Loan default predictability with explainable machine learning

Abstract

This paper studies loan defaults with data disclosed by a lending institution. We comprehensively compare the prediction performance of nine commonly used machine learning models and find that the random forest model has an efficient and stable prediction ability. Then, we apply an explainable machine learning method, i.e., SHapley Additive exPlanations (SHAP), to analyze the important factors affecting loan defaults. Moreover, we conduct an empirical study and find that the significant influencing factors are clearly consistent with those suggested by SHAP: the older the lender and the longer their working experience, the lower the risk of loan default.