percentile level. etc. it does not make any attempt to assess the probability of loss. The longer the duration. lines. For marketdriven instruments. The weighted average term of a security’s cash ﬂows. for an unused or partly used facility it is the full amount of the facility.) distressed exchange. In the U. and • Transfers and clearing might continue even with a senior debt impairment. A generic term which includes loans. • Exposure is not usually a statistical concept. See Loan Equivalent Exposure. this would count as a default event since it can signiﬁcantly impair value.Glossary 175 current exposure. (See page 79. commitments. expected excession of a percentile level. (See page 137.) exposure. During a time of credit distress.. this is the full amount plus accrued interest. If there is an enforceable netting agreement with the counterparty. The amount which would be lost in a default given the worst possible assumptions about recovery in the liquidation or bankruptcy of an obligor. default probability. it would be the gross amount.) Reference . the larger the price movement given a 1bp change in the yield.. For a speciﬁed percentile level. Inclusion of the accrued value of the coupon in the quoted price of a bond For instance. Mullins & Wolff [89]. facility. (See page 10. foreign exchange. Any arrangement by which a bank accepts credit exposure to an obligor. For a loan or fully drawn facility. • For marketdriven instruments. “Credit distress” usually leads to either an omitted delayed payment or distressed exchange which would impair the value to senior unsecured debt holders.
foreign exchange. The property of a statistical distribution to have more occurrences far away from the mean than would be predicted by a Normal distribution. Labeling credit spread volatility as either market or credit risk is a question of semantics. A ﬁnancial letter of credit (also termed a standby letter of credit) is used to assure access to funding without the immediate need for funds and is triggered at the obligor’s discretion. Standalone obligors have some likelihood of each possible credit quality migration. but who now have much higher percentage risk due to recent downgrades. A trade letter of credit is typically triggered by a non credit related (and infrequent) event. the “beneﬁciary”. also known as the “account party”.176 fallen angels. Included in this was a draft of a master agreement by which institutions outlined their rights to net multiple offsetting exposures which they might have to a counterparty at the time of a default. A credit portfolio loss distribution will typically be leptokurtotic given positive obligor correlations or coarse granularity in the size / number of exposures. Between two obligors there is some likelihood of each possible joint credit quality migration. A promise to lend issued by a bank which agrees to pay the addressee. (See page 36.). 2 Xi – x 4 ( N – 1)( N – 3) N – 2N + 3 . (Institutional Swap Dealers Association) A committee sponsored by this organization was instrumental in drafting an industry standard under which securities dealers would trade swaps. – 3 =  σ x  ( N – 1)( N – 2)( N – 3) N ( N – 2)( N – 3) i=1 N KX ∑ Since the unconditional normal distribution has a kurtosis of 3. (See page 46). See leptokurtosis. Characterizes relative peakedness or ﬂatness of a given distribution compared to a normal distribution. letter of credit. excess kurtosis is deﬁned as Kx3 > 0. There are different types of letters of credit. whose large exposures were created when their credit ratings were better. The credit risk to the issuer of traded instruments (typically a bond.) kurtosis. The probabilities are commonly inﬂuences by the correlation between the two obligors. etc. ISDA. excess kurtosis is deﬁned as Kx3. CreditMetrics™—Technical Document . CreditMetrics addresses market price volatility as it is caused by changes in credit quality. Obligors having both relatively high percentage risk and relatively large exposure. leptokurtosis (fat tails). but also swaps. This means that a downside conﬁdence interval will be further away from the mean than would be expected given the standard deviation and skewness. A project letter of credit is secured by a speciﬁc asset or project income. under speciﬁed conditions on behalf of a third party. joint probabilities. It is the fourth moment of a distribution. Since a normal distribution has a kurtosis measure of 3. issuer exposure.
Sum of observation values divided by the number of observations. A transition matrix model is an example of a Markov process. pronounced “mu. forwards.) marginal statistic. called autocorrelation.) Reference .). (See page 71. Impact of a given asset on the total portfolio standard deviation. marketdriven instruments. Derivative instruments that are subject to counterparty default (e. options. This exposure/uncertainty is captured by calculating the netted mean and standard deviation of exposure(s).. The face amount of any loan outstanding plus accrued interest plus. the ability to trade in volume without directly moving the market price. A statistical measure of central tendency. when the replacement cost of the contract exceeds the original value). A mean calculated across a sample from a population is referred to as X . often measures as the ratio of current assets to current liabilities. For marketdriven instruments. often measured as bid/ask spread and daily turnover. This is on a much longer scale than another similar measure. See dirty price. Markov process.. etc.) mean. The distinguishing feature of these types of credit exposures is that their amount is only the net replacement cost – the amount the position is inthemoney – rather than a full notional amount. There are two separate meanings: • At the enterprise level. Credit quality migration describes the possibility that a ﬁrm or obligor with some credit rating today may move to (or “migrate”) to potentially any other credit rating – or perhaps default – by the risk horizon.e. while means calculated across the entire population – or means given exogenously – are referred to as µ. migration. A statistic for a particular asset which is the difference between that statistic for the entire portfolio and that for the portfolio not including the asset. There are two types of means. The statistical tendency in a time series to gravitate back towards a long term historical level. there is an amount at risk to default only when the contract is inthemoney (i.) 1 x = N ∑ xi i=1 N mean reversion. • At the security level. marginal standard deviation. It is the ﬁrst moment of a distribution. and these two behaviors are mathematically independent of one another. the ability to meet current liabilities as they fall due. loan exposure.Glossary 177 liquidity. (See page 24. swaps.g.” (See page 15. market exposure. A model which deﬁnes a ﬁnite set of “states” and whose next progression is determinable solely by the current state. (See page 129. (See page 47).
