Tuning White Box Model With Black Box Models : Transparency in Credit Risk Modeling

In the context of credit risk modeling, more sophisticated machine learning models, such as boosted regression trees, often show higher predictive accuracy than linear regression models. Nevertheless, linear regression remains popular in the credit risk industry, mainly because the lack of explainability and interpretability of machine learning models is incompatible with the requirements of financial regulators. In this study, we propose the use of linear regression with automated variable selection optimized using machine learning. Formally, the most influential variables and interactions in the machine learning model, identified using Shapley values, are used as predictors of linear regression. This approach allows us to capture the correct linear and nonlinear effects that can arise in credit data while preserving the intrinsic interpretability of linear regression. Based on Monte Carlo experiment and an empirical analysis using two real-life credit default datasets, we show that linear regression with the optimal variable set can predict credit risk significantly more accurately than regressions that use standard variable selection techniques and is competitively comparable with the best machine learning methods