Risk MGT CRD Risk

Agenda
Facoltà di Scienze Bancarie, Finanziarie e Assicurative
• Overview
Credit Risk & Scoring – the main drivers of Credit Risk
• PD estimation: backward-looking
models statistical models or “scoring models”
– Discriminant analysis
– Logit/probit models
Courtesy of Andrea Sironi – Neural networks
– Genetic algorithms
© M. Anolli - Risk Management Page: 275
The three (main) factors driving

Credit risk guidelines: Basel
future losses on a single credit
Facoltà di Scienze Bancarie, Finanziarie e Assicurative Facoltà di Scienze Bancarie, Finanziarie e Assicurative
EAD and LGD are • A. Establishing an appropriate credit risk environment

conditional upon ~ ~ ~ – Principle 1: The board of directors should have responsibility for
default; D is not Li D EAD LGD approving and periodically (at least annually) reviewing the credit
risk strategy and significant credit risk policies of the bank. The
Loss given strategy should reflect the bank’s tolerance for risk and the level
of profitability the bank expects to achieve for incurring various
default credit risks.
Default One minus the – Principle 2: Senior management should have responsibility for
binomial variable Exposure at default recovery implementing the credit risk strategy approved by the board of
directors and for developing policies and procedures for
(1 = defaulted, 0 = (think of credit rate identifying, measuring, monitoring and controlling credit risk.
alive), its mean is the commitments, or = 1 - RR Such policies and procedures should address credit risk in all of
the bank’s activities and at both the individual credit and portfolio
PD, probability of over-the-counter levels.
default contracts) – Principle 3: Banks should identify and manage credit risk inherent
in all products and activities. Banks should ensure that the risks of
NOTES: products and activities new to them are subject to adequate risk
1. This is a simplified binomial scheme, management procedures and controls before being introduced or
undertaken, and approved in advance by the board of directors or
that does not account form downgrade effects and maturity its appropriate committee.
2. Unexpected losses on a portfolio also depend on correlations
Page: 276
© M. Anolli - Risk Management © M. Anolli - Risk Management 277
Credit risk guidelines: Basel Credit risk guidelines: Basel
• B. Operating under a sound credit granting process • C. Maintaining an appropriate credit administration, measurement and
– Principle 4: Banks must operate within sound, well-defined credit- monitoring process
granting criteria. These criteria should include a clear indication of – Principle 8: Banks should have in place a system for the ongoing
the bank’s target market and a thorough understanding of the administration of their various credit risk-bearing portfolios.
borrower or counterparty, as well as the purpose and structure of – Principle 9: Banks must have in place a system for monitoring the condition
the credit, and its source of repayment. of individual credits, including determining the adequacy of provisions and
reserves.
– Principle 5: Banks should establish overall credit limits at the level – Principle 10: Banks are encouraged to develop and utilise an internal risk
of individual borrowers and counterparties, and groups of rating system in managing credit risk. The rating system should be
connected counterparties that aggregate in a comparable and consistent with the nature, size and complexity of a bank’s activities.
meaningful manner different types of exposures, both in the – Principle 11: Banks must have information systems and analytical
banking and trading book and on and off the balance sheet. techniques that enable management to measure the credit risk inherent in
– Principle 6: Banks should have a clearly-established process in all on- and off-balance sheet activities. The management information
place for approving new credits as well as the amendment, system should provide adequate information on the composition of the
renewal and re-financing of existing credits. credit portfolio, including identification of any concentrations of risk.
– Principle 12: Banks must have in place a system for monitoring the overall
– Principle 7: All extensions of credit must be made on an arm’s- composition and quality of the credit portfolio.
length basis. In particular, credits to related companies and – Principle 13: Banks should take into consideration potential future changes
individuals must be authorised on an exception basis, monitored in economic conditions when assessing individual credits and their credit
with particular care and other appropriate steps taken to control portfolios, and should assess their credit risk exposures under stressful
or mitigate the risks of non-arm’s length lending. conditions.
© M. Anolli - Risk Management 278 © M. Anolli - Risk Management 279
Credit risk guidelines: Basel PD estimation

• D. Ensuring adequate controls over credit risk • To estimate PD, one must preliminarily define:
– Principle 14: Banks must establish a system of independent, ongoing
assessment of the bank’s credit risk management processes and the – what default is
results of such reviews should be communicated directly to the board of • “unable or unwilling to pay”
directors and senior management.
– Principle 15: Banks must ensure that the credit-granting function is • subjective judgment versus objective conditions
being properly managed and that credit exposures are within levels
consistent with prudential standards and internal limits. Banks should • The probability of default can then be assessed by:
establish and enforce internal controls and other practices to ensure that – Backward-looking models
exceptions to policies, procedures and limits are reported in a timely
manner to the appropriate level of management for action. • Use past information (e.g., financial ratios, news,
– Principle 16: Banks must have a system in place for early remedial etc.)
action on deteriorating credits, managing problem credits and similar
workout situations. • Can rely on human judgment and/or automated
• E. The role of supervisors algorithms
– Principle 17: Supervisors should require that banks have an effective – Discriminant Analysis, Logit and Probit models, Neural
system in place to identify, measure, monitor and control credit risk as Networds, Genetic Algorithms and more
part of an overall approach to risk management. Supervisors should
conduct an independent evaluation of a bank’s strategies, policies, – Forward-looking models
procedures and practices related to the granting of credit and the • Include expectations on future developments
ongoing management of the portfolio. Supervisors should consider
setting prudential limits to restrict bank exposures to single borrowers or • Usually based on market data (more later)
groups of connected counterparties.
© M. Anolli - Risk Management 280 © M. Anolli - Risk Management Page: 281

Automated rating systems Automated rating systems/2
0.05%
0.9%
Financial
Default 3 Financial 3 1.3%
statements: e.g., statements: e.g.,
liquidity ratios probability liquidity ratios Score 1.7%
2.5%
1.3% 33 3%
5%
Central Credit Central Credit
Registry: e.g., Registry: e.g., 10%
decrease in the decrease in the
# of banks An array of quantitative data # of banks An array of quantitative data
on the borrower is transformed on the borrower is transformed
Internal data: into a default probability. This, in turn, Internal data:
into a score, leading to a specific rating grade.
e.g., overdrafts can originate a binary judgment (A/R) e.g., overdrafts
The behavior of each grade is monitored over time,
or the assignment to a specific rating grade. so that a PD can be Page:
inferred.
© M. Anolli - Risk Management Page: 282 © M. Anolli - Risk Management 283
Discriminant analysis: Discriminant analysis:

the idea behind all technicalities the idea behind all technicalities
• The bank’s borrowers are described through a Good firms show low average values.
set of variables Moving away from the means,
Interest expenses over turnover

– E.g. 2, 10, 80 financial ratios the probability of belonging
• The borrowers belong to two (ore more) to this group decreases.
groups The shape of the probability
curves depends on the
– Good firms versus bad ones
(firms that will default one year later) variance/covariance
of the two variables
• The above-mentioned variables take
(here, positively
structurally different values ( and ) in the
correlated)
two groups
– From the variables’ values one can infer the
group to which a firm belongs
Unauthorized overdraft over total exposure
Page: 285
© M. Anolli - Risk Management Page: 284 © M. Anolli - Risk Management
“Bad” firms show higher This firm clearly

average values, while belongs to the

the variance/covariance “green” group
structure is similar
to the “good” group.
P = 98% P=2%
Unauthorized overdraft over total exposure Unauthorized overdraft over total exposure
© M. Anolli - Risk Management Page: 286 © M. Anolli - Risk Management Page: 287

P = 1% P=99%

This firm clearly
belongs to the red
group Grey area
Unauthorized overdraft over total exposure Unauthorized overdraft over total exposure
Discriminant analysis can be
DA: the computation of a score
used in 2 ways
• The score can be seen as a “weighted • To estimate a probability of being good

average” of the scores reported by a or bad (probability of default)
student in a set of different subjects – This is just a “quick & dirty” probability,
– Unimportant variables are weighted close to overlooking qualitative variables and
zero drawing on a multinormality assumption
– Important ones get high weights • To summarize into a “synthetic”
– Counterproductive variables get negative variable D all information that was
coefficients dispersed among n natural variables
• Weights are chosen in such a way that – E.g., interest expenses over turnover (IET)
the good firms’ scores be as far as and unauthorized overdrafts over total
possible from the bad firms’ ones exposure (OE)…
From natural variables… …to a synthetic one (score)

D(x) 1 x1 2 x2 IET xIET OE xOE

0.9
OF over turnover
0.8
0.7
0.6
Good
x1: interest expenses
0.5
IET
0.4
0.3
0.2
cattive
bad Bad
0.1
good
buone
0
0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70
0 2 4 6 8
x2: unauthorised overdrafts
SAover total credit exposure
D(x)
OE
Page: 292
© M. Anolli - Risk Management © M. Anolli - Risk Management Page: 293
The choice of the weights i The choice of the weights i
Consider the artificial variable D = D(x): Which are the best weights?
The ones that better divide the good ones from the bad ones.
D D(x) γx Formally, the ones which maximize the standardized distance
If x1 and x2 are the two vectors with mean values for “good” and “bad”, then between the good and bad means (centroids)
the score for the good and bad ones takes the following values:
D1 D2 γ x1 γ x2
D1 γ x1 D2 γ x2 γ Max
D D
While the variance of D, as for any linear combination of random variables, is
equal to: For simplicity, we maximize the square of the above
2 2
2
γ Σx γ γ x1 γ x2 γ x1 γ x 2
where
D
is the variance/covariance matrix of the natural variables xi.

Max γ
2
γ Σx γ
q
x D
How to use the scores to

Centroids
classify
Take a point
which is in
between the two
centroids centroids and
good
Buone classify as good a
Buone
Good company if its
score is higher
than , bad if its
Cattive
bad score is higher
than
Cattive
Bad 0 2 4
z 6 8
D(x)
0 2 4 6 8 D1 γ x1 D2 γ x2
D(x)
It can be shown that the weights used to compute D(x) are chosen in such a way
D1 D2 1
that the “good” and “bad” scores tend to gather as close as possible around γ x1 x2
their respective centroids 2 2 Page: 297
How to use the scores to
Example
classify
• Consider the data for 38 companies

– 24 are good
• One year after the data collection were not failed
wrong
Buone – 14 are bad
• One year after data collection they were defaulted
• For each company we analyze
Cattive – Interest Expenses on Turnover (IE/T)
– overdrafts over total exposure (O/TE)
Given company i, with
0 2 4 6 8
natural variables xi, • Note that, for abnormal companies, the
classify it as good if: data were not recorded when the financial
D xi γ xi crisis was already apparent, but some
Bad Good
months earlier (e.g. one year earlier)
divides good ones and bad ones, and is called cut-off point
Page: 298
Table 1:
a simplified example
x2:
unauthorised
x2:
unauthorised
Example
overdrafts overdrafts
x1: interest
Facoltà di Scienze Bancarie, Finanziarie e Assicurative over total x1: interest over total Facoltà di Scienze Bancarie, Finanziarie e Assicurative
“Good” expenses over credit “Bad” expenses over credit
companies: turnover exposure companies: turnover exposure
Company 1 0% 0% Company 25 74% 36%
0.9
OF over turnover
Company 2 72% 40% Company 26 85% 10%
Company 3 75% 31% Company 27 67% 42%
0.8
Company 6 1% 2% Company 30 72% 64% 0.7
Company 8 42% 3% Company 32 81% 32% 0.6
Company 9 36% 12% Company 33 60% 51%
x1: interest expenses

Company 10 12% 9% Company 34 72% 0% 0.5
Company 11 65% 25% Company 35 58% 6%
Company 12 16% 9% Company 36 64% 11% 0.4
Company 13 45% 5% Company 37 55% 21%
Company 14 0% 0% Company 38 65% 47% 0.3
Company 15 65% 0%
Company 16 16% 2% 0.2
Company 17 70% 33% cattive
bad
Company 18 29% 15% 0.1
Company 19 0% 32%
good
buone
Company 20 0% 0% 0
Company 21 54% 19%
0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70
Company 22 9% 0%
Company 23
Company 24
0%
57%
4%
24%
x2: unauthorised overdrafts
SAover total credit exposure
Mean values: 29,1% 11,3% 67.4% 31.2%
Scores D(x) for the sample Scores D(x) for the sample
companies
companies
Similarly, we can compute the score for every company

zi γ xi 6.09 x1,i 4.84 x2,i
“Good” “Bad”
companies: zi companies: zi
Company 1 0,000 Company 25 -6,237
Negative weight to both variables high financial charges on Company 2 -6,295 Company 26 -5,657
Company 3 -6,065 Company 27 -6,085
turnover reduce the company’s score, and the same applies to Company 4 -0,526 Company 28 -6,106
unauthorised overdrafts Company 5 -0,097 Company 29 -6,349
The first one receives a higher weighting, confirming its greater Company 7 -1,850 Company 31 -4,920
discriminant power Company 8 -2,733 Company 32 -6,479
Using the values found above, the score of any company can be Company 10 -1,167 Company 34 -4,352
calculated. For example, for company 2 this is
The two groups values are “clustered” around their respective centroids (-2.32
and -5.61). However, they are not perfectly separate, as some healthy
companies have relatively low scores, and viceversa
z2 6.09 0.27 4.84 0.05 1.85
Computing the cut-off

The model ability
and the classification rule
1 1 0.29 0.67 Classification

γ ( x1 x 2 ) 6.09 4.84
2 2 0.11 0.31 Good Bad Total
1 19 5 24
( 6.09 0.96 4.84 0.42) 3.97 Good 50.0% 13.2% 63.2%
2
True 4 10 14
Given company i, with natural variables xi, classify it as groups Bad 10.5% 26.3% 36.8%
good if:
23 15 38
D xi γ xi Total 60.5% 39.5% 100.0%
D(x i ) 6.09 xOF ,i 4.84 xSA,i 3.97

Rate of correct classification (hit rate): 76.3%
The Altman Z score Wilks’ Lambda
• Formally, D(x) minimizes the ratio

The best-known discriminant score applied to credit risk is the one between residual deviance and total
developed by Ed Altman in 1968 for listed US companies. It is a deviance (Wilks’ Lambda):
function of 5 independent variables
zi 1.2 xi ,1 1.4 xi , 2 3.3 xi ,3 0.6 xi , 4 1.0 xi ,5 (d i d1 ) 2 (d i d 2 )2
x1 = working capital/total assets i g1 i g2
x2 = retained profits/total assets N
x3 = earnings before interest and tax/total assets (d i d )2 Lambda can be seen as
x4 = mkt value of assets/book value of long-term debts i 1 a goodness-of-fit
measure (in sample);
x5 = turnover/total assets (d i D ( x i )) from 1 (total
The greater the z score of a company, the better its quality uselessness) to 0 (strong
effectiveness).
Cut-off point at 1.81
Costs associated Costs associated

with different errors
with different errors
The costs of the two mistakes can

… then … then be deeply different. This explains
pays does not why credit analysts are so
“conservative” when issuing new
back pay … then
pays
… then
does not loans.
back
Assigned Assigned
pay
Can a DA model be made
to the OK -70% to the OK -70% more conservative?
“good”… “good”…
Assigned
Yes, by forcing it to choose based
on the lower expected error cost,
Assigned to the
“bad”…
-2% OK not base on the higher probability
to the -2% OK This alternate criterion still lends

