Oliver, Wyman & Company 1 December 1999

Credit portfolio management
odern credit risk management techniques were initiated by the banking industry’s desire to avoid a repeat of its late ‘80s and early ‘90s default experience. The heavy credit losses during this period, driven by a poorly controlled rush to build market share at the expense of asset quality and portfolio diversification, threatened the solvency of even well capitalised institutions. The need to better understand portfolio credit risks was reinforced by the publication of the Bank for International Settlements’ (BIS) capital adequacy guidelines in 1988. These guidelines, whilst specifying minimum regulatory capital requirements, were inadequate to provide an accurate measure of the risk/reward characteristics of a credit portfolio. Banks therefore started to develop more sophisticated credit risk management techniques that recognised both the credit risk of individual exposures and the degree to which

By Thomas Garside, Henry Stott and Anthony Stevens

these risks were diversified. Banks leading the development of credit risk management techniques quickly discovered that credit pricing was highly inefficient. Typically pricing within a loan portfolio would be almost flat across the credit risk spectrum, generating huge skews in customer profitability. Initial efforts focused on mitigating these skews by calculating riskadjusted profitability (eg risk adjusted return on [risk-adjusted] capital) by sub-portfolio and then using these measures to create riskadjusted loan pricing tools. Leading

banks thus started to rationalise pricing in both loan and bond portfolios, and moving under-performing assets off their balance sheets. Consequently banks that had not developed risk-adjusted performance measures started to suffer from negative selection, often accepting significantly underpriced assets from more sophisticated institutions. In parallel to developing aggregate risk-adjusted performance measures, leading banks were also starting to quantify credit risk at finer levels of detail. Credit portfolio models were developed which could differentiate credit risk along multiple dimensions (credit grade, industry, country/region etc) and, for large corporate exposures, on a name-by-name basis. These credit portfolio models have positioned leading institutions to take advantage of the increasing liquidity of the credit markets and to adopt a far more active approach to credit portfolio management than was previously possible. Historically, credit portfolio management had focused on the monitoring of exposure by broad portfolio segment and, if necessary, the imposition of exposure caps. The creation of a stand-alone credit portfolio management function, armed with sophisticated portfolio models and with a controlling mandate over assets held on the balance sheet, now enabled the credit portfolio to be optimised independent of origination activity, (figure 1). Active credit portfolio optimisation has enormous poten-

1. Optimisation of origination and portfolio management activities
Line of business
product and delivery optimisation

Banks leading the development of credit risk management techniques discovered that credit pricing was highly inefficient’

Portfolio management
mark-to-market credit portfolio transfer of Pr (loss) assets
optimisation of syndication/ sales loss distribution

origination opportunities

sales/product teams



asset syndication/ disposals asset purchases asset swaps credit derivatives



Credit portfolio methodology I Measuring credit risk correlations. for example. Concentration describes the ‘lumpiness’ of the credit portfolio (eg why it is more risky to lend £10m to 10 companies than to lend £0. Effects of concentration and correlation on credit risk credit risk concentration of portfolio specific risk: driven by concentration II. two alternative approaches can be used when valuing the portfolio: G Loss-based method. In general. NPV-based methods are most applicable to bond portfolios and large corporate portfolios where meaningful markets exist for either the physical assets or credit derivatives. and if it downgrades it is assumed to be worth less than par. The different credit risk profiles generated for the same portfolio using loss-based and NPV-based methods are shown later in this article (figure 8). to accurately model portfolio credit risk the correlation between exposures must first be measured. NPV-based methods are most applicable to bond portfolios and large corporate portfolios where meaningful markets exist for either the physical assets or credit derivatives’ 3.000 companies). Under this approach. Using this approach credit migration has no 2. G NPV-based method. Volatility of portfolio losses is driven by two factors – concentration and correlation (figure 3). However. Because expected losses can be anticipated. If the obligation upgrades then it is assumed to be worth more than par. where such markets do not exist a more meaningful risk profile is obtained using a loss-based method. Correlation describes the sensitivity of the portfolio to changes in underlying macro-economic factors (eg why it is more risky to lend to very cyclical industries such as property development). This seemingly simple statement conceals the complex string of calcula1 December 1999 2 correlation of borrower behaviour systematic risk: driven by correlation size of portfolio 02 . Loss-based calculations have the advantage of requiring less input data (margin and maturity information. The value of the obligation can be calculated using either using market credit spreads (where applicable) or by marking-to-model using CAPM or similar method. the embedded value of an exposure is assumed to be realisable. As discussed previously.1m to 1. many institutions are starting to run both methods in parallel. Under this approach an exposure is assumed to be held to maturity. When quantifying credit risk. they should be regarded as a cost of doing business and not as a financial risk. The exposure is therefore either repaid at par or defaults. particularly for portfolios where securitisation is possible. Credit risk is conventionally defined using the concepts of expected loss (EL) and unexpected loss (UL) (figure 2). Definition of credit loss volatility credit losses credit losses effect on the book value of the obligation.ERisk. there being substantial volatility (unexpected loss) about the level of expected loss. correlation effects will dominate. EL time (years) frequency ‘ 1 diversification of credit risk In general. It is this volatility that credit portfolio models are designed to quantify. Obviously credit losses are not constant across the economic cycle.com tial to enhance profitability. In all but the smallest credit portfolios. is not required) and being simpler to compute. and thus worth the recovery value of any collateral. I Credit risk measurement framework. For the vast majority of commercial bank exposures. Using only very basic optimisation techniques a typical institution might expect to reduce the economic capital consumed by its credit portfolio by 25%–30%.

