Interpretability of Machine Learning versus Statistical Credit Risk Models

Abstract

Model interpretability is important in the banking industry for three reasons: certain US regulations require creditors to provide consumers with the reasons for taking adverse action (reason codes) on their credit applications; model users want to understand the reasoning behind model predictions; and identification of bias and reinforcement of stakeholders’ trust in the model. In this article, the authors compare the interpretability of an XGBoost versus a logistic model in predicting the probability of default for a credit card customer. They conclude that (1) the reason codes of an XGBoost model and a comparable logistic model are similar, (2) reason codes generated by XGBoost are more trustworthy from the customer’s perspective, and (3) nonlinearity of XGBoost is unlikely to have a significant impact on reason code(s).