“bad”… itself to threshold setting
and grey areas
New rule: classify based on the Different classification rules
costs of wrong classification Table 3:
Facoltà di Scienze Bancarie, Finanziarie e Assicurative Facoltà di Scienze Bancarie, Finanziarie e Assicurative Decisions prompted by the model
based on different cut-off points
“Good” Cut-off point “Bad” Cut-off point
Classify as good if companies ’ ” companies ’ ”
The change from ’ to ” Company 1 Company 25
Company 2 Company 26
makes the model more
C (1 | 2) p (2 | x) C (2 | 1) p (1 | x) selective (a loan is
Company 3
Company 4
Company 27
Company 28
now granted to only 8 Company 5 Company 29
Cost of Cost of Company 6 Company 30
Probability Probability companies out of 38) Company 7 Company 31
classifying as of being bad classifying as of being good we have “taught” Company 8 Company 32
good if it is bad bad if it is good our model to be more Company 9 Company 33
prudent Company 11 Company 35
If customers were to Company 12 Company 36
i.e., if: provide more
C (1 | 2)
2 collateral, the cost of a Company 15

γx log type II error would fall Company 16
1C ( 2 | 1)
Company 17
the cut-off point ” Company 18
depends partly on the Company 19
Company 20
loan’s LGD Company 21
Note: as usually C(1|2)>C(2|1), then ”> ’: the cutoff point moves toward Company 22
the right, i.e. classifying a company as good is more difficult Company 23
© M. Anolli - Risk Management Page: 310 Company 24
The selection of the Discriminant analysis:

variables
hypotheses
• The selection of discriminant variables can follow 1) Var/cov matrices of the independent variables are equal
2 procedures for the two groups of companies empirical data
• Simultaneous or direct method suggest the opposite heteroscedasticity between
groups
– The model is constructed on an a priori basis. In
other words – More sophisticated versions of the model can be used
(heteroscedastic or quadratic discriminant analysis)
– Variables are selected on the basis of theoretical – Require estimate of larger n. parameters if too few
reasoning data points, estimates will be imprecise offset
• Stepwise method advantages
– Variables are selected from a list of “candidates”, based – Also, functions based on quadratic discriminant analysis
on their discriminating capacity are less readable and hence less useful in practical terms
– Including all variables and subsequently removing those 2) Formulas which convert the score to a PD assume
with lower discriminating power backward elimination multivariate normal distribution for independent variables
– Including a single variable and progressively adding - Unrealistic many economic/financial indicators are
those which most improve the discriminating power structurally limited between 0 and 100%
forward selection - Banks prefer to calculate PDs with different methods,
• Stepwise methods are very powerful but can lead e.g. actual default rates associated with given score
to the inclusion of variables whose economic ranges
significance is unclear to be used under the
supervision of an expert
Discriminant analysis: Regression models:
summing up the idea behind the technicality
• Customers are divided into two (or more) groups • The PD is a function of a company’s
that are supposed to have different means.
features:
• Such groups are described based on past data, so
that new data can be assigned to one of them
• DA is a descriptive tool: no model for the process
leading to default is present
• The assumption of an equal variance/covariance
Ratios
matrix across groups is not totally realistic, PD
although it helps keeping things simple
• Probability estimates require “multi-normality”,
which is seldom supported by empirical data function 1.3%
Central
Credit Register
Internal
Procedures
Linear probabilistic model Linear probabilistic model

• Variables that lead to the default of a company, and • Major drawback: y (PD) may take on values
their weights, are identified with a simple linear
regression outside the 0-100% range
1. Sample selection sample formed by large n. of
companies is selected and divided into two groups, • If the value y generated by the model (based
identified by a binary state variable (dummy) y, which on the customer’s xi,js) is above 100% or
only takes value 0 or 1 (healthy vs default) below 0%, it is usually truncated at those
2. Selection of independent variables. For each company
i, m significant variables (xi1, xi, … xij, … xim) are limits
recorded economic/financial indicators, measured
prior to default • Further problems the variance of the
3. Estimating coefficients: OLS m residuals of the model linear is not constant,
yi j xi , j i but suffers from heteroscedasticity
j 1
imprecise coefficient estimates linear form
4. Estimate probability of default model used to
estimate PDof new companies applying for bank loans never used in practice
• non-linear functions are preferred, as in
the probit and logit models
Logit and probit models: Logit and probit models:
the idea behind the technicality the idea behind the technicality
• In the logit model, the linear relationship is adjusted

• Not just a simple function: through an exponential transformation (logistic)
– We need one generating values between 1
0% and 100% yi f ( wi )
•
1 e wi
where the independent variable wi is given by the linear
• It’s then useful to use: function of the financial indicators xij already seen
– A standard normal probability function m
• PROBIT wi j xi , j
j 1
– A logistic function
• Combining the two, and adding the usual random
• LOGIT disturbance term, we obtain the logit model
1
yi i
jxj
1 e j
Example: logit with 3 variables Logit & probit

• Transformations alternative to the logistic

one can be used, provided that their
codomain is always between 0 and 1
Probability • For example, a normal cumulative density
of default function could be used final model is
x1 called “normit”, or more commonly “probit”
• Main difference between logit and probit
1 lies in the fact that the logistic function has
x2 P 1 x1 21 x 2 3 x3
1.3% “fatter” tails; in practice, this does not
1 e produce any significant differences
between the two models unless the sample
x3 includes numerous observations with
extreme values of wi
The problem is to estimate alfa and beta,
which can be done with maximum likelihood
Page: 320
Logit & probit,
Inductive models
what are they used for
• Credit-scoring models attempt to identify the

• The PD can then be used… fundamental relationships which explain the
– …to decide whether to lend or not, setting a economic/financial health of a company
structural characteristics of a company choice
cutoff point (or a grey area) of relevant variables reflects some a priori choice
– As an indicative “rating” of the uality of the based on economic reasoning
• Inductive models use a purely inductive process:
borrower if, starting from a data sample, a certain
empirical regularity is found (e.g. that many
• Similarities with the DA “score” abnormal companies present values in variable xj
– Under certain conditions, logit and DA above some cut-off point k), this regularity is
used, in an uncritical, “agnostic” way, to forecast
coincide future defaults by other companies a purely
empirical approach is used
• Inductive models are often “black boxes”, which
can be used to generate results rapidly, but
whose logic may not be fully understood.
Neural networks: Neural networks:

Data on borrowers get analyzed with many The individual functions need to be fitted
different functions which are interdependent “training the network”
Judgement on Judgment on
creditworthiness creditworthiness
F(n1,n2)
Neural networks: Genetic algorithms:
• The network is trained through an • Like neural networks, they require an

iterative process on a sample by iterative process
comparing the judgments produced by • Replicate the logic of biology:
the network itself and the actual state – Charles Darwin: survival of the fittest
of the sample companies (default or
• In each iteraction a finite number of
not)
algorithm is tested
• There is an overlearning risk, i.e. the – These are judged based on the
networks learns the sample but looses correspondence between the generated
the ability to generalize results and the true values of the sample
• For this reason one generally keeps a – The worst get eliminated
portion of the sample for testing – The best are then recombined with each
purposes other
Genetic algorithms: Genetic algorithms:

• A new generation of algorithms is

generated, which likely to be better than Generation of Valuation Does a
the first ones The initial of the satisfactory Solution
population solutions solution exist? yes
• One goes on from generation to
no
generation, adding “cromosomes” (small
mutations) randomly (piccole mutazioni)
which are not present in their parents ? Generation
Selection
• When the new generations do not add of a
Source: Pomante (1999)

Problem
population Recombination
value to the algorithm ability with respect
to the previous ones, the entire process is Mutation
stopped
Page: 329
Outline of this session
• Models based on market data

Capital Market Approaches – General features
for the Estimation of PD – Bond spreads and risk-neutral PDs
– Share prices: Merton’s model
– Share prices: Moody’s KMV
Courtesy of Andrea Sironi Creditmonitor
Using bond spreads Corporate bond spreads: let’s

and share prices start with simple compounding
– For some (large) corporate borrowers, some

specific information sources can be used 1 r 1 p 1 r* 1 p 1 r d
• The bond market (spreads) r = risk-free rate (government bond)
• The stock market (prices and volatilities)
p = probability of default
– Such sources
r* = return of the risky bond (corporate bond)
• Can be handled through “quantitative” models
that embed all market opinions on all aspects d = spread (corporate bond return – risk free rate)
(also on qualitative ones) of the firm’s Hp: zero recovery (k=0)
management and perspectives
• Are forward looking (while this does not apply to
ratios based on financial statement data, and Example:
even to CCR indicators) 1 r 1.04
• Are prone to the same “illusions” and “bubbles” p 1 p 1 0.95%
that all capital markets usually suffer from 1 r d 1.05
Corporate bond spreads Corporate bond spreads
Equilibrium spread given a PD Hp: p=1%

If recovery rate (k) >0
If k=0 p1 r
1 r 1 p 1 r d p k1 r d d
1 p
r d r d 1.04
p dk 0 0.01 1.0505%
1 r d 1 k 1 r d 1 k 0.99
Example: r=4%, d=1%, k=50% If k>0 d

1 r
1 r
1 p1 k
1%
p 1.9% 1.04
1 5% 1 0.5 dk 50% 1.04 0.5225%
1 0.04 1 0.5

Let’s consider these data

From the previous spot rates we can derive
forward rates
Maturity Risk Free Rates Risky Rates Spreads
Maturity Risk free rates Risky rates Spread
1 4.00% 5.00% 1.00% 1 4,00% 5,00% 1,00%
2 4.10% 5.20% 1.10% 2 4,20% 5,40% 1,20%
3 4,40% 6,10% 1,70%
3 4.20% 5.50% 1.30%
4 4,60% 6,71% 2,11%
4 4.30% 5.80% 1.50% 5 5,30% 7,82% 2,52%
5 4.50% 6.20% 1.70%
© M. Anolli - Risk Management

Page: 336 © M. Anolli - Risk Management
Page: 337
Forward rates Spreads
9,00% 3,00%
8,00% 2,50%
7,00% 2,00%
Risk free rates Spread Spot
6,00% 1,50%
Risky rates Spread Forward
5,00% 1,00%
4,00% 0,50%
3,00% 0,00%
1 2 3 4 5 1 2 3 4 5
Higher forward spreads positive slope of the spread

curve expectations of increasing spreads

Using forward spreads one can estimate the Using the probability of default of the first
marginal probabilities of default of different year (1.9%) and the one of the second year
future time periods (2.28%), it is possible to compute the
For example, it is possible to estimate the cumulative probability of default for the two
marginal probability of default of the second year period
year One first has to estimate the marginal
survival probability 1-p
d 5.4% 4.2% 0.012

1 p2 2.28% s 1 p
1 r d 1 k 1 5.4% 1 50% 1.054 0.5

Page: 341
T T
Marginal survival probabilities Generalizing 0 sT t st 1 0 pT 1 t st 1
t 0 t 0
0 1 s 1 0 p1 1 0.019 98.1%
s
1 2 1 1 p2 1 0.0228 97.72%
Marginal Marginal Cumulative Cumulative
Probability of Survival Survival Default
Year Default Probability Probability Probability
0 2s s s
0 11 2 98.1% 97.72% 95.86% 1 1,90% 98,10% 98,10% 1,90%
2 2,28% 97,72% 95,86% 4,14%
3 3,21% 96,79% 92,79% 7,21%
4 3,94% 96,06% 89,13% 10,87%
0 p2 1 0 s2 1 95.86% 4.14% 5 4,66% 95,34% 84,97% 15,03%

Page: 343
Corporate bond spreads Using corporate bond data

Advantages • We can also use continuous rates. This means

Uses market data objective that, e.g., for a one-year bond:
– 91 (1+9.89%) = 100 becomes 91e9.43% = 100
“forward looking” expected market – To switch from yearly compounded rates to
default rates continuous rates, we simply compute:
9.43% = ln(1+9.89%)
Weaknesses • Then, assuming k=0
Assumption expectation theory is true:
spreads reflects market expectations er 1 p er d
p 1 e d
Spreads may also reflect other factors:

liquidity premia, transaction costs – Advantage of continuous compounding: the
PD implied in bond rates does not depend
Assumption of risk neutral investors on the level of the rates (i.e. on r and r*),
Only applicable to companies with bonds but solely on the spread between them
outstanding in the capital markets
Page: 344
Using corporate bond data Using corporate bond data
• If k>o, we get • When we look at longer time horizons

(T years)
er 1 p pk e r d
1 p1 k er d
e rT T 1 pT pT k e rT dT T 1 pT 1 k e rT dT T
1 e dT T
1 e d pT
p 1 k
1 k
• If d is equal to 1% and k is 50%, we • This gives us the cumulative PDs associated
get: with the various maturities
0.01
1 e • From cumulative PDs we can compute
p 1.99% cumulative and marginal survival rates
1 0.5 marginal PDs
Historical data and bond

prices According to Standard & Poor’s:
• As an alternative source to estimate
349
Page:
1 2 3 4 5 7 10
default probabilities one may also use
historical data provided by rating AAA 0,00 0,00 0,07 0,15 0,24 0,66 1,40
agencies AAA 0,00 0,02 0,12 0,25 0,43 0,89 1,29
• By doing so, one would find out that PD A 0,06 0,16 0,27 0,44 0,67 1,12 2,17

estimates based on the bond spread BBB 0,18 0,44 0,72 1,27 1,78 2,99 4,34
model tend to be significantly and BB 1,06 3,48 6,12 8,68 10,97 14,46 17,73
systematically higher than historical B 5,20 11,00 15,95 19,40 21,88 25,14 29,02
ones
CCC 19,79 26,92 31,63 35,97 40,15 42,64 45,10
• Let us see an example: A-rated bonds…

According to bond spreads: Risk-neutral probabilities
• A-rated bonds usually paid (before the crisis) a • Spread-based PDs are risk-neutral probabilities. In fact, we
spread of about 50 basis points over 5-year are saying that
Treasuries
• Hence, the present value of the loss expected over (1) It is identically the same to get 100
the next 5 years is 2.47% (= 1 - e-0.005×5) of the risky (with a spread of 0,50%) dollars
face value or 97.529 riskless (no spread) dollars.
• Expected loss is given by PD*LGD, that is, PD*(1-

RR)
• If we assume that RR is 30% (close to empirical 97.529e r 5 100e ( r 0.5) 5
100
values recorded on distressed bond markets), then
the PD implied by a 2.47% is about 3.53% [= 100
2.47%/(1 - 30%)] e0.5 5 E(.)=97.529
• This is more than 5 times higher than historical PD 97.529
(0.67%) 1 100
• Why? 0.5 ln 30
– Liquidity considerations? Low cyclical 5 97.529
expectations? Or… (2) What is 97.529 dollars? The expected value
from a lottery consisting in buying the risky bond.
And we are treating them as a certain value (asking
for no spread on them)
Page: 351
Risk-neutral probabilities Outline of this session