ERisk. callable liabilities expected default frequency probability distribution of future asset value callable liabilities. Indeed. The Merton model assumes that a firm will default if.10 0.5 -1 -0. as this would require repeated observations over a given time-period during which each company would either default or survive.05 0 -0. Unfortunately this approach is unsuitable except for the most basic of portfolio analysis for two main reasons. with insufficient data on underlying credit risk rating. equity and other derivative desks are positioned to take advantage of resulting arbitrage 0.9 asset value 2 1 defaults asset value 1 2 defaults 6. Asset correlations have the benefit of being more easily observable (from equity prices.15 0. if macro-economic factors are chosen. volatility of asset values 1. The Merton model has also been successfully adapted to describe credit risk correlations in financial institution portfolios that contain corporate exposures. In figure 5 the more correlated the movements in the two companies’ assets the greater the ‘twist’ in the joint asset value distribution. to measure default correlation (as required for loss-based measures) between two companies is impossible.15 -0.20 -0.5 3 3.5 2 2. and then transforms them into a default probability. over a 12month period.00 -1.14 0. Clearly such analysis is impossible in practice. current market value of assets 3.5 0. Not only could a credit portfolio manager potentially hedge credit risk via equity or ‘macro-economic’ derivatives. of enabling intuitive stress testing and scenario analysis of the credit portfolio.5 4 -0. Factor models usually produce better prospective correlation estimates than direct observation and have the additional benefit.3 aggregate time series to infer credit risk correlation. using aggregate time series produces unstable results over time. Firstly.7 2. An example of a macro-economic factor model is shown in figure 6. A more attractive solution to calculating credit risk correlation is to use a causative default model that takes more observable financial quantities as inputs. Similar difficulties exist when trying to estimate correlation between changes in credit rating or bond spreads.05 -0.9 0.5 -1. The simplest solution is to use Probability % market value 2.9 3.3 2. fall and ultimately default together. The Merton default model tions that are actually necessary.10 0. This enables asset correlation to be transformed into credit risk correlation (figure 5). the market value of assets falls below the value of 1 and 2 default 3. The ‘connection’ of credit risk to underlying macroeconomic risk factors has significant implications for credit risk management and the future development of credit markets. industry and geographic distribution of the portfolio. balance sheet analysis etc) and their correlations have been shown to be stable over time. Complexity arises as it is extremely difficult to calculate credit risk correlations directly.02 0. Hence the greater the probability that the credit quality of the two firms will rise.com 4. The most widely used model for commercial lending portfolios being the Merton default model (figure 4).04 0.10 -0. Joint probability density function for company asset values 0.20 0.06 0. Illustrative macro.08 0.20 0.3 -0.5 1 1.5 0 0.economic factor model average factor weight using factor models in the same way that an equity ‘beta’ is estimated. but professional market-makers should ensure that credit.25 equity index exchange rate index interest rates unemployment property index 1 December 1999 03 .16 0. The correlation of model inputs themselves are best measured 1 year time 5. Credit risk correlation could then be calculated from the number of times both companies defaulted simultaneously. Secondly.1 1. aggregate time series are usually available only at a very high level.12 0.18 0.

