Measuring the Default Risk of Small Business Loans: Improved Credit Risk Prediction Using Deep Learning

Abstract

This paper proposes a multilayer artificial neural network (ANN) method to predict the probability of default (PD) within a survival analysis framework. The ANN method captures hidden interconnections among covariates that influence PD, potentially leading to improved predictive performance compared to both logit and skewed logit models. To assess the impact of covariates on PD, we introduce a generalized covariate method that accounts for compositional effects among covariates and employ stochastic dominance analysis to rank the importance of covariate effects across both the ANN and logit model approaches. Applying the ANN method to a large dataset of small business loans reveals prediction gains over the logit models. These improvements are evident for short-term prediction horizons and in reducing type I misclassification errors in the identification of loan defaults, an aspect crucial for effective credit risk management. Regarding the generalized covariate effects, our results suggest that behavior-related covariates exert the strongest influence on PD. Moreover, we demonstrate that the ANN structure stochastically dominates the logit models for the majority of the covariates examined.