• We assumed that the buyer of a bond was indifferent to receiving

the (certain) amount for a treasury bond or the (risky) amount • Models based on market data
for a corporate bond, as long as the expected value for the two
was the same – General features
• However, in order to trade a risk-free investment for a risky one,
investors demand a premium R – Bond spreads and risk-neutral PDs
• er R 1 p* 1 k er d
– Share prices: Merton’s model
• The value of p that solves the equation without R is higher than – Share prices: Moody’s KMV Creditmonitor
that value p* that solves this equation
• Default probabilities based on historical data are “real-world”
probabilities
• In the real world, better expectations on default risk are offset by
the fact that investors require a risk-premium to take part in a
lottery
• Never forget that “risk-free” and “real-world” (risk-averted)
probabilities are different but may lead to the same price
• Risk neutral PDs may be used in a number of asset-pricing

models (e.g. to estimate the fair value of a credit default swap)
Merton’s model Merton’s model
• Merton’s model uses stock prices as input and

seeks to determine the equilibrium bond spread, in The value of a borrower’s assets follows a geometric brownian
addition to estimating PD motion
• Original model has been subsequently enhanced in
order to eliminate certain assumptions and to dV
make it easier to apply in practice
• Simple intuition: a company defaults when the
dt v dz dt v dt
value of its assets becomes lower than the value of V
its liabilities
• Shareholders have the option of defaulting and is the expected instantaneous yield on the assets and dz (equal
giving the company to the creditors, when the
value of liabilities is greater than the value of the to the product of a standard normally-distributed term e and the
company’s assets square root of time) is a random disturbance
• Merton’s model and the variants discussed below
are generally referred to as structural models Practically speaking, the “asset returns” evolve stochastically,
structural traits of a company that determine its and the uncertainty as to their future path increases with the
PD, i.e. the value of assets, the value of debt
(leverage), and the volatility of asset values time horizon
Page: 355
Merton’s model Merton Models in a nutshell / 2

Past evolution Possible future evolutions Probability

Credit risk concerns the possibility that the value of the distribution of
all possible
company’s assets, V , will be less than the future values
T
Log of the assets’ value

repayment value of the loan, F, as of the maturity of
the debt (T). This possibility increases as the
following increase
Default
1. the ratio B0/V0, i.e. the company’s leverage at T=0; probability
Value of debt (log) p
2. the volatility of the company’s asset yield, measured as the
standard deviation of the asset yield V;
3. the debt’s maturity
Today In one year’s time
Merton’s model Merton model
• A company’s PD can be expressed as the

probability that VT < F (area under the normal • Shareholders have the option of handing over
distribution)
• All other things being equal, this area increases as:
their company to creditors rather than
– the beginning market value of assets (V0) decreases repaying the company’s debt they can trade
– the nominal value of the debt (F) increases VT for F when the former is lower than the
– the volatility of the market value of assets increases latter
– the debt’s maturity increases
• This can be considered as a put option (i.e.
• Given T, the first 3 variables represent all relevant
information: the option to sell at an agreed upon price) that
– the outlook for the company, its industry, and the the company’s lenders (e.g. a bank) have
economy, which affects the company’s expected granted the shareholders put option on the
future cash flows, the PV of which represents the
current market value of its assets (V0); value of the company’s assets, where the
– the company’s financial risk i.e. leverage; strike price is equal to F, and maturity T
– Its business risk, i.e. the volatility of the asset
returns
Merton model Merton model

Figure 4 – The payoff profile of a lender bank

• In order to hedge the credit risk of the loan,
the bank could, in turn, acquire a long put
option on the value of the company’s assets
Payoff to the Debtholders
(V) with the same maturity as the loan (T) and

a strike price equal to the debt repayment (F)
Payoff at time 0 and T on a loan and a long put option
Payoff at time 0 Payoff at T
if VT<F if VT>F
Loan -B0 VT F
Put option -P0 F-VT 0
Total -(B0+P0) F F
V1 F V2 Asset Value (VT)
rT
P0 B0 Fe
Merton model Merton model: the value of debt
r *T Market value of debt

P0 N ( d1 )V0 Fe N ( d2 )
rT
P0 B0 Fe
One can estimate rT
P0 N ( d1 )V0 Fe N ( d2 )
The value of debt
The equilibrium spread (risk premium)
The borrowing firm PD rT rT 1
B0 Fe 1 N ( d2 ) N ( d1 )V0 Fe N (d 2 ) N ( d1 )
L
The value of debt is larger the lower is leverage and the lower
is debt maturity
Merton model: the equilibrium spread Merton model: PD

Spread p Pr VT F
1 1
r* r d ln N (d 2 ) N ( d1 )
T L p Pr VT F 1 N (d 2 ) N ( d2 )
Risk premiums at various levels of leverage and asset yield volatility (T=1; r=5%)
5% 10% 15% 20% 25% 30%
Example
A
L
50% 0.000% 0.000% 0.000% 0.002% 0.029% 0.149%
L=85.61% (V=100,000; F=90,000)
60% 0.000% 0.000% 0.002% 0.044% 0.243% 0.700% = 10%
70% 0.000% 0.001% 0.052% 0.355% 1.032% 2.063% r = 5%
80% 0.000% 0.050% 0.506% 1.494% 2.873% 4.519%
90% 0.033% 0.795% 2.272% 4.070% 6.036% 8.112%
100% 2.015% 4.069% 6.165% 8.301% 10.478% 12.696%
p Pr VT F 1 N (d 2 ) N ( d2 ) 6.63%
Term structures of the credit spreads (d)
PD higher: for two different types of borrower
(a) The lower the value of assets and the higher the value of debt (leverage)
financial risk 4,5%
(b) The higher the asset volatility business risk
4,0%
3,5%
Probability of default (p) and risk premium (d) by maturity
Maturity L = 90%; Sigma ( 3,0%
V) = 20% L = 75%; Sigma ( V) =10%
(years)
Spreads
p (cumulative d (spread) p (cumulative PD) d (spread) 2,5% Low quality (high PD)
PD)
2,0% High quality (low PD)
1 33.48% 4.07% 0.24% 0.01%
2 40.86% 3.69% 2.48% 0.06%
1,5%
3 44.79% 3.37% 5.77% 0.13%
4 47.47% 3.12% 9.04% 0.19% 1,0%
5 49.52% 2.93% 12.00% 0.24% 0,5%

6 51.19% 2.77% 14.64% 0.28%
7 52.61% 2.64% 16.98% 0.31% 0,0%
1 2 3 4 5 6 7 8 9 10
8 53.85% 2.53% 19.06% 0.33%
9 54.95% 2.44% 20.93% 0.35% Maturity (years)
10 55.95% 2.36% 22.61% 0.36%

Marginal default probabilities - Moody's

Merton model 1970-2005
Longer maturities lead to declining risk premiums

when the probability of default is high
This is due to the fact that, for poor credit quality
companies there is a significant risk that they will
not “survive” the first year. However, if they do
get past the first year, the probability of becoming
insolvent in subsequent years declines significantly
While a high asset volatility does cause many
companies to default during the first year, it also https://www.moodys.com/sites/products/DefaultResearch/2006200000425249.pdf
leads an equally significant number of borrowers to
considerably reduce their leverage, which, in turn,
reduces the probability of default in the coming
years
Model advantages Model drawbacks

Just one debt (zero-coupon)
It clearly identifies the key relevant
variables that determine PD: (i) leverage Only looks at default risk (no migration risk)
financial risk; (ii) asset volatility Some key variables cannot be empirically
observed (asset value and volatility)
(business risk)
Constant risk free rate
It allows to estimate, through objective Arbitrage-free logic possibility to buy and
and clear criteria, the borrowing sell the option’s underlying asset
company PD and the risk premium a (company’s assets)
lender should demand for the credit risk
it faces
Merton models and recovery Merton Model:

rates: some drawbacks
parameter estimation
• The present value of total assets and their

• Liquidation costs are not accounted for volatility cannot be measured empirically
– The bank cashes in the whole value of the • However, shares can be seen as call options
corporate assets written on a firm’s total assets
• The normal distribution underestimates • Within a simplified framework, the value of
extreme events shares when debt must be paid back, is
– Making it very unlikely that the firm’s value max(V’T - Ft, 0)
may drop significantly below the value of • where V’T is the final value of corporate assets
debt and Ft is the face value of the debt
• However, the line of reasoning underlying • This is the payout profile of a call option,
those models is conceptually correct which can be thought to be a European one,
for the sake of simplicity
Merton Model: The value of equity
and the value of assets
Since equity capital can be • The current value of equity, E, can be linked
seen as a call on total assets to the current value of corporate assets (V)
(having strike price equal to and their volatility (σV) by means of an options
Value the debt) VT and V can be pricing formula, such as the Black-Scholes
of equity worked out from the model:
(call) behavior (value and
volatility) of equity (which
rT
is empirically observable).
E0 V0 N ( d1 ) Fe N (d 2 ) 1
Value of Assets where, as usual:

(underlying)
2
Ft
ln V0 / F r V /2 T
d1 and d 2 d1 V T
However, this can be done only for listed companies. V T
The volatility of equity Merton Model:

and the volatility of assets
• Two equations and two unknowns the

• Moreover, since the value of equity E(Vt) is a unknowns appear multiple times in both equations
function of the asset value and of time, Ito’s the system must be solved iteratively: we
lemma can be applied to derive a mathematical choose two initial estimates of the unknowns and
link between the volatility of equity and that of iteratively change both V0 and sV until the two
total assets: equations generate the values of E0 and E that
have been observed empirically on the stock
V0 market
E N ( d1 ) V
2
E0
where, as usual: VA
rT
V 2 E E0 V0 N d1 Fe N d2 1
ln r V
T V0
E F 2 N d1 V
Δ N(d1 ) and d1
A
E 2
E
E0
V V T
Structural-form models Merton Models:
versus reduced-form models
children and grandchildren
• Merton’s Model is a structural one: the process • Merton’s intuition has lead to
driving to default is endogenous to the model a number of subsequent results:
– The value of total assets crosses a lower – KMV
bound • Kealhofer, McQuown & Vasicek
• However, several alternative models have – Creditmetrics
adopted a reduced form: the process driving
– Basel 2 (the uni-factorial, infinitely granular
to default is exogenous, not included into the
model proposed e.g., by Finger or Gordy)
model. However, it can be “hinted at” through
the bond prices observed on the capital –…
markets
– E.g.: Duffie and Singleton, Das and Tufano,
Jarrow, Lando and Turnbull
Moody KMV’s Moody’s KMV

• Three aspects of the standard Merton Probability

Possible future evolutions
model are refined distribution of
– Not all corporate debt is thought to be due all possible
at the end of the first year: F is replaced by future values
Distance to
the so-called default point DP default, expressed
in units of
Log of the asset value

– Since a normal distribution leads to standard
unsatisfactory empirical results, (e.g., deviations
downward biased PDs, upward biased RRs)
it is replaced by an empirical, non- Default
probability
parametric distribution
DEFAULT POINT (log)
– The link between (volatile) market data and
PD estimates is made steadier by means of
a “sticky” mechanism
Today In 1 year’s time
Moody KMV’s Moody KMV’s
• KMV uses the default point (DP), equal to all short-

term debt (STD) plus 50% of long-term debt (LTD) DD (a) (b) (c ) = (b) / (a)
(approximate # of # of defaulted Default frequency
1 value) companies companies
DP STD LTD
2 1 9000 720 8%
2 15000 450 3%
• Distance to default is equal to the difference 3 20000 200 1%
between asset value and the default point,
expressed as a multiple of the standard deviation 4 35000 150 0,4%
of assets 5 40000 28 0,07%
V0 DP 6 42000 17 0,04%
DD
V0 V
• The second phase in KMV’s PD calculation Empirical data suggest a fairly precise empirical correlation
procedure is based on the empirical link between
DD and actual past rates of default for various between DD and past default frequencies. Once a company’s DD
DD ranges, empirical estimates of percentage of
defaulted companies is known, this correlation can be used to calculate the associated
PD (expected default frequency)
Moody KMV Enron Credit Risk Measures
Facoltà di Scienze Bancarie, Finanziarie e Assicurative Facoltà di Scienze Bancarie, Finanziarie e Assicurative EDF Equivalent
DD = 1
DD = 1 DD = 2 DD = 3 DD = 4 DD = 5 DD = 6
9.000 comp.s 15.000 comp.s 30.000 comp.s 35.000 comp.s 40.000 comp.s 42.000 comp.s Rating
720 defaults 450 defaults 200 defaults 150 defaults 20 defaults 17 defaults
CC
9.000 companies
p = 8% p = 3% P = 1% P = 0,43% p = 0,07% p = 0,04 bp
CCC
720 defaults
p = 8% B
p (expected default frequency)
10,00%
BB
1,00% BBB
0,10% AA
AAA
0,01%
0 1 2 3 4 5 6
DD (distance to default)
Worldcom Credit Risk Measures
KMV: the advantages
Z" SCORES AND EDF'S FOR WORLDCOM Facoltà di Scienze Bancarie, Finanziarie e Assicurative
(Q4'1999 - Q1'2002)
7.00 0.01
• Forward-looking model based on market data
EDFs adapt quickly to the changing financial
AAA
S&P Rating A-
BBB
6.00 BBB conditions of the companies being evaluated as
AA opposed to agency ratings, which tend to react
with a significant delay
BEQ*
0.10
BB
5.00 Z" Scores BB-
B+ B+ • Does not require the borrower to issue any listed

BBB
(and liquid) corporate bonds
EDF Score
4.00
• Measures PDs (EDFs) on a single-borrower basis,

Z" Score
EDF BB
CCC+
while historical rating-agency data have to pool

1.00
together issuers of different (although similar)

3.00
CCC-
quality
2.00
CC 10.00 • EDFs do not swing significantly as economic cycles
change, as opposed to the empirical default rates
CCC
D
1.00
associated with agency ratings it is a point-in-
time rating, so it is closer to the banks’ point of
0.00
D
100.00 view than agency ratings are
– Problem: can its high sensitivity to economic cycles
Q4'99 Q1'00 Q2'00 Q3'00 Q4'00 Q1'01 Q2'01 Q3'01 Q4'01 Q1'02 Q2'02 Q3'02
Quarter- Year
Z" UnAdj Z" Adj:3.85B Z" Adj:7.2B&50B EDF
raise any unwanted effect (e.g., procyclicality of the
rating system)?
*BEQ = Z" Score Bond Equivalent Rating
Sources: Compilation by the author (E. Altman, NYU Stern), the KMV
(Moody's) Website and Standard & Poor's Corporation.
Page: 386
KMV: its drawbacks

