| Summary: | The term <i>credit scoring</i> refers to the application of formal statistical tools to support or automate loan-issuing decision-making processes. One of the most extended methodologies for credit scoring include fitting logistic regression models by using WOE explanatory variables, which are obtained through the discretization of the original inputs by means of classification trees. However, this Weight of Evidence (WOE)-based methodology encounters some difficulties in order to model interactions between explanatory variables. In this paper, an extension of the WOE-based methodology for credit scoring is proposed that allows constructing a new kind of WOE variable devised to capture interaction effects. Particularly, these new WOE variables are obtained through the simultaneous discretization of pairs of explanatory variables in a single classification tree. Moreover, the proposed extension of the WOE-based methodology can be complemented as usual by balance <i>scorecards</i>, which enable explaining why individual loans are granted or not granted from the fitted logistic models. Such explainability of loan decisions is essential for credit scoring and even more so by taking into account the recent law developments, e.g., the European Union’s GDPR. An extensive computational study shows the feasibility of the proposed approach that also enables the improvement of the predicitve capability of the standard WOE-based methodology.
|
|---|