(See page 36. See rank order. netting by novation: The legal obligation of the parties to make required payments under one or more series of related transactions are canceled and a new obligation to make only the net payment is created. The ﬁrst moment is the mean which indicates the central tendency. There are international jurisdictions where the enforceability of netting in bankruptcy has not been legally tested. See transition matrix. move persistence. 2. moments (of a statistical distribution). The alternative would allow the liquidator to choose which contracts to enforce and which not to (and thus potentially “cherry pick”). equivalently.. netting. An approach to estimating the expected default frequency of a particular ﬁrm. the shape of any distribution can be described by its (inﬁnitely many) moments. In general though. this amount is not itself a cash ﬂow. For swaps. The ﬁnancial institution that extends credit on a facility which may later be held by another institution through. see Counterparty. Credit exposure arises – not against the notional – but against the present value (market replacement cost) of inthemoney future terminal payment(s). The fourth moment is the kurtosis which indicates the degree of central “peakedness” or. notional amount. The technique of estimating the likelihood of credit quality migrations. a loan sale. CreditMetrics™—Technical Document . Statistical distributions show the frequency at which events might occur across a range of values. obligor. (ii) a bond issuer. There are at least three types of netting: closeout netting: In the event of counterparty bankruptcy.) originator. all transactions or all of a given type are netted at market value. 4. positive autocorrelation). A party who is in debt to another: (i) a loan borrower. 3. The face amount of a transaction typically used as the basis for interest payment calculations. The third moment is the skewness which indicates any asymmetric “leaning” either left or right. settlement or payment netting: For cash settled trades.) option theoretic. (See page 5. monotinicity. etc. The statistical tendency in a time series to move on the next step in the same direction as the previous step (see also. (v) a contractor with unﬁnished performance. this can be applied either bilaterally or multilaterally and on related or unrelated transactions. The second moment is the variance which indicates the width. 1. The most familiar distribution is a Normal “Bell Shaped” curve. It applies Robert Merton’s modeloftheﬁrm which states that debt can be valued as a put option of the underlying asset value of the ﬁrm.178 migration analysis. the “fatness” of the outer tails. (iii) a trader who has not yet settled. (iv) a trade partner with accounts payable. for instance.
Expression in percent terms of the impact of a given asset on the total portfolio standard deviation. N S x = ( N – 1)( N – 2) N ∑ i=1 Xi – x 3 . CreditMetrics does not utilize this ﬁgure because it is not possible to aggregate tail statistics across a portfolio. etc. As two industries “load” – are inﬂuenced by – common factors. A statistical measure which characterizes the asymmetry of a distribution around its mean. percent marginal standard deviation. (iii) oil prices. A generic term referring to some facility which a client can use – or refrain from using – without canceling the facility.) percentile level. sector loadings. then such terms reduce the volatility of value across all nondefault credit quality migrations by keeping the instrument close to par.Glossary 179 peak exposure. the likelihood that the portfolio market falls below the 99th percentile number is 1%. A schedule of credit spreads listed by credit rating that are applied to either a loan or CreditSensitive Note (CSN) upon an up(down)grade of the obligor or issuer. Positive skews indicate asymmetric tail extending toward positive values (righthand side). the maximum (perhaps netted) exposure expected with 95% conﬁdence for the remaining life of a transaction.. (See page 129. “trade credit. It is the third moment of a distribution. Mathematicians term this property of data. monotonicity. (See page 42). skewness. This means that the conﬁdence interval out on the downside tail will be further Reference .) rank order.” that are at risk to the extent that the customer may not pay its obligation in full. σ  X The distribution of losses across a credit portfolio will be positively skewed if there is positive correlation between obligors or the size / number of exposures is coarsely granular. Non interest bearing short term extensions of credit in the normal course of business. revolving commitment (revolver). For marketdriven instruments. a ﬁrm or industry group is said to be dependent upon underlying economic factors or “sectors” such as: (i) the market as a whole. (See page 16. they will have a higher correlation between them. If the spreads are speciﬁed at market levels. serial correlation.) receivables.g. A measure of risk based on the speciﬁed conﬁdence level of the portfolio value distribution: e. Negative skewness implies asymmetry toward negative values (lefthand side). since it is not the case that these “peaks” will all occur at the same time. For correlation analysis. (ii) interest rates.) pricing grid. (See page 66. See autocorrelation. (See page 67. A quality of data often found across credit rating categories where values consistently progress in one direction – never reversing direction.
