Credit Default Prediction Using Logistic Regression and XGBoost: A Multi-Scenario Evaluation with SHAP-Based Interpretability Analysis

Abstract

This study compares Logistic Regression (LR) and Extreme Gradient Boosting (XGB) for credit default prediction using 554,349 fintech loan records with a 7.95 percent default rate. Twelve scenarios covered 70-to-30 and 80-to-20 train-test splits, under both imbalance conditions, plus out-of-time validation from October 2023 to February 2024. Bootstrap analysis with 1,000 resamples showed consistent XGB superiority in ROC-AUC and PR-AUC, with non-overlapping 95 percent confidence intervals. In random splits, XGB achieved ROC-AUC of 0.788 to 0.790 versus 0.725 to 0.726 for LR and about 40 percent higher precision at similar recall. Under temporal validation, XGB reached 0.751 to 0.760 versus 0.711 to 0.714 for LR. LR preserved higher recall but produced far more false positives, while XGB retained higher precision and F1. SHAP identified internal score and loan term days as XGB’s dominant predictors, while LR rankings shifted more over time.