98% NPV-based -9% -8% -7% -6% -5% -4% -3% -2% -1% 0% % change in loss. from which its credit risk profile can be calculated (figure 7). Wyman & Co approach (and other proprietary models) Disadvantages complexity in portfolio model calculations G Data availability G Additional of intuitive understanding of underlying factors G Applicability to non-quoted companies/retail segments G Ability of factor structure to describe correlation effects (especially for non-corporate exposures) G Loss-based calculation only G Lack of macro-economic insights G Performance drop-off when specifying multiple portfolio segments G Loss III. Summary of credit portfolio models Risk measure Advantages G Intuitive understanding of G NPV-based underlying factors G Applicable to all portfolio segments Macro-economic factor models G Links to economic forecasting models and provisioning G Ease of stress testing/ scenario analysis KMV Portfolio Model G Loss-based G Simplified portfolio G NPV-based model calculations Orthogonal factor sets based on G Data availability for equity prices quotedcompanies JP Morgan CreditMetrics/ G NPV-based G Data availability CreditManager G Easy to understand factors Equity indices CSFP CreditRisk+ G Loss-based G Simplicity G Speed of calculation “Market risk” volatility approach – no factor structure G Loss-based Model/factor Oliver. A lossbased example of such an analysis is shown in figure 8 where. with a corresponding fall in credit quality and increase in default rate. I Credit risk concentrations and portfolio optimisation.8% of total exposure is required. I Simulation methods.ERisk.00% 1 December 1999 04 .00% 99.02% default probability). I Summary of credit portfolio models. Conversely. •factor weights %change in value. In figure 6 a positive factor weight indicates that a positive change in that factor produces an increase in asset value. Portfolio model applications Having discussed the inner workings of credit portfolio models we can now illustrate their uses by examining a number of management applications. These models are summarised in table A.com 7. % change in NPV 2% 99. to achieve a Aa1/AA+ credit rating (equivalent to a 0. There are a number of currently available credit portfolio models that are distinguished by their correlation structures and choice of risk measure.% losses loss factor-to-factor correlations 10% 5% –5% 0% %change in value. This is calculated from the tails of the credit risk distribution by determining the probability that a reduction in portfolio value exceeds a critical value. Whilst the risk of small credit portfolios can be calculated analytically. I Solvency analysis. The most obvious application of a credit portfolio model is to calculate economic capital.98% loss-based 99. a negative factor weight indicates that a positive change in that factor produces a decrease in asset value. Monte Carlo simulation is the standard method. Assessment of economic capital requirement loss expected NPV NPV 1% expected loss 99. Breaking 8. Monte Carlo simulation Monte Carlo simulation input datasets borrower/cluster details • exposure •collateral •margin •rating simulated portfolio loss and value distribution probability density scenario 1 DEquity = –20% DFX =+5% • • DUnemp=+2% scenario 2 scenario 3 • • portfolio response to scenario probability density change in value loss opportunities.% losses A. the large number of calculations required mean that for most portfolios it is better to employ a numerical simulation technique. These developments are likely to be a major driver of liquidity as these markets develop.% losses • • • change in value scenario N* DEquity =+15% DFX =+0. and can be thought of as a ‘state-of-the-world generator’ that generates all possible states of the economy and the resulting impact on the value of the credit portfolio. economic capital equivalent to 7. with a corresponding rise in credit quality and reduction in default rate.5% probability density change in value loss %change in value. In this way a distribution of all possible portfolio values is built up.5% • • DUnemp=–0.

Potential diversification benefits in a typical bank portfolio economic capital (£bn) 6% exposure by counterparty 10m economic capital 50m economic capital Economic capital as % of exposure 70 60 50 40 30 20 10 0 0 100 200 300 400 500 600 Exposure (millions) 700 800 reduction in reduction in name £5. An example of portfolio behav- iour under stress-test scenarios is shown in figure 12.7% of outstandings implied probability of scenario consistent with desired solvency standard * = 0.98%) solvency standard Property Crash Scenario Probability that losses exceed 7. banks must ensure that they understand the economic value of their portfolios and how this value can be maximised through efficient credit portfolio management.90% 1. simple optimisation techniques will substantially reduce economic capital requirements – typically reductions of 30% are achievable equivalent to annual savings of £288m (assuming a capital charge of 18%) for a portfolio of £100bn (figure 10).02% ∏1. Conclusion This article has described the underlying theory of credit portfolio models and illustrated their value in making more effective management decisions. I Sensitivity analysis and stress testing.6bn concentration correlation specific risk 4% systematic risk £4.05%* down the aggregate credit risk distribution to show the credit risk of each portfolio element allows risk concentrations and hence diversification opportunities to be identified (figure 9). Portfolio models can be used to calculate expected loss rates under different economic scenarios and thus drive dynamic provisioning estimates or loan loss reserving methodologies such as the SBC ACRA reserve. Stress testing credit portfolios Assessment of capital adequacy under property crash scenario Base Case: Economic capital AA+ (99.ERisk. The sensitivity of portfolio credit losses to changes in the underlying macro-economic risk factors can also be examined to determine whether a hedging strategy might be possible (figure 11). 1 December 1999 ‘ Banks must ensure that they understand the economic value of their portfolios and how this can be maximised through effective management’ 05 .7% 1. IV. Loan loss sensitivity 12.0bn 2% 0% existing portfolio diversified portfolio 11. Credit risk concentrations 10.com 9. the possibility of active credit portfolio management will increase dramatically and result in a fundamental shift in the way banks both originate and hold credit assets.90% 7. For most credit portfolios. In order to benefit from these new opportunities. An extension of this application is to use the models for ‘stress-testing’ to estimate possible changes in portfolio value conditional on extreme macro-economic scenarios. With the rapidly growing market in credit derivatives and portfolio securitisations.