Thus.” pronounced “sigma”) is the square root of the second moment of a distribution.) unexpected losses. A square table of probabilities which summarize the likelihood that a credit will migrate from its current credit rating today to any possible credit rating – or perhaps default – in one period. standalone standard deviation. a new credit rating arrived at the risk horizon. (See page 129. Following a process which includes a random element.) standalone percent standard deviation. A credit rating migration outcome. (See page 25. See letter of credit. in retrospect. A measure of the maximum potential change in value of a portfolio of ﬁnancial instruments with a given probability over a preset horizon.) valueatrisk (VaR). 1 N–1 N σx = ∑ ( Xi – x) i=1 2 The distribution of losses across a credit portfolio will (typically) have a standard deviation which is much larger than its mean and yet negative losses are not possible. Standalone standard deviation expressed as a percentage of the mean value for the given asset.) stochastic. (See page 15. See “receivables. Standard deviation of value for an asset computed without regard for the other instruments in the portfolio. (See page 129.) trade credit. A popular term for the volatility of losses but also used when referring to the realization of a large loss which.” transition matrix. (See page 70. was unexpected.) standard deviation.180 away from the mean than would be expected given the portfolio’s standard deviation alone. state of the world. it is not meaningful to think of a standard deviation as being a +/. (See page 24.) CreditMetrics™—Technical Document . A standard deviation (Greek letter “σ. (See page 60. A statistical measure which indicates the width of a distribution around the mean.range within which will lie X% of the distribution – as one would naturally do for a normal distribution.) standby letter of credit. (See page 5. This can be either for a single obligor on a standalone basis or jointly between two obligors.
) 2 σx VAR x = 1 = N–1 ∑ ( Xi – x) i=1 N 2 Reference . It is the second moment of a distribution. which is the square root of the variance. A statistical measure which indicates the width of a distribution around the mean. (See page 16.Glossary 181 variance. A related measure is the standard deviation.
182 CreditMetrics™—Technical Document .
183
Bibliography
Alici, Yurt. “Neural Networks in Corporate Failure Prediction: The UK Experience,” Working Paper, University of Exeter (1995). Altman, Edward I. “Rating Migration of Corporate Bonds: Comparative Results and Investor/Lender Implications,” Working Paper, New York University Salomon Center, (Dec1996). Altman, Edward I. “Corporate Bond and Commercial Loan Portfolio Analysis,” Working Paper, (Sep1996). Altman, Edward I. “Defaulted Bonds: Demand, Supply and Performance, 19871992,” Financial Analysts Journal, (May/Jun1993), pp. 5560. Altman, Edward I. “Revisiting the High Yield Bond Market,” Working Paper Series, New York University Salomon Center (Feb1992). Altman, Edward I. “Measuring Corporate Bond Mortality and Performance,” The Journal of Finance, (Sep1989), pp. 909922. Altman, Edward I. “The Anatomy of the HighYield Bank Market,” Financial Analysts Journal, (Jul/Aug1987), pp. 1225. Altman, Edward I. “Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy”. Journal of Finance, (Sep1968), pp. 589609. Altman, Edward I., and Joseph C. Bencivenga. “A Yield Premium Model for the HighYield Debt Market,” Financial Analysts Journal, (Sep/Oct1995), pp. 4956. Altman, Edward I., and Allan C. Eberhart. “Do Seniority Provisions Protect Bondholders’ Investments?” Journal of Portfolio Management, (Summer1994), pp. 6775. Altman, Edward I., and Robert G.Haldeman. “Valuation, Loss Reserves and Pricing of Commercial Loans,” Working Paper Series, New York University Salomon Center (May1992). Altman, Edward I., and Duen Li Kao. “The Implications of Corporate Bond Ratings Drift,” Financial Analysts Journal, (May/Jun1992), pp. 6475. Altman, Edward I., and Duen Li Kao. “Rating Drift in High Yield Bonds,” The Journal of Fixed Income, (Mar1992), pp 1520. Altman, Edward I., and Duen Li Kao. “Examining and Modeling Corporate Bond Rating Drift,” Working Paper Series, New York University Salomon Center (1991). Altman, Edward I., Giancarlo Marco, and Franco Varetto. “Corporate Distress Diagnosis: Comparisons Using Linear Discriminant Analysis and Neural Networks,” Working Paper Series, New York University Salomon Center (Dec 1993).
