| Abstract: | Scorecards used by consumer credit providers to assess the probability that an applicant will default are usually built for the population of potential applicants as a whole. This paper investigates whether it is permissible and worth-while to build a separate scorecard for each subpopulation of applicants. We review the legal requirements to find that it is permissible to use separate scorecards for many, but not all, personal characteristics. Second, using data supplied by a credit card organization separate scorecards were built for several subpopulations for each of twelve personal characteristics. The predicted performance of each was compared with that gained form estimating a scorecard for the full population using three methods for setting the cut-off scores in an `independent' way. These methods differ in the degree to which the cut-off scores are independent of information about other subpopulation, in the level of discrimination achieved between likely good payers and defaulters and in the degree to which each method is robust to new data. We conclude, first, that creating scorecards using subpopulations does not necessarily give better discrimination between likely good payers and defaulters. Second, none of the three methods examined to set the cut-off scores dominates the others using the three desirable properties described; trade-offs are required. Finally, subpopulation scorecards lead to the rejection of fewer applicants than scorecards built on full populations. |