KMV: the advantages
for continental Europe
Facoltà di Scienze Bancarie, FinanziarieTable
e Assicurative
9 – One-year transition matrix – Standard & Poor’s Facoltà di Scienze Bancarie, Finanziarie e Assicurative
Rating at year-end (%)
Initial AAA AA A BBB BB B CCC Default • It requires a lot of market data theoretically
rating
it can only be applied to listed companies
AAA 90.81 8.33 0.68 0.06 0.12 0.00 0.00 0.00
AA 0.70 90.65 7.79 0.64 0.06 0.14 0.02 0.00 – A private firm model exists, where a peer group
A 0.09 2.27 91.05 5.52 0.74 0.26 0.01 0.06 of listed company is required
BBB 0.02 0.33 5.95 86.93 5.30 1.17 1.12 0.18 • It will never be suited for firm clusters where
BB
B
0.03
0.00
0.14
0.11
0.67
0.24
7.73
0.43
80.53
6.48
8.84
83.46
1.00
4.07
1.06
5.20
no listed peer group is available
CCC 0.22 0.00 0.22 1.30 2.38 11.24 64.86 19.79 – Medium enterprises, small businesses
Source: Standard & Poor’s CreditWeek (1996).
• Models based on the contingent claims
Table 10 – One-year transition matrix – KMV
approach are based on the assumption that
Initial AAA AA
Rating at year-end (%)
A BBB BB B CCC Default
equity markets are informationally-efficient
rating suffers from the markets’ irrationality and
AAA 66.26 22.22 7.37 2.45 0.86 0.67 0.14 0.02
nervousness
AA 21.66 43.04 25.83 6.56 1.99 0.68 0.20 0.04
A 2.76 20.34 44.19 22.94 7.42 1.97 0.28 0.10 – One might argue that the problem is with the
BBB 0.30 2.80 22.63 42.54 23.52 6.95 1.00 0.26 capital markets, not with the model!
BB 0.08 0.24 3.69 22.93 44.41 24.53 3.41 0.71 – Reactivity is a “two-hedged sword”
B 0.01 0.05 0.39 3.48 20.47 53.00 20.58 2.01
CCC 0.00 0.01 0.09 0.26 1.79 17.77 69.94 10.13
Creditmonitor and the
Private Firm model
Data taken from
the financial statements
of listed firms
EDFs in
Asset Value
the
EBITDA
recent
Size financial
DD Asset Volatility
crisis
Default Point
EDF
Link
DD/EDF Data taken from
financial statements
Page: 390
Agenda
Portfolio models
– the main concepts behind credit VaR
Credit Portfolio Models – Non normality of credit risk
– Creditmetrics
• A multinomial approach and “mark-to-model”
• The risk of a portfolio
– CreditRisk+
• The analogy with insurance companies and the
Courtesy of Andrea Sironi mathematical framework
• The issue of correlations

Credit VaR models: common
Portfolio models
goals
Main portfolio models • To estimate the probability distribution

of all possible future losses
• CreditMetrics (J.P. Morgan) TM
• To use that distribution to isolate a

• CreditRisk+ (CSFP) TM
measure of Value at Risk associated

• Portfolio Manager (KMV) TM
with a given confidence level

• CreditPortfolioView (McKinsey) TM
• To estimate a measure of
– Expected loss
• We’ll focus on the first two – Unexpected loss
For a portfolio of loans, the normal

distribution cannot be reasonable: In the case of loans, moreover:
396
397
Page:
Page:
2 • Returns are not 2 • Since the default
Same Vera
True normal, and Vera
True is a rare event, no
and Normal
Normale significantly skewed 1.5 Normal
Normale historical
1.5
databases exist to
to the left
measure the

1 • They reflect: 1 correlation
– Limited earnings between a given
0.5 that are highly likely 0.5 pair of borrowers.
– A limited probability
0 of huge losses 0
-150% -100% -50% 0% 50% -150% -100% -50% 0% 50%
Credit VaR models: Credit VaR models:
time horizon the time horizon
Objective Key factors Ideal horizon

• Time horizon Risk measurement and
control
Effective liquidity of positions and
holding period of the bank
Residual life of
exposures
– Can be set to just one common value for all
loans (e.g., 1 year): Simplification Consistency with the time horizon 1 year
• This makes it easier to collect, file and use the used in estimating PD and other risk
parameters (e.g., PD, EAD, correlations, parameters
transitions) needed to feed credit VaR models Measurement of the risk- Frequency of the budgeting process 1 year
• This is the time period needed to raise new adjusted performance Frequency of the reporting process
capital when the bank experiences high losses (RAP) of the various bank Capital allocation 1 year
units
– Alternatively, it can be set – for each asset – to 1 year
its true liquidation horizon: Consistency between risk Time necessary to collect new 1 year
and capital capital
• This means that a 5-year loan for which no
secondary market is available has a risk horizon Implementing corrective Average portfolio turnover period 1 year
of 5 years, not of 12 months actions on the portfolio
Pricing Maturity of exposures Residual life of
exposures
Frequency of rate revisions 1 year
Credit VaR models: Credit VaR models:

definition of loss
correlations
• Definition of loss: • Correlations among loans

– Only default (“default mode paradigm”) – Can be modeled explicitly
• Credit risk can then be modeled by means of • The values of firm A and firm B change together
binomial models • This means that their defaults tend to occurr
• Loans can be kept at “book value” together and their PDs tend to change together
• BUT: an important source of losses can be – Can be modeled implicitly
overlooked, above all if the model adopts a short • For every state of the economy, two loans can be
risk horizon (e.g., one year) thought of as independent
– Any change in value (“mark-to-market”) • However, their PDs depend on the state of the
• Rather, it is a “mark-to-model” paradigm economy
• Two future states are not enough – For example, a recession increases the PDs of both
loans
• Loans must be evaluated based on their market
value – The uncertainty on the future state of the
economy, therefore, makes two loans correlated
Creditmetrics: default is just one
Creditmetrics among a set of credit events
• The multinomial approach

– the role of ratings and transition matrices • Credit events = all events that can alter the value
– Wilson’s critique and CreditPortfolioView of a loan
– spread curves and mark-to-model • The default is only the most significant one:
• The “present value in future” – The new value of the loan is equal to the value that
can be reasonably recovered (e.g. il 30%)
– Expected value, percentiles, VaR
– The normal distribution is a false friend
• But all changes in the borrower’s rating are credit
events, in that they change the fair value of the
• The case of 2 or more loans
loan
– Joint transitions, VaR, diversification effects
– The future cash flows from downgraded loans have
– Estimating joint transitions: multinomial Merton to be discounted using higher rates (higher
• Montecarlo simulations spreads) and their present value decreases
CreditMetrics™ CreditMetrics™
6 steps Model inputs

1. Market value of exposures Time horizon
2. Estimate migration probabilities Rating system (S&P, Moodys, internal)
3. Estimate recovery rate Transition matrix
4. Market values associated to different
rating grades Recovery rates
5. Distribution of market values at the Forward spreads associated to
end of the year different rating grades
6. Portfolio risk
Second step: estimate migration Third step: recovery rates

probabilities
One-year transition matrix
RATING AT YEAR-END (%)

INITIAL RATING AAA AA A BBB BB B CCC Default
Category Senior Senior Senior Subordinated Junior
AAA 90.81 8.33 0.68 0.06 0.12 0 0 0.00
Secured Unsecured Subordinated Subordinated
AA 0.70 90.65 7.79 0.64 0.06 0.14 0.02 0.00
Mean 53.80 51.13 38.52 32.74 17.09
A 0.09 2.27 91.05 5.52 0.74 0.26 0.01 0.06
Std. dev. (%) 26.86 25.45 23.81 20.18 10.90
BBB 0.02 0.33 5.95 86.93 5.30 1.17 0.12 0.18 Source: Gupton, Finger and Bhatia (1997).
BB 0.03 0.14 0.67 7.73 80.53 8.84 1.00 1.06
B 0.00 0.11 0.24 0.43 6.48 83.46 4.07 5.20
CCC 0.22 0.00 0.22 1.30 2.38 11.24 64.86 19.79
Source: S&P CreditWeek (15 April 1996)
Page: 407
Fourth step: market values associated to Example: BBB bond, 5 years, 6% fixed
coupon
different rating grades
Remains BBB (probability = 86.93%)
One-year forward zero coupon rate curve (%)
6 6 6 106
Maturity 1 year 2 years 3 years 4 years FV1,BBB 6 107.53
Rating class
(1 4.10%) (1 4.67%) 2 (1 5.25%)3 (1 5.63%) 4
AAA 3.60 4.17 4.73 5.12 If downgraded to BB
AA 3.65 4.22 4.78 5.17
A 3.72 4.32 4.93 5.32 6 6 6 106
FV1, BB 6 102.01
BBB 4.10 4.67 5.25 5.63 (1 5.55%) (1 6.02%) 2 (1 6.78%)3 (1 7.27%) 4
BB 5.55 6.02 6.78 7.27
B 6.05 7.02 8.03 8.52
Loss 5.52 = 107.53-102.01
CCC 15.05 15.02 14.03 13.52
Page: 409
Fifth step: distribution of market value changes Vj Probability, Cumulative

We can compute the value at pj (%) probability (%)
risk associated with a certain ci pj
Distribution of one-year market values of a BBB bond
confidence level, by V j Vi
State at year-end (j) Present value Probability, pj Vj = -53.27 0.18 0.18

in a year's time (%) FVj – E(FV) “cutting” the distribution of -23.45 0.12 0.3
(FVj) value changes at the desired -8.99 1.17 1.47
AAA 109.35 0.02 2.28 percentile -5.07 5.3 6.77
AA 109.17 0.33 2.10 0.46 86.93 93.7
VaR 99% = 8.99 1.57 5.95 99.65
A 108.64 5.95 1.57
VaR 95% = 5.07 2.1 0.33 99.98
BBB 107.53 86.93 0.46
2.28 0.02 100
BB 102.01 5.3 -5.07
B 98.09 1.17 -8.99
If we had used a parametric
CCC 83.63 0.12 -23.45 approach based on the normal
VaR99% 2.32 2.9 6.75
Default 53.80 0.18 -53.27 distribution, we would have found VaR 95 % 1 . 64 2 . 9 4 . 77
Mean, E(FV)= pjFVj 107.07 quite different values
The distribution of forward values

Sixth step: portfolio VaR
Example: 2 independent bonds with rating
A and BB
Probability they both remain in their initial
grating grade 80.53% x 91.05% =
73.32%
Probability they both default 0.06% x
1.06% = 0.00%
Proceeding this way, one can construct a
joint transition matrix
Problem: in reality migrations are not
independent
Page: 413
Probability of joint migration of two issuers with ratings A and BB, assuming • Assumption of independence not realistic
independence of the relative migration rates rating changes and defaults are partly the
result of common factors (e.g. economic
Issuer A
cycle, interest rates, changes in commodity
AAA AA A BBB BB B CCC Default
prices, etc.)
Issuer BB 0.09 2.27 91.05 5.52 0.74 0.26 0.01 0.06
AAA 0.03 0.00 0.00 0.03 0.00 0.00 0.00 0.00 0.00 • CreditMetrics™
AA 0.14 0.00 0.00 0.13 0.01 0.00 0.00 0.00 0.00 – uses a modified version of the Merton model,
A 0.67 0.00 0.02 0.61 0.40 0.00 0.00 0.00 0.00 where not only defaults but also migrations
BBB 7.73 0.01 0.18 7.04 0.43 0.06 0.02 0.00 0.00 depend on changes in the value of corporate
BB 80.53 0.07 1.83 73.32 4.45 0.60 0.20 0.01 0.05
assets (asset value returns, AVR)
B 8.84 0.01 0.20 8.05 0.49 0.07 0.02 0.00 0.00 – estimates the correlation between the asset
CCC 1.00 0.00 0.02 0.91 0.06 0.01 0.00 0.00 0.00
value returns of the two obligors
Default 1.06 0.00 0.02 0.97 0.06 0.01 0.00 0.00 0.00 – based on that correlation, derives a distribution
of joint probabilities
Page: 415
CreditMetrics™ Example BB
CreditMetrics™
A “multinomial” Merton model with default and migrations

In Merton’s model, the estimate of Zdef involves
0,45 an analysis of corporate debt, the current value
of corporate assets and its volatility
0,4
In Creditmetrics (reduced-form model), all
80,53%
0,35
AVRTs are derived from the probabilities of the
0,3
Z-BB
transition matrix graphically, each transition
Z-B
0,25 probability is equivalent to the area below the
relevant section of the asset distribution
f(x)
0,2 Z-BBB
8,84% 7,73% For example, Zdef is selected such that the
Z-CCC
0,15 0,67
default probability (area under the curve left of
the value Zdef) is the PD of the transition matrix
Z-A
1,00% 0,1 0,14
Z-def
1,06%
0,05
Z-AA
0,03
(1.06%) in the case of a BB bond
-2,30 -2,04 -1,23
0 1,37 2,39 2,93 3,43 Z def
-4 -3 -2 -1 0 1 2 3 4
© M. Anolli - Risk Management f (rBB )drBB F ( Z def ) PD 1.06% Page: 417
Since the probability density function of the AVRs is Probability of migration and associated AVR thresholds
normal, the condition becomes
Z def
for a BB company
f (rBB )drBB F ( Z def ) PD 1.06%
State at Transition Cumulative AVRT
Similarly, ZCCC will be selected such that the area year end probability probability (Zj)
included between Zdef and ZCCC is equal to the (j) (pBB j)
probability of migration from BB to CCC
Default 1.06% 1.06% -2.3
N ( Z def ) 1.06% Zdef = N-1(1.06%) -2.3 CCC 1.00% 2.06% -2.04

B 8.84% 10.90% -1.23
BB 80.53% 91.43% 1.37
Z CCC BBB 7.73% 99.16% 2.39
f (rBB )drBB N ( Z CCC ) N ( Z def ) p BB CCC 1% A 0.67% 99.83% 2.93
Z def
AA 0.14% 99.97% 3.43
AAA 0.03% 100.00%
Page: 419
The same logic can be followed for an A rated If the stand.zed AVRs of the two companies are
company described by a std normal the joint AVR
Migration probabilities and relative thresholds for a BB company
distribution is described by a standardized
bivariate normal x 2 2 xy y 2 A
State at Transition Cumulative AVRT 1 2 (1 2
)
end-year probability probability (Zj) f ( x; y; ) e
2
(j) (pA j) 2 1
Default 0.06% 0.06% -3.24 Its cumulative density function (probability that x
is less than X and, at the same time, y is less than
CCC 0.01% 0.07% -3.19
Y) is given by the following double integral
B 0.26% 0.33% -2.72
Y X x 2 2 xy y 2
BB 0.74% 1.07% -2.30 1 2 (1 2
)
Pr x X; y Y e dxdy N ( X ;Y ; )
BBB 5.52% 6.59% -1.51 2
2 1
A 91.05% 97.64% 1.98
AA 2.27% 99.91% 3.12
both functions depend on the parameter
correlation between the asset value returns
AAA 0.09% 100.00% Page: 420 © M. Anolli - Risk Management
Page: 421
The estimation of the between
CreditMetrics™
the asset values of two borrowers
Joint probabilities of two obligors (with A and BB ratings)