Reference
184
Bibliography
Altman, Edward I., and Scott A. Nammacher. “The Default Rate Experience of HighYield Corporate Debt,” Financial Analysts Journal, (Jul/Aug1985), pp. 2541. Altman, Edward I., and Paul Narayanan. “Business Failure Classiﬁcation Models: An International Survey,” Working Paper, New York University, (May 1996). Asarnow, Elliot. “Best Practices in Loan Portfolio Management,” Journal of Lending and Credit Risk Management, (Mar1996), pp. 1424. Asarnow, Elliot. “Measuring the Hidden Risks in Corporate Loans,” Commercial Lending Review, Vol. 10, No. 1, (Winter 199495), pp. 2432. Asarnow, Elliot, and David Edwards. “Measuring Loss on Defaulted Bank Loans: A 24Year Study,” The Journal of Commercial Lending, (Mar1995). Asarnow, Elliot, and James Marker. “Historical Performance of the U.S. Corporate Loan Market: 19881993,” The Journal of Commercial Lending, Vol. 10 No. 2, (Spring 1995), pp.1332. Asquith, Paul, Robert Gertner, and David Scharfstein. “Anatomy of Financial Distress: An Examination of JunkBond Issuers,” Quarterly Journal of Economics, (Aug1994), pp. 625657. Asquith, Paul, David W. Mullins, and Eric D. Wolff. “Original Issue High Yield Bonds: Aging Analysis of Defaults, Exchanges, and Calls,” The Journal of Finance, (Sep1989), pp. 923952. Austin, Donald G. “Use Migration Analysis to Reﬁne Estimates of Future Loan Losses,” Commercial Lending Review, Vol. 7, No. 2, (Spring 1992), pp. 3443. Banks, Eric. Volatility and Credit Risk in the Capital Markets, (MacMillan Publishers Ltd., Merrill Lynch, Tokyo, 1993). Barclay, Michael J., and Clifford W. Smith. “The Priority Structure of Corporate Liabilities,” Journal of Finance, Vol. L, No. 3, (Jul1995), pp. 899917. Bennett, Thomas L., Stephen F. Esser, and Christian G. Roth. “Corporate Credit Risk and Reward,” Working Paper, Miller Anderson & Sherrerd (1993). Bensoussan, Alain, Michael Crouhy, and Dan Galai. “Stochastic Equity Volatility Related to the Leverage Effect,” Applied Mathematical Finance, Vol. 1, No. 1, (Sep1994), pp. 6385. Bernstein, Peter L. Against the Gods: The Remarkable Story of Risk. (New York: John Wiley & Sons, Inc., 1996). Blume, Marshall E., Howard Butcher III, and Donald B. Keim. “The Risk and Return of LowGrade Bonds: An Update,” Financial Analysts Journal, (Sep/Oct1991), pp. 8589.
CreditMetrics™—Technical Document
185
Blume, Marshall E., Donald B. Keim, and Sandeep A. Patel. “Returns and Volatility of LowGrade Bonds 19771989,” The Journal of Finance, Vol. XLVI, No. 1, (Mar1991), pp. 4974. Cantor, Richard, and Frank Packer. “Determinants and Impact of Sovereign Credit Ratings,” Economic Policy Review, Vol. 2, No. 2, Federal Reserve Bank of New York, (Oct 1996), pp. 3753. Cantor, Richard, and Frank Packer. “Multiple Ratings and Credit Standards: Differences of Opinion in the Credit Rating Industry,” Federal Reserve Bank of New York, Research Paper #9527, (Dec1995). Carty, Lea V., and Dana Lieberman. “Corporate Bond Defaults and Default Rates 19381995,” Moody’s Investors Service, Global Credit Research, (Jan1996). Carty, Lea V., and Dana Lieberman. “Defaulted Bank Loan Recoveries,” Moody’s Investors Service, Global Credit Research, Special Report (Nov1996). Chen, Yehning, J. Fred Weston, and Edward I. Altman. “Financial Distress and Restructuring Models,” Financial Management, Vol. 24, No. 2, (Summer1995), pp. 5775. Cook, Douglas O., and Lewis J. Spellman. “Firm and Guarantor Risk, Risk Contagion, and the Interﬁrm Spread among Insured Deposits,” Journal of Financial and Quantitative Analysis, (Jun1996), pp. 265281. Cooper, Ian A., and Antonio S. Mello. “The Default Risk of Swaps,” The Journal of Finance, Vol. XLVI, No. 2, (Jun1991), pp. 597619. Crabbe, Leland. “A Framework for Corporate Bond Strategy,” The Journal of Fixed Income, (Jun1995), pp. 1525. Crabbe, Leland, and Mitchell A. Post. “The Effect of a Rating Downgrade on Outstanding Commercial Paper,” The Journal of Finance, Vol. XLIX, No. 1, (Mar 1994), pp. 3956. Cumby, Robert E., and Martin D.D. Evans. “The Term Structure of Credit Risk: Estimates and Speciﬁcation Tests,” Working Paper, Georgetown University and New York University (Revised May1995). Das, Sanjiv Ranjan, and Peter Tufano. “Pricing CreditSensitive Debt When Interest Rates, Credit Ratings and Credit Spreads are Stochastic,” The Journal of Financial Engineering, Vol 5, No. 2, (Jun96), pp. 161198. DeGroot, Morris H. Probability and Statistics, Second Edition. Addison Wesley Publishing Co., Reading, Massachusetts (1986). Duffee, Gregory R. “Estimating the price of default risk,” Working Paper, Federal Reserve Board (Sep1995).
