The degradation of the scorecard over the business cycle

The published literature on credit scoring has not compared the characteristics of those who default, nor the discriminating power of individual variables used to predict default, under different economic conditions. Similarly, scorecards constructed by credit-scoring agencies are estimated from data relating to two or three consecutive years for applications over three to five years before. The aim of this paper is to explore the effects of changes in a scoring function over time on the classification of applicants into those likely to default and those not likely to default. Linear discriminant analysis is applied to a training sample of 26,043 applicants for a bank credit card to estimate empirically a model of their repayment behaviour in 1989 and 1990. The variables that have additional statistically significant discriminating power over others are broadly similar between the two years, although some differences exist. Using a holdout sample of 17,084 cases which are thought to be representative of a profile of applications to the data-supplying organization, we cross-tabulate the number who would be accepted and rejected using the 1989 model with the corresponding predictions using the 1990 model. The characteristics of those who would be accepted using the 1989 model but rejected using the 1990 model are identified. Differences in the predicted classification of a case may be due to differences between the two years in the functions estimated and/or to difference in the prior probabilities of default. We consider the proportion of applicants who would be accepted in one year but not in the other, if the prior probabilities are adjusted to give the same rejection rate in both years, and discuss their characteristics.