• Creditmetrics using an approach
assuming correlation between asset value returns of 20% - Values % by large building blocks
– Correlations are first estimated among a
Issuer A
Issuer BB AAA AA A BBB BB B CCC Default Total
large set of industries and countries (“risk
AAA 0.00 0.00 0.03 0.00 0.00 0.00 0.00 0.00 0.03 factors”)
AA 0.00 0.01 0.13 0.00 0.00 0.00 0.00 0.00 0.14 – For each borrower, a set of weights must be
A 0.00 0.04 0.61 0.01 0.00 0.00 0.00 0.00 0.67 specified, expressing his sensitivity to
BBB 0.02 0.35 7.10 0.20 0.02 0.01 0.00 0.00 7.69
different risk factors and to idiosyncratic
BB 0.07 1.79 73.65 4.24 0.56 0.18 0.01 0.04 80.53
risk
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 – Combining those weights and the risk factor
Default 0.00 0.01 0.90 0.13 0.02 0.01 0.00 0.00 1.07 correlations, an estimate of the pairwise
Total 0.09 2.29 91.06 5.48 0.75 0.26 0.01 0.06 100.00 correlation of two firms can be obtained
Source: Gupton, Finger and Bhatia (1997).
The estimation of the between The estimation of the between

the asset values of two borrowers the asset values of two borrowers
• In practice: the asset returns are proxied by the Company A B

return on stock indices E.g., Companies A and B Industry / Country
• The AVR of a firm is decomposed into one or USA
1
-
rA - Banking 50%
more systematic components (connected with the 1, A I 1 2, A I 2 A A - Insurance 2 40% -
dynamics of country- or industry-specific stock Italy
indices, e.g., chemical, banking, automotive, rB I
3, B 3 B B - Automotive 3 - 80%
etc.), plus an idiosyncratic term which is typical of Idiosyncratic risk 10% 20%
Total 100% 100%
each individual company
Since the idiosyncratic component is not correlated to any
rj 1, j I 1 2, j I 2 ... n, j I n j j country/industry index, the correlation between company A and
company B boils down to
I1,I2,… In = common factors (country/industry indices)

A, B 1, A 3, B 1, 3 2, A 31, B 2,3
j = specific component for company j
Page: 425
Montecarlo simulations Montecarlo simulations
• When the number of obligors increases, • A large number of scenarios is generated. The
number is so high (e.g. 20,000) that the empirical
the analytical computation of joint distribution obtained in the end is a good
probabilities becomes more and more approximation of the “true” (theoretical) one
complex • For each scenario, a change in the asset values of
– 2 borrowers = 16 cases n obligors is generated. This is made by means of
random draws, but the random numbers generator
– 3 borrowers = 64 cases
takes into account the correlation among different
• Then, a different approach can be more borrowers
effective to estimate the distribution of • For each firm, the asset value change is compared
the future portfolio values to its thresholds. After checking for rating changes
(or default) the value of the loan is computed, and
so is the total value of the portfolio
MC simulation: a simple (but N. scenario 1 2. For every scenario

Impresa
Firm 1 1 +34%
general) example with 2 loans: Impresa
Firm 2 2 +20%
the change in the
firm’s asset value is
compared to the
2 asset value
429
Page:
Scenario # 1
1. 20.000 random asset thresholds (AVRTs),
Firm 1 Moves from B to A … and translated into a
value changes are generated
Firm 2 Stays in A … rating (or default)
for firms 1 e 2
(the values shown below
suggest that the two firms’
3. For every credit event N. scenario 1
asset values are strongly and
the value of the loan Impresa
Firm 1 1 IlLoan
credito
value:vale
108.7108.7
20%
positively correlated) is computed (based on Impresa
Firm 2 2 IlLoan
credito
value:vale
106.3106.3
-15% the forward rates or
-50% -25% on the recovery rate) Totale
Total IlPortfolio
portafoglio vale 215
value: 215
0% -50%
25% Note steps 2 e 3 are repeated
50%
Scenario # 1 2 3 4 … 20.000 for each of the 20,000 generated at step 1
Firm 1 +34% -8% +1% -20% … +43% N. scenario 1 2 3 4 … 20.000
Firm 2 +20% -10% +0% -22% … +52% Totale
Total 215 212.3 215.2 214.7 … 203.2
N. scenario 1 2 3 4 … 20.000 A MC simulation
Totale
Total
215 212.3 215.2 214.7 … 203.2
with more than 2 loans
430 4. The 20,000 values generated in the previous steps are now • To generate all asset value changes, a
Page:
ranked in increasing order, and used to compute the expected

value, the standard deviation, the percentile and the VaR bivariate normal distribution is not
84%
enough. A multivariate normal, e.g.
83% with n=10, must be used
10%
8% • Instead of one correlation coefficient
6% between 2 firms, a correlation matrix is
needed, reporting correlations between
4%
2%
0% each firm and the other 9
102 122 142 162 182 202
• Steps 2, 3 and 4 remain unchanged
Due to the high number of observations, the sample distribution
Is very close to the theoretical one
Creditmetrics: advantages and

Creditrisk+
disadvantages
– Uses objective and forward looking market • The insurance approach

data
– Interest rate curves and stock indices correlations • Estimating the number of defaults
– Evaluates the portfolio market value – The poisson distribution
– Takes into account migration risk
• but: • Estimating losses
– Needs a lot of data: forward rates, – Banding
transition matrices
• Injecting correlation into the model
– Assumes the bank is price-taker
– Assumes stable transition matrices – Conditional PDs
– Proxies correlations with stock indices and their macroeconomic drivers
– Maps counterparties to industries and
countries in an arbitrary and discretionary
way
Creditrisk+ and the “insurance”
(actuarial) approach The insurance approach 27
February
• Banks and insurance companies are 2006
Today the customers pay a
similar because they trade immediate premium against future risks….
payments for future ones
• Credit can be seen
Bank
as an insurance contract:
– the mark-up is a premium
– The default is a “contractual right”
of the borrower
Risk-free rate Costs Premium
Lending rate Page: 435

The insurance approach 27 The insurance approach

February
2009
Tomorrow the bank pays the
cost of their defaults
Bank
Bank
Build up the
reserves
Premi
Premi
Guarantees Losses Premi
Premi Reward the Face losses
Premi
Premia shareholder’s
Lent principal capital
Estimating the number of defaults:
The insurance approach
some useful mathematical tools
• If loans can be seen as insurance contracts,

then actuarial methods that were originally 90
• This chart depicts a
developed for the pricing and provisioning of 80 possible distribution of
insurance policies can be borrowed by bank 70 future defaults
risk managers 60 • Can we summarize it
• BUT the correlation among individuals must be in a more efficient

50
treated with care 40 way? Yes, by means of
– It can be ignored for simple insurance contracts 30 the probability
(e.g., life insurance) 20 generating function
– It must be injected into the model for more 10
sophisticated contracts (e.g. fire insurance, 0

0 1 2 3 4 5
loans)
Page: 439
The Poisson variable An example

• Given an expected number of defaults

(e.g. 4 over a portfolio of 100
counterparties with PD 4%) the Example with 3 loans
probability of having n defaults can be i Borrower Default
approximated by
n probability (pi )
e 1 Rossi 1%
p ( n)
n! 2 Bianchi 2%
• The approximation requires the default 3 Verdi 0.5%
events to be independent… # of expected defaults ( ): 0.035
– Not realistic, we’ll get back to it
• …and works well only if the PDs are low
Page: 441
Note: we traded precision
(continued) An example for ease of computation
0 , 035
e 0,0350 0, 035
p ( 0) e 96,56% • The p(n)s are greater than zero even
0! when n>3
0 , 035 • The quality of the approximation
e 0,0351 0, 035
p (1) 0,035e 3,38% declines when the pis are not small
1! enough
• Let us have a look at an example, to
p (2) 0,059% p (3) 0,001% see both of these drawbacks
Page: 442
Another example: (continued) Another example:

Example of bad approximations

Example of bad approximations Default probabilities of the borrowers
Default probabilities of the borrowers Rossi 25.0%
Bianchi 50.0%
Rossi 25.0%
Verdi 12.5%
Bianchi 50.0%
Probability of n defaults occurring
Verdi 12.5% Approximated True
Probability of n defaults occurring 0 41.7% >> 32.8%
Approximated True 1 36.5% << 48.4% Extremes
0 41.7% 32.8% 2 16.0% 17.2% are
36.5% 48.4% 3 4.7% > 1.6% overestimated
1
2 16.0% 17.2%
3 4.7% 1.6% 98,7%
A more realistic example Results for this example
i Borrower Default probability (p i ) 90% Probability

1 Rossi 1%
2 Bianchi 2% 85%
3 Verdi 0.50%
4 Gialli 2%
80%
5 Neri 1%
6 Mori 1% 20%
7 Grossi 1%
8 Piccoli 2% 15%
9 Astuti 2.50%
10 Codardi 2% 10%
11 Stupazzoni 0.50%
12 Molinari 2% 5%
# of
13 Vasari 1%
defaults
# of expected defaults ( ): 0.1850 0%
Page: 446 0 1 2 3 4 5 6 7 8 9 10 11 12Page:

13 447
© M. Anolli - Risk Management © M. Anolli - Risk Management
From the number of defaults

to the amount of losses The answer is called banding
• The model above foresees the number

of defaults • Fix a measurement unit L (e. g., 10,000 €)
• Yet, credit risk models (and the pricing, • Divide all exposures Li by L and round
provisioning, capitalization schemes them up, getting standardized values vi
that rely on these models) require an • Also re-write expected losses using the
estimate of the probability distribution new measurement unit. They become
of future losses i= i/L with i piLi
– Do not round up, as it would lead to useless
• Are the tools above – which refer to loss of precision
discrete random variables – still • Group into a single bucket (“band”) all
suitable for losses (a continuous loans of equal size vi
random variable)?
A practical example: A practical example (continued)
Name Probability Exposure standardized Expected Rossi 1% 11,000 1

of default exposure loss
(pi) (Li) ( vi) ( i) Bianchi 2% 12,000 1
Rossi 1% 11,000 1 110 Verdi 0.50% 11,000 1
Gialli 2% 9,500 1
11,000 / 10,000 1
11,000 x 1%
(recovery rate=0)
All into band # 1
Note: this is where recovery
i 110 rates fit into the
i 0,011 Creditrisk+ model, in a
L 10000© M. Anolli - Risk Management

simple, deterministic way
Page: 450 © M. Anolli - Risk Management Page: 451
For every band we can

Practical example (continued) compute
1 Rossi 1% 11,000 1 110

2 Bianchi 2% 12,000 1 240 i
j # of expected defaults in band j
3
4
Verdi
Gialli
0.50%
2%
11,000
9,500
1
1
55
190 i vi v j vi
5 Neri 1% 22,000 2 220
6 Mori 1% 21,000 2 210 m
7 Grossi 1% 19,500 2 195
While j # of expected defaults in the whole portfolio
8 Piccoli 2% 20,800 2 416
j 1
9 Astuti 2.50% 33,000 3 825
10 Codardi 2% 28,500 3 570
11 Stupazzoni 0.50% 31,000 3 155
12 Molinari 2% 30,800 3 616 j i Total expected loss in band j
13 Vasari 1% 29,000 3 290
i vi j
Each band is a small portfolio, with
In our example we get: losses proportional to defaults
Summary data for the 3 bands

n Probability of n defaults occurring
Exposure # of expected Expected e j
j in the j-the band, that is, of losing
p ( n) an amount nvjL
vj defaults j loss j n!
1 0.060 0.06
Alternatively, we can write:
2 0.052 0.10
3 0.082 0.25 Probability associated with
n
Totale =0.1934 0.41 e j
j nvj losses, each one of amount L,
p (nv j ) all coming from bank j
n!
In order to obtain the distribution of In our example, we get:

losses one needs to combine these p
Facoltà di Scienze Bancarie, Finanziarie e Assicurative Loss Probability
(nL) 90%
• Why? Think at the example of a 120,000 euro loss
(12L). This can derive from 0 - 82.41% 80%
– 12 defaults of band 1 1 10,000 4.90%

70%
– 6 defaults of band 2 2 20,000 4.44%
60%
– 4 defaults of band 3 3 30,000 7.01%
– 2 defaults of band 6… 4 40,000 0.52% 50%
5 50,000 0.37%
• All these cases must be combined in order to obtain the 40%
probability of a 120,000 euro loss in the entire portfolio 6 60,000 0.30%
30%
• This can be done deriving, for each band, a pgf similar 7 70,000 0.03%
to the one for each band and combining these pgf 8 80,000 0.02% 20%
– The pgf of the portfolio-sum is the product of each 9 90,000 0.01% 10%
individual pgf 10 100,000 0.00%
0%
… … … 0 1 2 3 4 5 6 9 30
30 300,000 0.00%
Page: 457
How to inject correlations Default probabilities of Messrs.
into the model Rossi, Bianchi and Verdi
4.5% 4.0%
• The probability distribution for future 4.0% 3.5%
losses has been derived in a (relatively) 3.5% 3.0%
painless way because of the assumption 3.0% Gdp chg.