Reference
186
Bibliography
Duffee, Gregory R. “On Measuring Credit Risks of Derivative Instruments,” Finance and Economics Discussion Series, Federal Reserve Board, Washington, D.C., (Sep1994). Dufﬁe, Darrell, and Ming Huang. “Swap Rates and Credit Quality,” The Journal of Finance, Vol. LI, No. 3, (Jul1996), pp. 921949. Dutta, Soumintra, and Shashi Shekhar. “Bond Rating: A NonConservative Application of Neural Networks,” IEEE International Conference on Neural Networks, (Jul1988), pp. II443458. Eberhart, Allan C., William T. Moore, and Rodney L. Roenfeldt. “Security Pricing and Deviations from the Absolute Priority Rule in Bankruptcy Proceedings,” The Journal of Finance, Vol. XLV, No. 5, (Dec1990), pp. 14571469. Eberhart, Allan C., and Richard J. Sweeney. “Does the Bond Market Predict Bankruptcy Settlements?” The Journal of Finance, Vol. XLVII, No. 3, (Jul1992), pp. 943980. Episcopos, Athanasios, Andreas Pericli, and Jianxun Hu. “Commercial Mortgage Default: A Comparison of the Logistic Model with Artiﬁcial Neural Networks,” Working Paper, Clarkson University, Smith Barney Inc., City University of New York (1995). Foss, Gregory W. “Quantifying Risk in the Corporate Bond Markets,” Financial Analysts Journal, (Mar/Apr1995), pp. 2934. Franks, Julian R., and Walter N. Torous. “An Empirical Investigation of U.S. Firms in Reorganization,” The Journal of Finance, (Jul1989), pp. 747769. Fridson, Martin S., and Yan Gao. “Primary versus Secondary Pricing of HighYield Bonds,” Financial Analysts Journal, (May/Jun1996), pp. 2027. Ginzburg, Alex, Kevin J. Maloney, and Ram Willner. “Risk Rating Migration and the Valuation of Floating Rate Debt,” Working Paper, Citibank, Dartmouth College, Sanford C. Bernstein & Co. (Nov1993). Gollinger, Terri L., and John B. Morgan. “Calculation of an Efﬁcient Frontier for a Commercial Loan Portfolio,” Journal of Portfolio Management, (Winter 1993), pp. 3946. Golubchin, Leonid. “The Pure Economic Theory of Default Risk for Sovereign Debt,” Working Paper, City University of New York (Preliminary Draft, 29Dec94). Golubchin, Leonid. “The ARCH Structure of the Sovereign Debt Risk Spread,” Working Paper, City University of New York (Preliminary Draft, 29Dec94). Hamilton, Robert, Barry Howcroft, Keith Pond, and Liu Zhonghua. “Predicting the Fate of Insolvent Companies in Administrative Receivership — Using Statistical and Neural Classiﬁcation Methods,” Working Paper, Loughborough University, (Dec1995). Hradsky, Gregory T., and Robert D. Long. “HighYield Default Losses and the Return Performance of Bankrupt Debt,” Financial Analysts Journal, (Jul/Aug1989), pp. 3849.
CreditMetrics™—Technical Document
Carl. “The Spread Curve and the Risk/Return Decision. pp. (Jul1995). L. pp. “Corporate Debt Value.” The Journal of Fixed Income. Jarrow. Douglas J. “A Simple Approach to Valuing Risky Fixed and Floating Rate Debt..” The Journal of Fixed Income.” The Journal of Finance. (1995). (Mar1995). Hayne E. “Are Bank Regulators Ready for Credit Scoring of Commercial Loans?” Commercial Lending Review.. 4350.. No. Douglas J. Leibowitz. 612. Schwartz. Jónsson.187 Hurley.” Working Paper. 3. Francis A. Matthias. “The Effectiveness of Downgrade Provisions in Reducing Counterparty Credit Risk. and Stuart M. Fridson. University of Maryland (Oct1996). Lucas. Reference . Stephen. 4.” Forthcoming. pp. Stanley Kogelman. Longstaff. No. Vol. “Optimal Capital Structure. (Sep1994). Review of Financial Studies.” The Journal of Finance. 94101.. L. Leland. 1. Maden. (Renaissance Risk Publications. Bond Covenants. (Jun1995). “Valuing Credit Derivatives. “Managing Default Risk in Portfolios of Derivatives. and Martin S. and the Term Structure of Credit Spreads. “Corporate Distress Diagnosis — An International Comparison. “Forecasting Default Rates on HighYield Bonds.” Derivative Credit Risk: Advances in Measurement and Management. Vol. Leland.” The Journal of Fixed Income.” Journal of Portfolio Management. Jarrow. William J. (Fall 1995). Dilip B. “Pricing the Risks of Default. Kealhofer.” The Journal of Finance. XLIX. 5385. pp. “On the Pricing of Bond Default Risk. “Default Correlation and Credit Analysis. Martin L. and Lewis D. pp.” University of Freiburg.. (Winter 94/95).. 789819. Schwartz. Longstaff. (Jun1995). Turnbull. Jón G.” The Journal of Fixed Income. Turnbull. 1995). David Lando. 6670. 9871019. 6977. (Jun1995). (Winter 1996). 3241. Johnson. pp. pp. 12131252. Lucas. Francis A. pp.. “A Markov Model for the Term Structure of Credit Risk Spreads. and Haluk Unal.. and Klaus Bjerre Toft. pp. 3. Ronald N.” The Journal of Fixed Income. Vol. Robert A. (Jul1996). pp. “Pricing Derivatives on Financial Securities Subject to Credit Risk. and Eduardo S. pp. (revised Aug1995). Bader. and Eduardo S. No. (Jun1996). and Optimal Capital Structure. 9295. Vol. Robert A. LI. pp. “FixedIncome Risk Modeling in the 1990s. Hyndman. Kerling. No. Endogenous Bankruptcy. (Mar1995). and Stuart M.” Journal of Portfolio Management. 7687. Hayne E.” The Journal of Finance. and Lawrence R. Kahn.
