GENERALIZED BANKRUPTCY MODELS APPLIED TO PREDICTING CONSUMER-CREDIT BEHAVIOR

Although it is well known that many financial variables do not behave according to the assumptions of commonly used statistical methodologies, it has been generally assumed that the robustness of these methodologies assures adequately accurate results. Recent work in the econometrics literature has explored the use of estimation methodologies that are appropriate for variables that do not have symmetric or normal distributions. This article explores the application of the EGB2 family of distributions to the problem of predicting bankruptcy and offering consumer credit. The distributions of the forecasts of good and bankrupt accounts violate many of the assumptions of the multiple discriminant analysis methodology used in most commercial bankruptcy forecasting models. The importance of the criterion used to classify good and bankrupt accounts is also considered. Since there is a positive payoff to the correct classification of good accounts, minimizing the number or expected costs of misclassifications does not maximize the expected economic return. The theoretical development is demonstrated on a data base of more than 16,000 observations of demographic and credit variables of bankrupt and solvent households. The performance of the various classification models is tested on a holdout sample of more than 11,000 observations. Although the generalized distribution is found to be statistically superior to discriminant analysis, the managerial importance of this superiority is seen to vary with the characteristics of the bankruptcy forecasting problem.