Explainable Credit Risk Assessment: Comparing Logistic Regression, XGBoost, and LightGBM with SHAP and LIME Analysis

This study investigates whether gradient boosting models (XGBoost and LightGBM), combined with SHAP and LIME explainability, offer meaningful advantages over Logistic Regression in credit scoring. Using the German Credit dataset (n = 1,000), three models were compared through stratified five-fold cross-validation, held-out test evaluation, and pairwise McNemar’s tests. None of the three models achieved a statistically significant improvement over any other (all p > 0.05). Logistic Regression achieved the highest recall on the test set (0.800), followed by XGBoost (0.717) and LightGBM (0.567). SHAP analysis revealed consistent nonlinear risk patterns across both ensemble models, including duration threshold effects and compensatory attribute dynamics inaccessible through linear coefficients. A comparative analysis of SHAP and LIME showed broad agreement on top predictors but meaningful differences in stability and explanation structure. The findings suggest that on small, structured datasets, Logistic Regression remains the most practical production choice, while gradient boosting with SHAP serves a complementary analytical role for credit policy design.

Explainable AI, Credit Risk, Logistic Regression, XGBoost, LightGBM, SHAP, LIME