2.5%
that all loans are uncorrelated 2.5%
2.0%
Verdi
2.0% Bianchi
• We now show how correlations can be 1.5%
1.5%
Rossi
added to the basic model 1.0% 1.0%
0.5% 0.5%
0.0% 0.0%
95 96 97 98 99
© M. Anolli - Risk Management Page: 458 Page: 459
Default probabilities of Messrs. The logic driving

Rossi, Bianchi and Verdi this enhanced model
• Individual PDs are not constant, but

swing around a long-term average 1 Rossi 1%
because of macroeconomic factors 2
3
Bianchi
Verdi
2%
0.50%
• They can be seen in the following way

90%
4 Gialli 2% 80%
5 Neri 1% 70%
60%
6 Mori 1%
50%
7 Grossi 1%
~
40%
x
8 Piccoli 2% 30%
pRossi ~
20%
9 Astuti 2.50%
pRossi x
10%
10 Codardi 2% 0%
0 1 2 3 4 5 6 9 30
11 Stupazzoni 0.50%
12 Molinari 2%
13 Vasari 1%
Stochastic disturbance
Average long-term
value 1. For every possible draw of x, the individual PDs and the
future loss distribution can be worked out accordingly
Page: 461
2. Step 1 is repeated looping over all possible 3. The “weighted average” of those scenarios (each one conditional
values of x and generating n scenarios, each one to a given value of x) provides the unconditional distribution
with its own probability distribution: of future losses
90%
80%
1
Facoltà di Scienze Bancarie, Finanziarie e Assicurative 70%
60%
462
463
Page:
Page:
50%
40%
1 Rossi 1%
2 Bianchi 2% 30%
3 Verdi 0.50% 20%
4 Gialli 2%
10%
5 Neri 1%
90%
6 Mori 1% 0%
0 1 2 3 4 5 6 9 30
7 Grossi 1% 80%
8
9
10
Piccoli
Astuti
Codardi
2%
2.50%
2%
90%
1 90% 2 90% 70%
~
80% 80% 80%
11 Stupazzoni 0.50% 60%
2 N
70% 70% 70%
x1
12 Molinari 2% 60% 60% 60%
90% 50%
p ( x1 ) p ( x2 ) ... p ( xN )
13 Vasari 1%
~
50% 50% 50%
1 Rossi 1%
80% 40%
2 Bianchi 2% 40% 40% 40%
x3
3 Verdi 0.50% 70% 30% 30% 30%
4 Gialli 2% 30%
60% 20% 20% 20%
5 Neri 1% 10% 10% 10% 20%
~
6 Mori 1% 50%
0% 0% 0%
7 Grossi 1% 40% 0 1 2 3 4 5 6 9 30 0 1 2 3 4 5 6 9 30 0 1 2 3 4 5 6 9 30
10%
x2
8 Piccoli 2%
30%
9 Astuti 2.50% 0%
1 Rossi 1% 20%
10 Codardi 2% 0 1 2 3 4 5 6 9 30
2 Bianchi 2%
11 Stupazzoni 0.50% 10%
3 Verdi 0.50%
4 Gialli 2% 12 Molinari 2%
0%
5 Neri 1% 13 Vasari 90% 1% 0 1 2 3 4 5 6 9 30
6 Mori 1% 80%
7 Grossi 1%
70%
8 Piccoli 2%
9 Astuti 2.50% 60%
10 Codardi 2%
50%
11 Stupazzoni 0.50%
3
40%
12 Molinari 2%
13 Vasari 1% 30%
20%
10%
© M. Anolli - Risk Management 0%

0 1 2 3 4 5 6 9 30
In the final distribution A tiny example:

risk has increased
Default probabilities of 2 borrowers in 2 possible states of the world
• Compared to the “basic model”, where (a) Boom
individual PDs are thought to be known Bianchi
without error, now we have two risk
Defaults Survives Total
sources instead of just one::
Defaults 0.08% 1.92% 2%
– Will Rossi & friends really default?
Rossi Survives 3.92% 94.08% 98%
– But then, what is the probability that they may
default? Total 4% 96% 100%
• Extreme events are now more likely (b) Recession
• In other words, the model now shows the Bianchi
effects of the correlation linking Rossi, Defaults Survives Total
Bianchi, Verdi and friends Defaults 0.60% 5.40% 6%
– Portfolio diversification works less well Rossi Survives 9.4 0% 84.60% 94%
Total 10% 90% 100%
Page: 465
CreditRisk+: advantages
Unconditional distribution and disadvantages
– Simple inputs: PDs and exposures (book

Bianchi value) net of recovery are enough
Fails Survives Total • The “correlated” version requires also sensitivities
to the economic cycle factors
Fails 0.34% 3.66% 4%
• An analytical solution exists
Rossi Survives 6.66% 89.34% 96% • Possibility to obtain the distrbution of losses
Total 7% 93% 100% without recurring to simulation techniques
– But:
0,34% > 7% x 4% (0,28%) • Only looks at default risk
• Does not consider migration risk
> 1% • Assumes constant exposures
recovery risk
does not consider
Page: 466
Outline of this session

Alternative measures of LGD

LGD estimation and recovery risk
The "traditional" approaches on LGD
estimation: results from previous research
Bank loans: what’s different?
The value of time
courtesy of Andrea Sironi
Estimating gross recoveries through a
multivariate model
© M. Anolli - Risk Management 469

Alternative measures of LGD Market LGD
Three methods for estimating recovery rates Usually computed as the ratio of post
Market LGD direct observation of default market price - from 15 to 60 days
defaulted debt prices in the market after default - to face value
Workout LGD PV of ex-post realised Suitable measure of loss for investors who
cash-flows back to the time of default sell their debt immediately after default
Implied LGD inference of recovery rates Not necessarily an accurate measure of
from non defaulted bond prices or from loss if the debt is worked out by the bank
CDS premia to eventual resolution
470 471
© M. Anolli - © M. Anolli -
Workout LGD Workout LGD

Particularly suited for bank loans risk management The actual recovery rate RR on a
and supervision: liquid secondary market of defaulted exposure (including the value
defaulted loans typically does not exist of time and recovery costs) can be
Consistent with Basel II definition of LGD
computed as:
Sensitive to changes in bank-specific workout
discounted net value administrative
processes (e.g. length of time, effectiveness, etc.) of the recovery workout costs
Commonly used in banking practices
DNR FR FR AC T
A suitable discount rate must be selected which RR (1 r )
should be consistent with the ex ante risk of the EAD EAD FR
workout process exposure at default
face value
of the recovered
amount
472 473
Market implied LGD Market implied LGD
Derived from risky (but not defaulted) bond prices A simple example
using a model of credit risky debt valuation
The spread above risk-free (i.e. Treasury) bonds is an 1 i 1 r 1 p 1 r pk r i
k 1
indicator of the risk premium demanded by investors p1 r
However, this spread reflects both PD and LGD, as well i = risk free rate 5%
as liquidity premiums
Only recently have models been developed which allow
r = risky rate 6%
one to separately identify these two parameters from p = probability of default
bond spreads (Bakshi, Madan and Zhang (2001) and
Unal, Madan and Guntay (2003). 1.5%
They find that on average, recovery rates obtained in k = recovery rate
this way lie systematically below the “physical”
recovery rates 6% 5%
k 1 37.11%
It is typically a risk neutral measure 1.5% 1 6%
474 475
Drivers of recovery risk

Empirical evidence
from “traditional” studies
Most LGD studies draw on data on Variables affecting recoveries

defaulted bonds can be classified into three groups
defaulted syndicated loans Debt characteristics (where most researchers
focus)
The recovery rate (1 – LGD) is usually Instrument type: e.g., bank loans versus bonds
computed as the ratio of post-default value Seniority
Collateral and covenants
to face value Borrower characteristics
A secondary market of defaulted exposures Leverage (for unsecured exposures)
must be available (“market LGD”) Liability structure (“debt cushion”, i.e., % of
Expected, not actual, recoveries are used liabilities
which are junior to the exposure being analyzed)
These can underestimate actual values, Average recovery rate for the industry
as risk-averse investors require a discount Environmental variables (cycle, markets)
Such a bias is practically demonstrated Current and expected default frequency
by the existence of vulture funds Cycle indicators (e.g., GDP, demand, etc.)
476
Industry-specific conditions at the time of default 477
Debt characteristics: debt type Debt characteristics: seniority
Bank loans clearly outperform bonds: why?
1) Banks monitor borrowers more closely and
timely than bond investors? Senior Senior Senior
2) Is there a sample bias? Bank loan data are Study Secured Unsecured Subordinated Subordinated
available only for large syndicated exposures, usually Fons (1994) 65% 48% 40% 30%
assisted by covenants or claims on specific assets. Altman & Kishore (1996) 58% 48% 34% 31%
Van de Castle & Keisman (1999) 66% 49% 37% 26%
Hu & Perraudin (2002) 53% 50% 38% 33%
Bank loans Average 61% 49% 37% 30%
Subordinated
Seniority might look as a
key driver of bonds’ recovery risk.
Senior However…
Bonds Subordinated
Senior Unsecured
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% Senior Secured
0% 10% 20% 30% 40% 50% 60% 70%

Source: Acharya et al., 2003 - Recoveries based on the market price of the
financial instruments received in settlement of the defaulted debt
478 479
© M. Anolli -
© M. Anolli -
Risk
Debt characteristics: seniority Debt characteristics: seniority

Figure 4 – Recovery rates for different seniorities: 1982-1999

95%-confidence intervals for average values
70%
90,00%
60% 80,00%
70,00%
50%
60,00%
50,00%
40%
40,00%
…average figures hide a lot
30,00%
30% of volatility: individual
values may fall quite apart 20,00%
20% from their mean, and even 10,00%
Senior Senior Senior Subordinated the average value for 0,00%

Secured Unsecured Subordinated different seniority classes 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
may not be statistically
Based on Altman-Kishore, 1996 different from each other… Senior Secured Subordinated
480 481
Source: Moody’s
Debt characteristics: collateral Debt characteristics: size
Asarnow and Edwards (1995) look at Citibank’s middle
Source: Acharya et al., 2003 - Recoveries based on the

Current Assets
market and large corporate lending from 1970–93 and
market price of the financial instruments received in

All or Most assets
find no relation between loss given default and size of
loan
Real Estate Carty and Lieberman (1996), using Moody’s data on
syndicated lending, arrive at a similar negative result
Plants, property and equipment Thornburn (2000), in her study of Swedish small
settlement of the defaulted debt

business bankruptcies, also found that firm size doesn’t
Unsecured
matter in determining LGD
Secured However, empirical studies based on bank loans seem
to indicate a negative relationship between loan size
Unsurprisingly,
liquid assets Other assets and recovery rate: Grippa et al., 2007 (Italian banking
look easier to system) and Carvalho & Dermine, 2007 (Portuguese
seize and sell 0% 20% 40% 60% 80% 100%
482 bank). 483
© M. Anolli -
© M. Anolli -
Risk
Borrower characteristics: Recovery Rates for Financial

industry Institutions
1975 – 2007
© M.
Anolli -
Risk
Manage
ment
Average Weighted
Public Utility Recovery Average
Number of Number of Rate (%) Recovery
Finance Financial Sector Issuers Issues (Issues) Rate (%)
E.g., public utilities have a large
Industrial endowment of physical capital Banks 7 20 26.91 19.87
Insurance for which a buyer can be found. Mortgage Bankers 7 17 44.97 37.75
Transportation Real estate companies tend to Accident, Health & Life Insurance 7 19 26.95 19.79
default when the value of their Fire, Marine & Casualty Insurance 9 15 26.05 21.31
Real Estate assets (e.g., new commercial Savings Institutions 4 9 13.28 18.67
Thrifts buildings) is lower than expected.
Investors 3 4 24.00 23.51
Other non-bank Note that industry concentration may also Security Brokers & Dealers 2 8 43.59 42.30
Banking play a role, as it reduces the number of
Miscellaneous Credit Institutions* 7 31 70.17 68.52
potential buyers for distressed assets.
Securities All Financial Sectors* 46 123 39.61 40.90
*Removing FINOVA defaults from calculations (19 issues, Recovery Rate = 86% for all issues) results in an Average Recovery of 38.97%
and a Weighted Average Recovery of 30.69% in the Miscellaneous Credit Institutions Sector as well as an Average Recovery of 31.13% and
0% 20% 40% 60% 80% Weighted Average Recovery of 26.38% in all Sectors.
484 Source:NYU Salomon Center Default Database
485
Source: Hu and Perraudin (2002) © M. Anolli -
Risk
Recovery Rates for Financial Borrower characteristics:
Institutions financial ratios
Various financial ratios have been shown to be helpful

With the massive turmoil in credit markets in in explaining recoveries
2008, the data on defaulted financial services The assets to liabilities ratio acts like a coverage ratio of
the funds available versus the claims to be paid. Hence it
firms have increased has been used as an explanatory variable by e.g., the
Moody’s LossCalc model
The most noteworthy financial sector defaults Not meaningful for secured credits
were Washington Mutual, GMAC, Residential The relative seniority of debt (i.e., its rank within the
capital structure of the firm) has been used by both
Capital, and Lehman Brothers Moody’s and S&P’s
Absolute seniority alone may be misleading: preferred stock
The overall average recovery for financials (about might hold the highest seniority rank within a particular firm
40%) is below the average for all defaults. that has no funding from loans or bonds
The firm’s profit margin (EBITDA/Sales- Acharya et al.,
However, if we exclude one important outlier 2003)
(FINOVA), the arithmetic average for all financials A high value indicates that the firm defaulted because of
liquidity or financial problems, but retains a good operating
decreases significantly to 34%, while the profitability: hence, a potential buyer might be willing to pay a
fair price for it
weighted average decreases to 30% 486 487
Borrower characteristics: Years after issuance (age)

size
Average Price at Default by Number of Years After Issuance (1971-2008)
Empirical evidence of the relevance of Years to Default

1
No. Of Observations
174
Average Price
32,61
the borrower’s size is not clear 2 355 34,22
3 429 32,78
Eales & Bosworth (1998), and Hurt & 4 349 38,21
Felsovalyi (1998): inverse relationship 5
6
260
217
36,41
46,91
recovery rate – borrower size 7 153 44,83
8 85 47,12
Carty & Lieberman (1996), Bartlett 9 57 45,06
(2000), Thorburn (2000), Davydendo & 10
All
157
2236
38,92
37,75
Franks (2004) no clear effect of size Source: NYU Salomon Center.
on recoveries Although we observe some aging effect with RR increasing as the years to
default increases, there is not a great difference between the first and fifth
488 years (32–36%) or between the sixth through ninth years (47–45%) 489
Original bond rating before default Environmental variables:
GDP and industry conditions
Average Price after Default by Original Bond Rating

Altman, Resti and Sironi (2002) find that
No. of Observations Average Price ($)
AAA 14 82,55 recoveries are systematically lower when the
AA 29 65,68 real GDP growth is below its long-term
A 169 49,55 average
BBB 421 41,87 Acharya et al (2003) use industry-specific
BB 215 37,6
distress indicators:
B 1135 34,53
CCC 243 37,37 When the industry of the borrower is in distress,
All 2226 38,36 average recoveries are lower by 10-20%
The borrower’s competitors are not going to bid for
her assets, as they
As expected, the better the original rating, the greater the average recovery
rate, but only in the investment-grade range fallen angels. Once below are financially constrained
BBB, the weighted-average recovery rate varies narrowly, 34–37% 490 already suffer from overcapacity
491
Environmental variables: Environmental variables:

GDP and industry conditions default rates
US bond market data (Altman, Resti, Sironi, 2002)
show that recoveries are lower when default rates
12 70 increase.
10 60
65
A bad state of the economy, reducing the value of
1987
60 corporate
y = -2 .6 1 7 xassets,
+ 5 0 .9 might
y = increase 2 LGD
0 .5 6 0 9 x - 8 .7 as6 0well
564x + .6 1 as
2
default risk.
2
50 R = 0 .4 4 9 8 R = 0 .6 0 9 1
1993
8 55
1983
1997
Recovery ra
On the distressed bond market, the effect might be

Default rat
1996
40 50 1992 y = -1 1 .1 8 1 L n (x ) + 5 2 .3 3 2
y = 5 2 .7 3 9 x -0 .2 8 3 4
6
Default rate 1984 stronger, as more defaults increase
R 2 = 0 .5 8 1 5 2 the supply of
Recovery Rate (%)

R = 0 .6 0 0 4
30
Recovery Rate 45
1985
1988
defaulted issues, leading to a decrease in prices (on
1995
which RR estimates are usually based).
4 40 1994
1982
20 1989
1998 1991
35 1986
2 10
30
1999
2000
0 0 25
1990
82
84
86
88
90
92
94
96
98
00
20
19
19
19
19
19
19
19
19
19
20
0 2 4 6 8 10 12
Year D e fa u lt R a te (% )
Source: Altman et al. (2005) 492 493

