Superscorecards

Scorecards used in credit scoring are based on a simple linear model, in which each characteristic contributes a weight to the final customer score. This simple form has several important merits, including interpretability and ease of explanation. As a consequence of this, a large base of software and user expertise has accumulated for such models. Any alternative to this simple form should have comparable simplicity, and should ideally not require the replacement of the existing software and knowledge bases. We describe such an alternative, based on multiplying together two (or more) standard scorecards. Classifying a new customer using this superscorecard simply requires finding the customer's score on two standard scorecards, using existing software, and multiplying the results together. The standard model is a special case of the superscorecard so that, leaving aside issues of sampling variation and overfitting, the superscorecard cannot yield inferior results. It gains its power by a simple iterative estimation procedure, in which the parameters of the component scorecards are estimated simultaneously. We describe this algorithm and illustrate the comparative performance of logistic regression and the superscorecard on a set of current account data.