(Mar1994). Vol.” The Journal of Commercial Lending. 5470. “The Valuation of Default Risk in Corporate Bonds and Interest Rate Swaps. (Mar/Apr1994). pp. and Pricing: The Need for a New Shared Database. “Liquidation Values and Debt Capacity: A Market Equilibrium Approach.” Working Paper. Andrei. Frederick C. Shimko. and Peter Zangari. No. “Optimal Trade Credit Limits. (New York: Morgan Guaranty Trust Co. “Using Quantitative Methods to Support CreditRisk Management. “The Pricing of Risky Debt When Interest Rates are Stochastic.. Raghuram G.. “Diversiﬁcation and Risk in Banking: Evidence from Ex Post Returns. D. revised 13Feb96). No. (Spring1996). Lawrence. McAllister. “Some Empirical Estimates of the Risk Structure of Interest Rates. Shane. L. 959985. (1974). Smith. Vol. Ronn. Mingo. CreditScoring. pp. 19. Naohiko Tejima. 7989. 1. pp. No. 13511360. Queen. pp.. Oded. 7378. and John J. 5. pp.” Finance and Economics Discussion Series.” Financial Analysts Journal. Longerstaey. and Arthur Warga. “On the Pricing of Corporate Debt: The Risk Structure of Interest Rates. (1995). Douglas W. “What Do We Know about Capital Structure? Some Evidence from International Data. 11 No. pp. XLIV. Rajan. Van Deventer. 76. pp.. No.” The Journal of Fixed Income.” Commercial Lending Review. Skinner. 926.” The Journal of Finance. McManus. 5. pp. “Comovements of Lowgrade Debt and Equity Returns of Highly Leveraged Firms. No. Vol. “A Trinomial Model of Bonds with Default Risk. and Ehud I. Patrick H. “Forecasting losses on a liquidating longterm loan portfolio. 13431366. Maggie. 4. Vol. CreditMetrics™—Technical Document . 622. Douglas. Hilary. University of Texas at Austin (Oct1994. Washington. pp. Vol.” The Journal of Finance. 29. 14211460.. Robert C. Shleifer. and Richard Roll.188 Bibliography McAllister.” Financial Management. 9. Federal Reserve Board. “Firm Mortality: Using Market Indicators to Predict Survival. pp. (Dec1995). Vol. pp. 7185. and Edward C. Meyer. XLVII. and Douglas A. 1996) Sarig. and Donald R. Frank S. Jacques. Merton. Nielsen. (1994). and Robert W. (Jun1992). 5865.C. Scherr. Vishny. and Luigi Zingales. RiskMetrics™—Technical Document. pp. (May/Jun1987). 25. (Dec1989). Vol. 4th Edition. (Sep1992).” The Journal of Finance. 449470.” Financial Analysts Journal. “Commercial Loan Risk Management. Vol. L.” The Journal of Fixed Income.” Journal of Banking & Finance. Patrick H. Soren S.” The Journal of Finance. (Winter 199596). David C.. 1. (Sep 1993).
pp. Stevenson. 2631. Eric K. 4047. 3rd Edition.” Financial Analysts Journal.. A. (May/Jun1993). Fadil... (May/Jun1994). Linear Algebra and Its Applications. Standard & Poor’s. 4452. David J. Ward. (Jun1995). Wagner. Strang. Long. and J. (Spring 1995). and Gary L.. Bollier. 6165. 1988). (Orlando: Saunders College Publishing. No. Swank. “Risk and Return in Defaulted Bonds. Herbert S.” Working Paper. and Thomas H. 2. CreditWeek. Reference . pp. and Michael W. 412. 2333. 10. III. (Jun1996). “Bonds in Default: Is Patience a Virtue?” The Journal of Fixed Income. 1996). Eric H. pp. Griepentrog. Root. “The Pricing of Bonds in Bankruptcy and Financial Restructuring. “Assessing Financial Distress with Probabilistic Neural Networks. “Modern Portfolio Theory: Can It Work for Commercial Loans?” Commercial Lending Review.189 Sorensen. Vol. Gilbert. pp. and Thierry F. (April 15.” Financial Analysts Journal. Bruce G. “Pricing Swap Default Risk. Thomas A.. Tyree.” The Journal of Fixed Income. pp. pp. City University of London (1994).
