A Study of Reject Inference Techniques
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Credit Scorecard models are commonly built using data available within anorganization's transactional database. Such data, however, will only contain informationfor those applicants who were "accepted" or previously awarded credit by theorganization; data will not be available for those applicants who were "rejected." Theuse of reject inference to adjust credit scorecard models for the missing data representedby rejected loan applications is common practice, and several approaches are used intoday's financial industry.The use of mixture models presents an alternative approach for reject inference.In this approach, the underlying probability density function (PDF) for the totalpopulation of all applicants is modeled as a weighted average of the PDFs of the acceptedpopulation and the rejected population. The unknown parameters in this model will thenbe the mixing parameters of this weighted average. To estimate these parameters, we usethe Expectation-Maximization (EM) algorithm wherein the data associated with therejected applicants is treated as missing completely at random. Simulated data derivedfrom actual case studies are used to assess the effectiveness of the mixture modelapproach and other reject inference techniques.Next, an alternative reject inference model is developed using a logistic regressionapproach. The parameters of this alternative model are derived using the EM algorithm.A simulation study will be performed to determine if the alternative reject inferencemodel offers any advantages over the financial industry's current standard rejectinference approach.Finally, a case study using data from a real-world reject inference data set will beconducted to determine how well the newly developed reject inference technique usingthe EM algorithm performs when compared to the standard method of reject inference.