Deep Generative Models in Credit Scoring

Abstract

Banks need to develop effective credit scoring models to better understand the relationship between customer information and the customer's ability to repay the loan. The output of such a model is called the default probability and is used to rank loan applications in terms of their creditworthiness. The focus of this thesis is to develop novel credit scoring methodologies that solve well-known problems in the field and that bridge the gap between simple neural networks and advanced methodologies in deep learning applied to credit scoring.

We propose a new methodology to learn useful data representations of bank customers introducing a supervision stage, where we group the input data using the Weight of Evidence transformation, into the Variational Autoencoder framework. Our proposed method learns data representations that are able to capture the customers' creditworthiness in a well-defined clustering structure and that are well suited for marketing campaigns and credit risk assessment.

Further, we develop two novel Deep Generative Models that are able to infer the unknown customers' creditworthiness of rejected loan applications using probabilistic theory, which is a clear advantage over traditional approaches. Adding rejected applications improves the classification accuracy of our proposed models, and potentially solves the selection bias problem. We parametrize a Gaussian mixture model with neural networks to further improve the latent representation of customers information.

Finally, we address credit scoring as a multi-modal learning problem. That is, banks have multiple measurement-modalities that provide complementary information about customers. Hence, we develop a novel Deep Generative Model that learns shared data representations and maximizes mutual information between future credit data and its shared representation. Our proposed model is able to generate future credit data, based on application data, which can be used to support bank activities other than credit scoring.