Scoring bank loans that may go wrong: a case study

A bank employs logistic regression with state-dependent sample selection to identify loans that may go wrong. The data consist of some 20 000 loans for which a number of conventional accounting ratios of the debtor firm are known; after two years just over 600 have gone wrong. Inspection shows that the state-dependent sampling technique does not work because the data do not satisfy the standard logit model. Several variants on this model are considered, and it is found that a bounded logit with a ceiling of (far) less than 1 fits the data better. When it comes to their performance in an independent data-set, however, the differences between the various methods of analysis are negligible.