3.8 DEVELOPMENT OF NEW APPROACHES TO CREDIT RISK MANAGEMENT
3.8.3 Traditional Credit Risk Measurement Approaches
3.8.3.1 Background
It is hard to draw a clear line between traditional and new approaches, as many of the superior concepts of the traditional models are used in the new models. For the purposes of this historical review, the traditional credit models are segregated into three types: expert systems, rating sys- tems, and credit-scoring systems.17
3.8.3.2 Expert Systems
In an expert system, the credit decision is made by the local or branch credit officer. Implicitly, this person’s expertise, skill set, subjective judg-
ment, and weighting of certain key factors are the most important deter- minants in the decision to grant credit. The potential factors and expert systems a credit officer could look at are infinite. However, one of the most common expert systems, the “five Cs” of credit, will yield sufficient un- derstanding. The expert analyzes these five key factors, subjectively weights them, and reaches a credit decision:
• Capital structure. The equity-to-debt ratio (leverage) is viewed as a good predictor of bankruptcy probability. High leverage suggests greater probability of bankruptcy than low leverage, as a low level of equity reduces the ability of the business to survive losses of income.
• Capacity. The ability to repay debts reflects the volatility of the borrower’s earnings. If repayments on debt contracts proves to be a constant stream over time, but earnings are volatile (and thus have a high standard deviation), the probability is high that the firm’s capacity to repay debt claims is at risk.
• Collateral. In the event of a default, a lender has a claim on the collateral pledged by the borrower. The greater the proportion of this claim and the greater the market value of the underlying collateral, the lower the remaining exposure risk of the loan in the case of a default.
• Cycle/economic conditions. An important factor in determining credit risk exposure is the state of the business cycle, especially for cycle-dependent industries. For example, the infrastructure sectors (such as the metal industries, construction, etc.) tend to be more cycle dependent than nondurable goods sectors, such as food, retail, and services. Similarly, industries that have exposure to international competitive conditions tend to be cycle sensitive.
Taylor, in an analysis of Dun and Bradstreet bankruptcy data by industry (both means and standard deviations), found some quite dramatic differences in U.S. industry failure rates during the business cycle.18
• Character. This is a measure of the firm’s reputation, its willingness to repay, and its credit history. In particular, it has been established empirically that the age factor of an organization is a good proxy for its repayment reputation.
Another factor, not covered by the five Cs, is the interest rate. It is well known from economic theory that the relationship between the interest-rate level and the expected return on a loan (loss probability) is highly nonlinear.19At low interest-rate levels, the expected return could in- crease if rates are raised. However, at high interest-rate levels, an increase in rates may lower the return on a loan, as the probability of loss increases.
Credit Risk 155
This negative relationship between high loan rates and expected loan re- turns is due to two effects: (1) adverse selection and (2) risk shifting. When loan rates rise beyond some point, good borrowers drop out of the loan market, preferring to self-finance their investment projects or to seek equity capital funding (adverse selection). The remaining borrowers, who have limited liability and limited equity at stake—and thus lower rat- ings—have the incentive to shift into riskier projects (risk shifting). In up- side economies and supporting conditions, they will be able to repay their debts to the bank. If economic conditions weaken, they will have limited downside loss from a borrower’s perspective.
Although many financial institutions still use expert systems as part of their credit decision process, these systems face two main problems re- garding the decision process:
• Consistency. What are the important common factors to analyze across different types of groups of borrowers?
• Subjectivity. What are the optimal weights to apply to the factors chosen?
In principle, the subjective weights applied to the five Cs derived by an ex- pert can vary from borrower to borrower. This makes comparability of rankings and decisions across the loan portfolio very difficult for an indi- vidual attempting to monitor a personal decision and for other experts in general. As a result, quite different processes and standards can be applied within a financial organization to similar types of borrowers. It can be ar- gued that the supervising committees or multilayered signature authorities are key mechanisms in avoiding consistency problems and subjectivity, but it is unclear how effectively they impose common standards in practice.20 3.8.3.3 Rating Systems
One of the oldest rating systems for loans was developed by the U.S. Of- fice of the Comptroller of the Currency (OCC). The system has been used in the United States by regulators and bankers to assess the adequacy of their loan loss reserves. The OCC rating system allocates an existing loan into five rating buckets: four low-quality ratings and one high-quality rat- ing. In Table 3-4, the required loss reserve appears next to each category.
Over the years, the financial institutions have extended and en- hanced the OCC-based rating system by developing internal rating sys- tems that more finely subdivide the pass/performing rating category.
The OCC pass grade is divided into six different categories (ratings 1 to 6). Ratings 7 to 10 correspond to the OCC’s four low-quality loan rat- ings. These loan-rating systems do not exactly correspond with the bond- rating systems, especially at the lower-quality end of the spectrum (see Section 2.7.2.4 for a further discussion of bond-rating systems). One rea- son is the different focus of the approaches: loan-rating systems are sup-
posed to rate an individual loan (including its covenants and collateral backing). Bond-rating systems are more oriented toward rating the over- all borrower. This gap of one-to-one mapping between bond and loan rat- ing methodologies raises a flag as to the merits of those newer approaches that rely on bond data (spreads, transition matrices, etc.) to value and price loans individually and in a portfolio context.