Summary of main results Recovery Ratings from the Rating
Agencies
Agency Moody’s Standard & Poor’s Fitch
Main determinants of recovery rates are security
and seniority Ratings Loss Given Default Recovery Ratings Recovery Ratings
Type Ratings
Bank loans, being at the top of the capital Ratings LGD1 0-9% 1+ 100% RR1 91-100%
structure, typically have higher recovery than Scale LGD2 10-29% 1 100% RR2 71-90%
bonds. LGD3 30-49% 2 80-100% RR3 51-70%
Recoveries are systematically lower in recessions LGD4 50-69% 3 50-80% RR4 31-50%
LGD5 70-89% 4 25-50% RR5 11-30%
Industry seems to matter: tangible asset- LGD6 90-100% 5 0-25% RR6 0-10%
intensive industries, e.g. utilities, have higher
recovery rates than service sector firms, with Assets Non-financial US and Canadian secured All corporate,
some exceptions such as high tech and telecom Rated corporate bank loans to which it financial institutions
speculative-grade assigns bank loan ratings, to and sovereign
Size of exposure seems to have no strong effect issuers in the US senior secured loans in issuers rated in the
on losses for bonds, but seems to negatively Europe, and to any secured single B category
affect recoveries in the case of bank loans bonds issued along with and below
494 rated bank loans 495
© M. Anolli -
© M. Anolli -
Risk
Moodys’ Standard & Poor’s

4 Main steps
1. Establish priority of claim 8 main steps
Jr bonds are subordinated to Sr bonds, but may or may 1. Review transaction structure
not be subordinated to other unsecured obligations
Prioritize claims across affiliates 2. Review borrower’s projections
2. Assume a beta probability distribution for 3. Establish simulated path to default
potential Enterprise Value (EV) outcomes 4. Forecast borrower’s free cash flow at default
For most issuers, assume a beta distribution of EV based on simulated default scenario and
relative to total liabilities
default proxy
Corporate LGD distribution will have 50% mean and
26% standard deviation 5. Determine valuation
3. For each EV outcome, calculate LGDs for each 6. Identify priority debt claims and value
security class implied by absolute priority 7. Determine collateral value available to lenders
4. Expected LGD equals the probability-weighted 8. Assign recovery rating
averages of LGDs across EV outcomes 496 497
Market versus workout approach:
Fitch how frequency distributions change
3 Main steps % of cases

30% Suppose some kind of
Estimate the enterprise value (EV) bonds/loans are fully
25% recovered 50% of the
Establish the level of cash flow upon which it is most
times, fully lost
appropriate to base the valuation 20% otherwise.
Apply a multiple reflecting a company’s relative position
within a sector based on actual or expected market 15%
They will increase the
and/or distressed multiples
tails of the blue (WA)
Estimate the creditor mass, ie identify existing 10% distribution.
claims 5%
Claims taken on as a company’s fortunes deteriorate However, investors
Claims necessary to the reorganization process 0% expectations (hence,
Claims that have priority in the relevant bankruptcy market prices) will
0-5%
5-10%
10-15%
15-20%
20-25%
25-30%
30-35%
35-40%
40-45%
45-50%
50-55%
55-60%
60-65%
65-70%
70-75%
75-80%
80-85%
85-90%
90-95%
95-
code cluster around fifty cents
Distributing the EV the dollar: the central
part of the pink
The resulting value is allocated to creditors according to distribution (MA) will
jurisdictional practice A Workout approach
498 increase.499
A Market approach
© M. Anolli -
Risk
Bank loans: what’s different? Bank loans: what’s different?

As bank loans usually do not have a Additional problem: the distribution tends
secondary market, no “price at default is to be bimodal
available”
The whole recovery process must be tracked,
(“workout approach”) The mean is a poor indicator
This can be hugely demanding in terms of data
The value of time must be factored into the
computations, although there is no easy way to do
it High probability of LGD values close to 0% and
Frequency distributions become hugely 100%
different from those based on the “market
approach”
The amount of variance to be explained through
recovery drivers increases dramatically
500 501
Bimodal distributions make the Overall means have to be replaced
mean a poor indicator by many conditional means
35%
30%
25%
20%
15%
…
10%
5%
For uni-modal
distributions (like the …but it might be … and becomes 0%
normal ) the mean is unsatisfactory when clearly inadequate for
05
15
25
35
45
55
65
75
85
95
0
1
1
9
a reasonable of bimodal, that is U-
0.
0.
0.
0.
0.
0.
0.
0.
0.
distributions are
0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
summarizing uniform… shaped distributions
individual values…
A vector summarizing the
502
From to (x) 503
characteristics of the facility
Risk Risk
All relevant drivers must be

factored into a multivariate model
A multivariate model
Vector with the

characteristics of
Facility
Collateral Guarantees … Let us have a look at real-life results
type (although simplified for confidentiality reasons)
the facility and
the borrower We are going to explain gross recoveries (including
recovery costs) at face value (not discounted)
These can be discounted later,
based on our expectations on future interest rates
Logistic
Linerar
function 1
function
DNR FR FR AC T
1 RR (1 r )
0 EAD EAD FR
0
We are going to use two alternative frameworks
A linear model
Before we see how recoveries A logistic model
(x) can be explained, let us look
at how they must be computed
504 505
© M. Anolli -
© M. Anolli -
Risk
An example of linear multivariate Gross recoveries:
model for gross recoveries logistic versus linear models
Overdrafts
The models
Base value lead to similar
Facility 120%
Receivables results
(no extra collateral) Collateral/guarantees Linear
100%
Receivables wih a
Logistic
The effect of extra collateral The logistic
third-party full guarantee 80%
and/or guarantee on
expected losses (hence, on function is
Receivables with financial
Collateral (100% coverage) pricing) can be quantified 60% more correct,
in an easy and intuitive way yet less
Long term loan 40% readable
(no collateral) However, puzzling
results may prompt 20%
Long term loan with rejection from credit
third-party full guarantee
officers!
0%
Mortgage loan
100% coverage Mortgage Mortgage Generic C/account C/account …
loans loans c/account overdrafts overdraft
Mortgage loan 117%! with 100% with 150% overdrafts with 3rd pty 50%-covered
200% coverage
coverage coverage guarantee by collateral
0% 40% 80% 120%
506 507
Risk Risk
Recovery risk Recovery risk

Carty & Lieberman Altman & Kishore

DNR FR FR AC T
RR (1 r ) Seniority class Number Mean Std. dev. Number Mean Std. dev.
EAD EAD FR Senior Secured 115 53.80% 26.80% 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%
The ex ante estimated recovery rate is a function
Subordinated 226 32.74% 20.18% 214 31.34% 22.42%
of four stochastic variables:
Junior Subordinated 9 17.09% 10.90% - - -
The amount to be recovered (FR)
Source: Carty L.V. e Lieberman D. (1996a); Altman E. e Kishore, V.M. (1996)
The workout administrative expenses (AC)
The discount rate (r)
The duration (T) of the recovery process High volatility: in case of a uniform distribution the mean would be 0.5
and the standard deviation 29%
Recovery risk: risk that the actual recovery rate
will be lower than estimated
508 509
Recovery risk Recovery risk
Binomial approach to expected loss Example:

measurement PD = 0.5%
Event Loss Probability

LGD = 50% UL 0.5 0.005 0.995 3.53%
Default LGD PD If stochastic LGD
No default 0 1-PD
EL PD LGD 1 PD 0 PD LGD UL PD 1 PD LGD

2
PD 2 2
LGD PD (1 PD ) 2
LGD
hp. LGD constant
UL LGD PD 1 PD UL PD 1 PD LGD
2
PD 2
LGD
510 511
Recovery risk The value of time

Example: The actual recovery rate RR on a

PD = 0.5% defaulted exposure (including
the value of time and recovery
LGD = 50% costs) can be computed as:
Variance LGD = 20%
discounted net value administrative
2 of the recovery workout costs
UL 0.005 0.995 0.5 0.005 0.2 4.74%
Recovery risk becomes more important DNR FR FR AC T
RR (1 r )
the higher is PD EAD EAD FR
face value
of the recovered
amount
512 exposure at default 513
© M. Anolli - Risk Management © M. Anolli -
Choosing the discount rate:
Choosing the discount rate
historical versus future values
The choice is particularly relevant as LGD Should we use historical values observed
estimates tend to be very sensitive to between the default and the end of the workout
different levels of the discount rate. process?
Moral-Oroz (2002): empirical study on a This would lead to a backward-looking
mortgage portfolio for a Spanish bank an measure, which would not account for the
present (and future) market conditions.
increase of 1% of the discount rate leads to
When estimating LGDs on future bad loans, we
an increase of LGD of 8% are concerned with the interest rates that
The authors also show that LGD variability might prevail on the market after a new
coming from the use of different rates, all default has emerged
existing in the 900 days period of the Since a bank’s PDs usually imply a one-year
valuation can be very high: 20% maximum risk horizon, estimated LGDs actually refer to
change defaults that might emerge one year later.
514 515
Choosing the discount rate: Choosing the discount rate:

historical versus future values which risk-premium?
Should we use historical values observed In discounting recoveries, some kind of risk
between the default and the end of the workout process? premium should be included, to account for the
This would lead to a backward-looking measure, which would bank’s risk aversion
not account for the present (and future) market conditions. Different options (Maclachlan, 2007):
When estimating LGDs on future bad loans, we are concerned Risk-free rate or interbank rate (Bank of Italy, 2001)
with the interest rates that might prevail on the market after Fixed rate: e.g. 5% (Moral, Garcia-Baena, 2002) +
a new default has surfaced. sensitivity of LGD estimates to different levels of discount
Since a bank’s PDs usually imply a one-year risk horizon, rate (Moral-Orozo, 2002)
estimated LGDs actually refer to defaults that might emerge Rate on speculative-grade or distressed bonds (Araten et
one year later. al., 2004)
7,5%
One-year forward interest rates might one-year forward rates 1rT Rate applied to new originations for the lowest quality
be used as a quick forecast of the
7,0%
grade in which a bank originates similar transactions (OCC
future spot rates. 6,5%
2003)
If the expected duration of the 6,0%
spot rates, rT The rate agreed on the loan when it was a performing
recovery process is T years, a T-year 5,5% exposure (Edwards and Asarnow 1995, Carty et al 1998,
forward rate for investments starting in 5,0%
IASB 2004)
one year (1rT) should be derived by Lender’s cost of equity (Eales and Bosworth 1998)
current spot curves 4,5%
Maturity - T
Coupon rate (Friedman and Sandow 2003)
4,0%
1 2 3 4
516
5 Same rate as used for an asset of similar risk (FSA, 2003)
517
which risk-premium?
which risk-premium?
The discount rate should be consistent with the

Guha (2003): regardless of promised yield and
remaining term to maturity, secondary market prices of
risk of the new financial claim
equal ranking traded corporate bonds converge to a Risk-free rate
common fractional value of par after default after In reality recovery is uncertain and therefore
default the pre-default yield is no longer relevant risky
Defaulted debt should be seen as an asset worth the Contractual loan rate or rate applied to worst
expected share of the post bankruptcy pool of assets grade customers
Post default cash flows represent a financial asset
default results in a change of the nature of the financial it confuses pre- and post-default required
claim returns
The debt restructuring could result in the issuance of Lender’s cost of equity
risky assets such as equity or warrants, or less risky Underlying logic: shareholders cover the cost
ones such as notes, bonds or even cash the correct of recapitalising the bank. Problem: it
rate would be for an asset of similar risk (FSA, 2003). mistakenly replaces the systematic risk of the
518 defaulted debt with the risk of the bank 519

which risk-premium? risk-free versus risk adjusted values
The discount rate may be equal to the risk free rate is In discounting recoveries, some kind of risk premium
should be included, to account for the bank’s risk aversion
the lender holds full recourse to cash collateral One might use the spreads on speculative-grade or distressed
If the collateral is represented by a long term “default bonds
free” asset (e.g. Treasury bond), a risk premium may The rate agreed on the loan when it was
a performing exposure does not look especially advisable
still remain if the instrument’s value is correlated to the The original risk profile has changed!
market return
A risk premium might not be
Where collateral, unrelated to the firm’s default risk, is required, however, if the Probability d.f.
taken as security, the risk of recovery is related to the uncertainty on the actual
secondary market price of the collateral (bonds, stocks, recovery rate on future defaults
etc.), and hence its market beta, and not necessarily is already accounted for by other
means
the firm’s asset beta
E.g., when the whole probability
More generally, the discount rate should vary according distribution of nominal
to the sources of repayment, and multiple discount recoveries (not just their
conditional mean) was
rates may be applicable (individual asset class discount estimated and some
rates) 520
conservative percentile has been
5211
chosen 0 0,5
LGD on some class of loans
Conclusions Further readings
Main determinants of recovery rates include

security, seniority, industry and economic https://www.bancaditalia.it/pubblicazioni/note-
cycle stabilita/2017-
Recovery rates have a “bimodal” distribution 0007/en_Note_di_stabilita_finanziaria_e_vigilanza_N._7.
means are not relevant need for PDF?language_id=1
conditional means all relevant variables https://www.bancaditalia.it/pubblicazioni/note-
must be considered in a multivariate model stabilita/2017-0011/eng-note-stabilita-finanziaria-
Defaulted debt is a new type of financial vigilanza-N-11.pdf?language_id=1
asset for which an appropriate discount rate
has to be selected according to its risk
Recovery risk seems relevant. However, is it
diversifiable or is it linked in some way to
default risk? 522
Agenda
• Rating assignment
Rating and rating • Rating quantification
• Rating validation
systems

Rating: definition Internal and external ratings
• concise evaluation of creditworthiness • differences based on

• based on economic and financial structure of the – borrowers being evaluated
debtor, its market, its products, management, • bond issuers vs. ordinary borrowers:
industry… – available information
• external ratings, produced by rating acencies • comparable during rating assignment, different
• internal ratings (Basel 2 and 3), based on during the life of the contract (banks have valuable
additional pieces of info from the account turnover)
qualitative and quantitative methodologies – different incentive ssystems
• reputation (independent credit opinion) vs. value of
the investment; through the cycle vs. point in time
Agency ratings Rating and financial ratios

• history
• based on analytical models integrated by
subjective judgement
– quantitative analysis CAMEL or the like. 5 Cs: character
(integrity), capacity (sufficient cash flow to service the
obligation), capital (net worth ), collateral (assets to
secure the debt), and conditions (of the borrower and
the overall economy
– qualitative analysis: management, efficiency of
processes, reputation, geographical diversification,
marketing, market power
• aim: to evaluate business risk and financial risk

Rating and financial ratios Rating classes
Indicatori di bilancio per classe di rating

