3 Credit Assessment Models

Main categories of credit assessment models:

1. Heuristic models
2. Statistical models
3. Causal models

Heuristic models:
Heuristic models aim at documenting and leveraging previous experience shedding insight on the creditworthiness of a customer. The bases of this experience are the following:

• subjective experience and observations
• conjectured business interrelationships
• business theories

The main kinds of heuristic models are the following:

1. Rating questionnaire
2. Qualitative systems (e.g. BVR-I of the Federal Association of German Cooperative Banks)
3. Expert systems (e.g. CODEX, Commerzbank Debitoren Experten System)

Statistical models:
Statistical models aim at verifying hypotheses concerning potential creditworthiness criteria to using statistical methods on a sample of debtors representing a segment of the banks customers or potential customers. These hypotheses are statements as to whether on average a given recorded variable can be expected to be higher or lower for insolvent borrowers than for solvent borrowers.

The ability of a statistical model to correctly identify solvent and insolvent borrowers depends on the following:

• availability of a large enough data set to make statistically relevant inference,
• representativeness of the sample data set in relation to the target population.

The main kinds of statistical models are the following:

• Multivariate discriminant analysis (MDA)
• Regression models

Multivariate Discriminant Ananlysis
Objective: To distinguish solvent and insolvent borrowers as accurately as possible using a scoring function dependig on several independent creditworthiness criteria, or observed variables. Linear MDA produces a linear function (discriminant function) built as a linear combination of the values of the creditworthiness criteria.



A cutoff point is determined to discriminate likely defaulters from likely non-defaulters. This cutoff point determines a hyperplane ideally cutting through the n-dimensional space of independent criteria in such a way as to separate as well as possible the defaulters from the non-defaulters.

MDA requires the observed variables to be normally distributed. If the values of a particular variable are not normally distributed (as would happen with an ordinal qualitative criterion), then they must be rescaled and normalized. Further, MDA requires the groups to be discriminated to have the same variance/covariance matrices. In practice, however, this requirement is less important.

Regression analysis
Regression analysis as applied to the assessment of creditworthiness models the dependence of a binary variable (default, non-default) on a set of other independent variables. The advantage of regression analysis over MDA is that under some conditions it can be used to directly compute estimates of membership probabilities. That is, the regression function can be built to directly compute the probability of default; whereas, MDA models always explicitly need a calibration phase to produce estimates of the probability of default.

The two most common methods of regression analysis on binary dependent variables are the following:

• Logit regression
• Probit regression

Logit regression and probit regression differ on account the non-linear transformation function used to model the probability distribution of the binary variable. The first method uses the logistic funciton; whereas, the second uses the probit function.

The advantages of logistic regression over the multivariate discriminant analysis method are the following:

• Logistic regression does not require the independent variables to be normally distributed.
• The result of the logistic regression can be directly interpreted as the probability of group membership.
• Logistic regression tends to be more robust than MDA because of the lower demands it makes on input data.
• All else equal, results of a logistic regression model tend to be more accurate than those provided by the MDA method.

Neural networks
Neural networks are artificial simulations of the way the brain works when categorizing concepts. These software devices are able to process effectively any kind of data, both quantitative and qualitative, without reference to the distribution on input variables. Classification results tend to be good; however, this method yields a "blak box" which cannot be easily interpreted from an economic theoretic point of view. For this reason they are still relatively uncommon as bases of rating systems.