SHAP-McNemar stepwise feature selection for machine learning in credit risk modeling

Information from the abstract

Non-Performing Loans pose a significant risk to financial stability, which can affect economic stability; conversely, macroeconomic conditions also influence credit risk. A credit risk model using machine learning algorithms and explainable AI is widely used, while traditional wrapper feature selection methods lack statistical inference for the underlying population. This study proposes a credit risk model that predicts the probability of default using a machine learning algorithm incorporating macroeconomic determinants, and introduces a novel SHAP-McNemar stepwise feature selection framework that integrates SHAP values for candidate ranking and uses the McNemar mid-p test for statistically grounded removal and addition stepwise decisions. The framework is applied to penalized logistic regression and XGBoost models using U.S. small-business loan data, merged with lagged macroeconomic features. The results showed that the proposed method reduces the feature dimension from 38 to 7 for penalized logistic regression and to 19 for XGBoost; moreover, both SHAP-McNemar reduced models outperform established feature selection methods, while performing comparably to the full model. Interest rate, annual income, and inflation rate consistently influence the loan default prediction across both models. The SHAP-McNemar reduced penalized logistic regression model achieves performance comparable to or superior to XGBoost with substantially fewer features, which is particularly important for credit risk applications, where interpretability is essential for regulatory compliance. The proposed credit risk model and feature selection framework offer a transparent, interpretable, and statistically based approach, supporting effective credit risk management and regulatory compliance.

Why this record is monitored

This record has an Impact Signal of 85/100 based on recency, source, collaboration, and bibliographic signals. It prioritizes monitoring and is not a judgment of research quality.

Related topics: Financial Distress and Bankruptcy Prediction · Credit Risk and Financial Regulations · Imbalanced Data Classification Techniques