Building a Credit Risk Model using Transfer Learning and Domain Adaptation
Transfer learning has been successfully applied to the credit risk domain to predict the probability of default for “new to credit” individuals and small businesses. However when the source and target domains differ, we propose a domain adaptation approach to adjust the source domain features. We find that adaptation improves model accuracy in addition to the improvement by transfer learning. We propose and test a combined strategy of feature selection and an adaptation algorithm to convert values of source domain features to mimic target domain features. We find that transfer learning improves model accuracy by increasing the contribution of less predictive features. Although the percentage improvements are small, such improvements in real world lending would be of great economic importance. Our contribution also includes a strategy to choose features for adaptation and an algorithm to adapt values of these features.