190 Bibliography CreditMetrics™—Technical Document .
Pericli & Hu [95] expected default frequencies exposure reduction exposure size. 45. 61. 65 58 24 58 58 58. 26. bonds Carty & Lieberman [96a] Carty & Lieberman [96b] commitments to lend concentration risk conﬁdence interval calculation conﬁdence level correlation of two binomials Crabbe [95] credit derivatives Numerics A 78 59 60 58. Marco & Varetto [93] Asarnow & Edwards [95] Asarnow & Marker [95] Asarnow [96] Asquith. 78. 173 32 26. Mullins & Wolff [89] Austin [92] autocorrelation average shortfall average value Bank for International Settlements bellshaped distributions Bennett. modeling Dun & Bradstreet Dutta & Shekhar [88] Eberhart & Sweeney [92] Eberhart. 58. 138 59 133 8 133 8 57. 175 60 32 137 16 140 17 62 139 77 80 18. 60. 133 66 65 65 65. 67. 78 D credit distress credit exposure credit quality credit rating systems credit risk computing on different horizons in a portfolio limits portfolio pricing of credit risk measures credit scoring CreditMetrics steps to quantify risk Das & Tufano [96] debt classes default event deﬁnition mean estimate default probability discriminant analysis distressed exchanges distribution of portfolio value.191 Index Ward & Griepentrog Alici [95] allowance for loan and lease losses Altman [87] [88] [89] [92] Altman & Bencivenga [95] Altman & Haldeman [92] Altman & Kao [91] [92] [92b] Altman & Kishore [96] Altman & Nammacher [85] Altman. Esser & Roth [93] Bernstein [96] beta distribution beta distribution. 160 79 18 6 16 30 16 157 60 7 G . Moore & Roenfeldt [90] economic capital assessment Episcopos. 81 140 B E C F iv. absolute fallen angels Fons [91] Foss [95] FourFifteen Fridson & Gao [96] FX forwards Ginzburg. Maloney & Willner [93] Gollinger & Morgan [93] Greenspan. 71 59 58. 58. 62. 66 42 65 65 32 15 135 38 57 15 58 24 57. 60 64. 133 60 60. 81 59 78 133 138. 139 59 64 140 133 134. 78 58 59 79 44. 21. 60. 65. 43. 82. 62 62 58. 65 8 64. Alan 65.
47 72. 60 36 36 59 85 161 59 59. 137 92 optionality optiontheoretic approach percent marginal standard deviation percentile level performance evaluation portfolio credit risk portfolio effects principal components analysis probability of default. Lando & Turnbull [96] joint likelihoods joint probabilities Jónsson & Fridson [96] Kealhoffer [95] Keiser [94] Kerling [95] KMV Corporation letters of credit limit setting Lo Pucki & Triantis [94] loan commitment loans logistic regression logit analysis Lucas [95a] [95b] Madan & Unal [96] marginal risk marginal standard deviation marginal statistic market volatilities marketdriven exposure uncertainty marketdriven instruments Markov process Markowitztype analysis McAllister. 57. statistical RiskMetrics S S&P CreditWeek [96] Sarig & Warga [89] scenario generation scenarios default nondefault selforganizing feature maps Shimko. 62. 75 63 93 32 36. 81 19. 137 140 7 81 59 59 161 18. 64. 79 59 59 158 60 133 35. 133 60 iv. 42 77 77 77 79 79 8 98 40 133 iii 58 133 113 116 116 59 133 113. 40 129 118 8 8 19. 85.192 H J historical tabulation Hurley & Johnson [96] jackkniﬁng Jarrow & Turnbull [95] Jarrow. See Default event probit analysis K L R M Rajan & Zingales [95] receivables Recoveries mean estimate uncertainty recovery rate recovery rate distribution recovery rate estimates risk distinction between market and credit idiosyncratic marginal risk level. Tejima & VanDeventer [93] simulation Skinner [94] Smith & Lawrence [95] Sorenson & Bollier [94] standalone percent standard deviation standalone standard deviation Standard & Poor’s standard deviation standardized asset returns CreditMetrics™—Technical Document . 46 140 161 43 21. Patrick H. 60 116 113 137 58 155 N Neilsen & Ronn [96] netting neural network techniques O P 133 7 59 48 59 129 118. mean reversion Merton [74] Meyer [95] migration analysis Monte Carlo estimation simulation Monte Carlo simulation Moody’s Investors Service mutually exclusive outcomes 67 133 149 133 57. 118 133 60 133 129 129 58 15.
113 64. 81 115 66 78 time horizon choosing trade credit transition matrix Tyree & Long [94] U V 31 18. 25. 66. 75 59 60 5 62 161 59. Inc.Index 193 states of the world Stevenson & Fadil [95] Strang [88] subordinated debt Swank & Root [95] T 24. 42 20. 71. Reference . 81 unexpected losses valueatrisk W Z Wagner [96] White [93] Zeta Services. 64.
