Survival analysis and the credit-granting decision

Abstract

Abstract Credit-scoring systems have traditionally been built on cases which are classified into a number of states (usually two, goods and bads). Associated with each case are a number of measures or predictors (e.g. age, occupation. repayment history of previous loans, etc). The predictor variables are grouped into bands. for example, the variable ‘residential status could be grouped into the three bands ‘owner. ‘tenant ‘. and ‘others. An indicator variable (0, 1) is created for each band. The problem is now to construct a weighted sum of the indicator variables such that this sum or score distinguishes good from bad cases. Methods used to solve this problem include discriminant analysis, multiple regression analysis. and iterative techniques which maximize some ‘information function. More recently experts systems and decision trees have also come into the picture. Clearly the ability to predict which individuals from a set of applicants are likely to be bad is of great value. However, when a case is classified as bad no regard is taken of the length of time the account has been on the books. I think we would all agree that there is a relationship between the probability of an account being classified as bad and the time that the account has been on the books. In this chapter an attempt is made to model the ‘time taken for an account to be classified as bad. I suspect that we would be more interested in identifying those accounts that make a significant number of payments before being classified as bad. What is classified as significant will depend on a number of factors, for example, interest rates, type of loan. cost of administering each account. etc. We will use the methods of Survival Analysis to try and construct this model.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 35 citations exist upstream.