Given this trend toward finer internal ratings of loans, compared to the OCC’s regulatory model, the 1998 Federal Reserve System Task Force Report21 and Mingo22 give some tentative support for using an internal model ratings-based approach as an alternative to the OCC model, to cal- culate capital reserves against unexpected losses, and loan loss reserves against expected loan losses. For example, using the outstanding dollar value of loans in each internal rating class (1 to 10), a bank might calculate its capital requirement against unexpected loan losses as follows:
total class 1 loans ⋅0.2%
Capital requirement =
Σ 冤total class 10 loans ⋅+⯗+ 100%冥
(3.1)
The 0.2 percent for rating class 1 is just suggestive of unexpected loss rates and should be based on historic loss probabilities of a loan in class 1 moving to class 10 (loss) over the next year.23However, an important prob-
Credit Risk 157
T A B L E 3-4
Example for Loss Reserves Based on Rating System
Rating Bucket Loss Reserves, %
Low-quality ratings
Other assets especially mentioned (OAEM) 0
Substandard assets 20
Doubtful assets 50
Loss assets 100
High-quality rating
Pass/performing 0
NOTE: From a technical perspective, the 0 percent loss reserves for OAEM and pass loans are lower bounds. In practice, the reserve rates on these categories are determined by the bank in consultation with examiners and auditors, depending on some type of historical analysis of charge-off rates for the bank.
SOURCE: U.S. Office of the Comptroller of the Currency, EC-159 (rev.), December 10, 1979, www.occ.treas.gov/ftp/release.
lem remains, similar to the current 8 percent risk-based capital ratio of the BIS approach—the diversification in the loan portfolio is not considered.
The credit risks of each rating class are simply added up to calculate a total capital requirement.
3.8.3.4 Credit-Scoring Systems
Credit-scoring approaches can be found in virtually all types of credit analysis. The basic concept is generally the same: certain key factors are preidentified. They determine the loss probability of default and the recov- ery rate (as opposed to repayment), and they are combined or weighted into a quantitative score schema. The score can be literally interpreted as a loss probability of default. In other scoring systems, the score can be re- garded as a classification system: it allocates a potential or existing bor- rower into either a good group (higher rating) or a bad group (lower rating), based on a score and a cutoff point. Full reviews of the traditional approach to credit scoring, and the various methodologies, can be found in Caouette, Altman, and Narayanan24 and in Saunders.25 See Altman and Narayanan for a good review of the worldwide application of credit- scoring models.26One simple example of this new credit risk model type should cover the key issues supposedly addressed by many of these newer models. The Altman Z-score model is a classification model for corporate borrowers and can also be used to get a default probability prediction.27 Based on a matched sample by year, size, and sectors of defaulted and sol- vent firms, and applying the linear discriminant analysis, the best-fitting scoring model for commercial loans results in the following equation:
Z=1.2 ⋅X1+1.4 ⋅X2+3.3 ⋅X3+0.6 ⋅X4+1.0 ⋅X5 (3.2) where X1=working capital/total assets ratio
X2=retained earnings/total assets ratio
X3=earnings before interest and taxes/total assets ratio X4=market value of equity/book value of total liabilities ratio X5=sales/total assets ratio
If a corporate borrower’s accounting ratios Xi,weighted by the esti- mated coefficients in the Z function, result in a Zscore below a critical value (in Altman’s initial study, 1.81), the borrower would be classified as
“insufficient” and the loan would be refused.
A number of issues need to be discussed here. First, the model is lin- ear, whereas the path to bankruptcy can be assumed to be highly nonlin- ear, and the relationship between the Xi values itself is likely to be nonlinear. A second issue is that, with the exception of the market value of equity term in the leverage ratio, the model is essentially based on ac- counting ratios. In most countries, standards require accounting data only
at discrete intervals (e.g., quarterly) and are generally based on historic- or book-value accounting principles. It is also questionable whether such models can capture the momentum of a firm whose condition is rapidly deteriorating (e.g., as in the Russia crisis of October 1998). As the world becomes more complex and competitive, and the decision flow becomes faster, the predictability of simple Z-score models may worsen. Brazil of- fers a good example. When fitted in the mid-1970s, the Z-score model did a quite good job of predicting default even two or three years prior to bankruptcy.28However, more recently, even with low inflation and greater economic stability, this type of model has performed less well as the Brazilian economy has become more open.29
The recent application of nonlinear methods (such as neural net- works) to credit risk analysis shows potential to improve on the proven credit-scoring models. Rather than assuming there is only a linear and di- rect effect from the Xivariables on the Zcredit score (or, in the language of neural networks, from the input layer to the output layer), neural net- works allow for additional explanatory power via complex correlations or interactions among the Xivariables (many of which are nonlinear). For ex- ample, the five variables in the Altman Z-score model can be described by some nonlinearly transformed sum of X1and X2as a further explanatory variable.30 In neural network terminology, the complex correlations among the Xivariables form a “hidden layer” which, when exploited (i.e., included in the model), can improve the fit and reduce type 1 and type 2 errors. (A type 1 error consists of misjudging a bad loan as good; a type 2 error consists of misjudging a good loan as bad.)
Yet, neural networks pose many problems for financial economists.
How many additional hidden correlations should be included? In the language of neural networks, when should training stop? It is entirely possible that a large neural network, including large Nnonlinear trans- formations of sums of the Xivariables, can reduce type 1 and type 2 errors of a historic loan database close to zero. However, as is well known, this creates the problem of overfitting—a model that well explains in-sample data may perform quite poorly in predicting out-of-sample data. More generally, the issue is when does one stop adding variables—when the re- maining forecasting error is reduced to 10 percent, 5 percent, or less? Re- ality might prove that what is thought to be a global minimum forecast error may turn out to be just a local minimum. In general, the issue of eco- nomic meaning is probably the most troubling aspect of financial inter- pretation and use. For example, what is the economic meaning of an exponentially transformed sum of the GARCH-adjusted sales to total as- sets and the credit-spread-adjusted discount factor ratio? The ad hoc eco- nomic nature of these models and their tenuous links to existing financial theory separate them from some of the newer models that are discussed in the following chapters.
Credit Risk 159
3.8.4 Option Theory, Credit Risk, and the KMV Model