194 CreditMetrics™—Technical Document .
195 Part III: Applications .
196 CreditMetrics™—Technical Document .
197 Part III: Applications .
198 CreditMetrics™—Technical Document .
P. Morgan representative or: New York London Singapore Greg M. Numerical examples are implemented in Excel spreadsheets.P. J.P.com Michael Wilson (65) 3269901 wilson_mike@jpmorgan. These Excel spreadsheets are intended as a demonstration of the CreditMetrics credit risk management methodology.com Guy Coughlan (44171) 3255384 coughlan_g@jpmorgan. Separately.com CreditMetrics™—Technical Document . They have been designed as an educational tool and should not be used for the risk estimation of actual portfolio positions.199 Look to the J. Gupton (1212) 6488062 gupton_greg@jpmorgan. Morgan sells software which embodies the CreditMetrics methodology.com for updates of examples illustrating points from this technical document or useful new tools. Morgan site on the Internet located at http://jpmorgan. If you have any questions about the use of these spreadsheets contact your local J.
com Linda Bammann (1212) 3351085 linda.com CreditMetrics™ is based on. please contact the authors or any cosponsors listed below: J.P. Morgan Guaranty Trust Company is a member of FDIC and SFA. methodology. Opinions and estimates constitute our judgment and are subject to change without notice. Worldwide CreditMetrics™ Contacts For more information about CreditMetrics™. timeliness. Tavakoli (1312) 8284732 Philip Basil (44171) 6344482 Walter Bloomenthal (1312) 8281668 Bank of Montreal BZW Barry Campbell (1416) 8674809 Loretta Hennessey (1212) 6051541 Jo Ousterhout (1212) 4126893 Michael Dyson (44171) 9563045 Loren Boston (852) 29032588 Deutsche Morgan Grenfell KMV Corporation Swiss Bank Corporation Hugo Bänziger (44171) 5452562 David Nordby (1415) 7563337 edfs@kmv. Morgan entity in their home jurisdiction unless governing law permits otherwise. documentation or any information derived from the data (collectively the “Data”) and does not guarantee its sequence.P. Additional information is available upon request. Clients should contact analysts at and execute transactions through a J. Past performance is not indicative of future results.P. Morgan and any of the the cosponsors. you can call (1800) JPMINET in the United States. completeness or continued availability. the credit risk management systems developed by J. It also discloses our approach to estimating credit exposures by instrument type and a method of estimating correlations of credit quality comovements.com Robert Gumerlock (411) 2395739 robert.gumerlock@swissbank.P.jpmorgan. Morgan may hold a position or act as market maker in the ﬁnancial instruments of any issuer discussed herein or act as advisor or lender to such issuer.ubs. CreditMetrics™ – Technical Document: A manual describing the CreditMetrics™ methodology for estimating credit risks. All the above can be downloaded from the Internet at http://www. Finally. J. Morgan does not warrant any results obtained from the use of the CreditMetrics™ data. 1997 page 200 CreditMetrics™ Products Introduction to CreditMetrics™: An abbreviated document which broadly describes the CreditMetrics™ methodology for measuring portfolio credit risk. It fully speciﬁes how we construct the volatility of value due to credit quality changes for both standalone exposures and portfolios of exposures.P.bammann@swissbank.P. Morgan & Co.com or downloading the CreditMetrics™ data ﬁles. CreditMetrics™ Monitor: A semiannual publication which will discuss broad credit risk management issues. defaults or market movements.com (1212) 6483461 cmx_amer@jpmorgan. downgrades. The Data is calculated on the basis of historical observations and should not be relied upon to predict future credit upgrades.com/RiskManagement/CreditMetrics CreditManager™ PC Program: A desktop software tool that implements the methodology of CreditMetrics and produces valueatrisk reports and other analysis of credit risk such as those outlined in the CreditMetrics documents. Information herein is believed to be reliable.com (852) 29735646 cmx_asia@jpmorgan. but J.com Union Bank of Switzerland Hei Wai Chan (1212) 8215547 nycnh@ny.P. . the manual describes the format of the data set.com (44171) 3258007 cmx_euro@jpmorgan. accuracy. Trouble accessing the Internet? If you encounter any difﬁculties in either accessing the J. but differs signiﬁcantly from. Morgan does not warrant its completeness or accuracy. Examples are for illustrative purposes only. J. actual risks will vary depending on speciﬁc circumstances. CreditManager can be purchased from J. Morgan for its own use.CreditMetrics™—Technical Document First Edition.P. CreditMetrics™ data set: A set of historical statistics and results of academic and industry studies which will be updated periodically.jpmorgan.P. Morgan Americas Europe Asia pacific Cosponsors Bank of America Janet M. statistical questions as well as new software implementations and enhancements. Copyright 1997 J. Morgan home page on http://www. This material is not intended as an offer or solicitation for the purchase or sale of any ﬁnancial instrument. Incorporated.