A rule of thumb for reject inference in credit scoring

Credit scoring models are usually developed using the accepted Known Good-Bad applicants, called KGB model. Yet, the KGB model does not represent the entire Through-The-Door population. Reject inference attempts to correct this inherent flaw by using information of the rejected accounts. Augmentation methods are widely used methods of reject inference, among which Fuzzy Augmentation is the most accurate one. In this paper, we first establish an important property of Fuzzy Augmentation: If Fuzzy Augmentation is not incorporated with variable re-selection, it will produce the same results as the KGB model. We then propose a rule of thumb for Augmentation methods. Based on this rule of thumb, we present a two-phase Augmentation. This two-phase method works not only for Machine Learning in Python but also for the traditional approach using SAS. Moreover, it is user friendly in that the user can specify a factor to increase the bad rate of rejected accounts.