(valori medi relativi a 3 anni: 1998-2000) – Imprese USA
AAA AA A BBB BB B CCC
EBIT interest coverage 21,4 10,1 6,1 3,7 2,1 0,8 0,1
EBITDA interest coverage 26,5 12,9 9,1 5,3 3,4 1,8 1,3
Free operating cash flow/total 84,2 25,2 15,0 8,5 2,6 (3,2) (12,9
debt (%) )
Funds from operations/total debt 128, 55,4 43,2 30,8 18,8 7,8 1,6
(%) 8
Return on capital (%) 34,9 21,7 19,4 13,6 11,6 6,6 1,0
Operating income/sales (%) 27,0 22,1 18,6 15,4 15,9 11,9 11,9
Long-term debt/capital (%) 13,3 28,2 33,9 42,5 57,2 69,7 68,8
Total debt/capital (%) 22,9 37,7 42,5 48,2 62,6 74,8 87,7
Numero di società 8 29 136 218 273 281 22
Fonte: Standard & Poor’s, de Servigny, Renault (2004)
Internal ratings Rating quantification

• They share with external ratings the aim: to • on the basis of the rating of a subject it is
evaluate business risk and financial risk possibile to infer the evolution in the riskiness
• They differ : of subjects classified in previous periods in the
– # of classes (2 to 20 pass 0 to 6 fail); lower same rating class and so assess the PD
granularity for advanced banks • the rating data published by rating can be the
– iformation considered: scoring systems; industry basis for
analysis; qualitative analysis; credit registry or credit – mortality rates of the different rating classed and
bureau data time horizons
– default definition – rating migrations (frequency with which a given
– borrower to facility rating passage credit moves from a given rating class to another
– time of periodical reevaluation class) -> transition matrices

Marginal and cumulative default Marginal default rates
rates
Tassi di insolvenza marginali - Moody's 1970-1994

•from rating agencies’ data Anni dopo 1 2 3 4 5 6 7 8 9 10
’emissione
Classe di
ating
Def t T Aaa 0,01% 0,02% 0,03% 0,03% 0,07% 0,07% 0,11% 0,12% 0,14% 0,15%
MDRt 1 SRt CDRT 1 SRt
Aa1
Aa2
0,02% 0,05% 0,07% 0,08% 0,10% 0,10% 0,11% 0,12%
0,02% 0,08% 0,12% 0,12% 0,14% 0,12% 0,11% 0,11%
0,13%
0,13%
0,15%
0,15%
:
Pop t t 1 Aa3 0,03% 0,08% 0,13% 0,14% 0,17% 0,15% 0,12% 0,16% 0,19% 0,22%
A1 0,05% 0,09% 0,14% 0,16% 0,21% 0,18% 0,14% 0,20% 0,24% 0,28%
•MDRt = marginal default rate in year t; A2
A3
0,06% 0,09% 0,15% 0,18% 0,24% 0,20% 0,15% 0,24%
0,09% 0,18% 0,23% 0,30% 0,33% 0,30% 0,31% 0,38%
0,30%
0,42%
0,35%
0,42%
•Deft = defaults in year t; Baa1 0,13% 0,27% 0,31% 0,43% 0,42% 0,40% 0,47% 0,51% 0,53% 0,50%
Baa2 0,16% 0,36% 0,40% 0,55% 0,51% 0,49% 0,63% 0,64% 0,65% 0,57%
•Popt = overall population in year t; Baa3 0,70% 1,11% 1,11% 1,19% 1,15% 0,98% 0,93% 0,91% 0,90% 0,84%
Ba1 1,25% 1,85% 1,82% 1,84% 1,80% 1,47% 1,22% 1,17% 1,15% 1,11%
•SRt = survival rate in year t; Ba2 1,79% 2,59% 2,53% 2,48% 2,44% 1,96% 1,51% 1,44% 1,40% 1,39%
Ba3 3,96% 3,90% 3,53% 3,12% 2,71% 2,60% 1,81% 1,75% 1,50% 1,47%
•CDRt = cumulative default rate in year t B1 6,14% 5,21% 4,54% 3,75% 2,98% 3,25% 2,11% 2,05% 1,60% 1,55%
B2 8,31% 6,52% 5,54% 4,39% 3,24% 3,90% 2,41% 2,35% 1,70% 1,64%
B3 15,08% 6,82% 5,21% 3,80% 3,14% 4,43% 2,58% 1,69% 2,54% 2,01%
534 Fonte: Moody’s (1996)
Marginal default rates Marginal default rates


Marginal, cumulative and survival
annualized default rates
rates
• ADRT
T
SRt (1 ADRT )T
t 1
cumulative default rates: proxy of the PD until a future
date hence
T
CDR5 1 0,9909 0,9634 0,9807 0,9722 1 0,9103 8,98%
ADRT 1 T SRt
ADR5 (1 0,9103) 0, 2 1 1,86% t 1
Some problems Default data

1. default definition • value based

– ratio between nominal value of the
– late payment, past due (how many days?) defaulted positions in a given class and the
bankruptcy, debt restructuring total value at the beginning of the period
2. Default data • value weighted (more weight to the default of
large firms, ok if the aim is to assess the loss)
3. Sample employed to estimate default • number based
rates – ratio between the number of the defaulted
positions in a given class and the total
number at the beginning of the period
• equally weighted (ok if the aim is to assess the
PD of a single firm)
• NB: aging effect (stickiness in the
migration prob.)
• through-the-cycle, point-in-time
Transition matrix: 1yr Rating validation: CRITERIA
• PD monotonically increasing when the ratin

class worsens
• PD increasing as time increases
• PD in a given class steady over time
• higest values on the main diagonal
• migration rates higer towards the nearest
classes
• debtors in default have been in bad classes
since a long time
Receiver Operating
Contingency table
Characteristic – ROC - curve
• Summarizes type 1 and type 2 errors

• Subjests (Ck) correctly classified (hit
rate Hk) with respect to defaulted ones
𝑁 𝑁
ℎ𝑖𝑡 𝑟𝑎𝑡𝑒 (D): Ck
𝑁 Hk
D
• Frequency of subjects erroneously
N4 N1 classified (Wk) (false alarm, Fk):
• Sensitivity Specificity
N2 N4 N1 N3
Wk
N3
Fk
• Alpha error N2 Beta error N D
N2 N4 N1 N 3

Gini o Cumulative Accuracy
ROC curve
Profile (CAP)
• mesaures the consistency between

model forecasts and actual PDs
• firms are ranked from the worst to the
best on the horizontal axis and firms in
defaul on the vertical axis
• Gini accuracy ratio B
G
A B
– the closer to 1, the more accurate the
model
CAP curve
Applications of
credit risk models

Agenda Why using credit risk models?
• Applications of credit risk measurement

models
– Loan pricing
Applications: Higher
– Estimating risk-adjusted profitability
PD
• proper loan rates profit
•Gain market share (new
• risk-adjusted performance management, RAPM LGD
clients avoiding future losses)
Costs Models Output
• more transparent
– Setting risk limits VaR value creation
• better risk control
– Bank’s loan portfolio optimatization VaR
• better portfolio mix
More
accurate
marginal models
551
© M. Anolli - Risk Management 550 © M. Anolli - Risk Management
Incorporating the expected

Loan pricing
loss
• We’ll start by
including the (1 i ) (1 r )(1 PD ) PD (1 r ) 1 LGD
expected loss
Cost of
funds
component only
• We’ll then add the 1 i
(1 r ) 1
unexpected loss 1 PD LGD
Expected
loss component
– Stand-alone VaR?
Loan rate
– Marginal VaR? i PD LGD i ELR
r
(spread)
Unexpected
• We’ll not take into 1 PD LGD 1 ELR
loss
account the
operating costs
– Already included in If i = 4%,
any analytical PD = 1%, 4% 0,5%
Operating
accounting systems LGD = 50%
r 4,522%
costs
(ELR = 0,5%)
1 0,5%
The expected loss spread Incorporating unexpected loss:
(KEL)
the risk premium on capital
• In any risky investment, unexpected losses

require the bank to hold a capital “cushion”
(CaR or VaR)
– Stand-alone VaR (per euro)
ELR (1 i) • Prudent as it does not take into account the
r i K EL capital saving coming from diversification
1 ELR
– Marginale VaR (per euro)
• minimal, as it gives to the last loan alla the
advantages coming from diversification given to
If i = ITR = 4%, the imperfect correlation with all the other loans
PD = 1%, 0.5% (1 4%) • Incorporating the unexpected loss component
LGD = 50% K EL 0.522% requires to remunerate the capital needed to
(ELR = 0.5%) 1 0.5% cover this UL
– The cost of equity capital includes both the cost
of debt (internal transfer rate – ITR) and a risk
premium for the shareholders
554
How does the loan rate changes when Let’s look at the previous
including the unexpected loss
example:
For the part of loan (e.g. 8 cents for each euro) corresponding to VaR
(funded with equity capital) we have to pay Ke – ITR:
4% 0.5% 8% (12% 4%)
r 5.166%
ITR ELR VaR ( Ke ITR) 1 0.5%
r
1 ELR
…or – equivalently – part of the loan is funded with debt capital at the ITR
(ex., 92 cents for each euro) and the rest, corresponding to VaR (ex., If we attribute to this loan the benefits of diversification:
8 cents per euro) with equity capital, at a cost of Ke > ITR:
More aggressive
pricing: 4% 0.5% 6% (12% 4%)
(1 VaR) ITR VaR Ke ELR marginal VaR r 5.005%
r (ex. 6%)
1 0.5%
1 ELR
The unexpected loss spread or Factors affecting risk-
economic capital spread (KEC)
adjusted loan rates:
VaR Ke ITR • Client’s probability of default (PD)

r i K EL K EC
1 ELR – higher the longer (maturity) the transaction
In our example, using marginal VaR, we get: • Loss given default rate (LGD)
– depends, for example, on the guarantees
6% 12% 4%
K EC 0.482% • Economic capital absorbed by the
1 0.5%
individual loan
Summing up: – may depend on correlations
r = 5.005%
• The internal transfer rate of the relevant
maturity (ITR)
i (= ITR) = 4% KEL = 0.522% KEC = 0.482%
cost of funding Expected loss Economic capital 558

• Cost of equity capital for the bank (Ke)
Let’s change perspective: risk-

Raroc: an example
adjusted profitability and Raroc
• In the previous example, the bank is able to

• If the bank is not completely free in deciding the impose a 5.005% rate and the shareholders
loan rate (price taker), then it may not be able to achieve the desired risk premium (8% = 12% -
charge Ke 4%).
• One then has to start from the market rate and • What about if the bank only charges a 4.8% rate?
estimate the risk.-adjusted return for the
4.8% 0.995 4% 0.5%
shareholders, expressed as a net premium Raroc 4.6%
NP 6%
Raroc • Note that, as
VaR
NP (1 r )(1 PD ) PD (1 r ) 1 LGD (1 ITR ) Raroc Ke TIT
... r (1 ELR ) EL ITR
…this loan would destroy value
r (1 ELR) EL ITR
Raroc
VaR
Risk limits systems Risk limits systems
• It is possible to give to each operating unit • The use of credit VaR as management tool
(ex. branch or group of branches) a degree of also requires to address some relevant
independence expressed in terms of maximum organizational and operating problems
VaR
– Defining the risk taking units (how much to
– Compared to nominal values, this criterium
recognizes the different risk of different loans aggregate?)
and exposures (PD, LGD, maturity, etc.) and – The frequency with which limits are revised
the diversification benefit coming from each over time
loan
– The individual risk-taking unit can use its VaR – The way VaR limits are taken into account
limit with long term loans to high quality within the bank’s budgeting process
companies or short term loans to more risky
companies or require more guarantees (lower
LGD) or diversify more, …
Optimizing the portfolio mix Optimizing the portfolio mix

• From Credit VaR to more efficient • In the future:

portfolios
– Reducing concentration risk – credit derivatives, loan sales and secondary
– Exploiting all diversification opportunities market for loans
• Problems: – Differentiating the two main bank
– Low liquidity and transparency of the functions: origination and risk management
loans market • The “sales department” looks for market
– High geographic and industrial opportunities and client relationships
concentration of regional banks’
portfolios • Risk management sells/exchange in the
– Application typically limited to portfolio market the credit risk and optimizes the
“rotation” portfolio

Risk MGT CRD Risk

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Risk MGT CRD Risk

Uploaded by

Copyright:

Available Formats

Agenda

Facoltà di Scienze Bancarie, Finanziarie e Assicurative

© M. Anolli - Risk Management Page: 275

The three (main) factors driving

EAD and LGD are • A. Establishing an appropriate credit risk environment

© M. Anolli - Risk Management 278 © M. Anolli - Risk Management 279

Credit risk guidelines: Basel PD estimation

© M. Anolli - Risk Management 280 © M. Anolli - Risk Management Page: 281

Discriminant analysis: Discriminant analysis:

Interest expenses over turnover

“Bad” firms show higher This firm clearly

Interest expenses over turnover

Discriminant analysis: Discriminant analysis:

Interest expenses over turnover

• The score can be seen as a “weighted • To estimate a probability of being good

From natural variables… …to a synthetic one (score)

D(x) 1 x1 2 x2 IET xIET OE xOE

is the variance/covariance matrix of the natural variables xi.

How to use the scores to

• Consider the data for 38 companies

x1: interest expenses

Similarly, we can compute the score for every company

Computing the cut-off

1 1 0.29 0.67 Classification

D(x i ) 6.09 xOF ,i 4.84 xSA,i 3.97

• Formally, D(x) minimizes the ratio

Costs associated Costs associated

The costs of the two mistakes can

to the -2% OK This alternate criterion still lends

2 collateral, the cost of a Company 15

The selection of the Discriminant analysis:

Linear probabilistic model Linear probabilistic model

• In the logit model, the linear relationship is adjusted

Example: logit with 3 variables Logit & probit

• Transformations alternative to the logistic

• Credit-scoring models attempt to identify the

Neural networks: Neural networks:

• The network is trained through an • Like neural networks, they require an

Genetic algorithms: Genetic algorithms:

• A new generation of algorithms is

• When the new generations do not add of a

Source: Pomante (1999)

• Models based on market data

© M. Anolli - Risk Management Page: 331

Using bond spreads Corporate bond spreads: let’s

– For some (large) corporate borrowers, some

Equilibrium spread given a PD Hp: p=1%

Example: r=4%, d=1%, k=50% If k>0 d

Corporate bond spreads Corporate bond spreads

Let’s consider these data

5 4.50% 6.20% 1.70%

© M. Anolli - Risk Management

Forward rates Spreads

Higher forward spreads positive slope of the spread

Corporate bond spreads Corporate bond spreads

d 5.4% 4.2% 0.012

© M. Anolli - Risk Management

0 p2 1 0 s2 1 95.86% 4.14% 5 4,66% 95,34% 84,97% 15,03%

© M. Anolli - Risk Management

Corporate bond spreads Using corporate bond data

Advantages • We can also use continuous rates. This means

Spreads may also reflect other factors:

• If k>o, we get • When we look at longer time horizons

• Expected loss is given by PDLGD, that is, PD(1-