A Maximum Likelihood Approach for Reject Inference in Credit Scoring

Abstract

We model reject inference - inferring how a rejected credit applicant would have behaved had it been granted credit - using a maximum likelihood approach within the framework of missing data analysis. Contrary to other methods that impute missing data, this reject inference method embeds the missing data mechanism into model estimation directly. We test performance of three reject inference methods using real default data. Results show our method to be superior and to improve classification power for credit scoring in within-sample tests.

Keywords: reject inference, missing data not at random, maximum likelihood, credit scoring

JEL Classification: C14, C25, C29