Ablation Study of Class Imbalance Techniques for Credit Default Prediction with SHAP-based Explainability Analysis

<title>Abstract</title> <p>SMOTE has been the default solution for class imbalance in credit scoring since 2002. An alternative exists—XGBoost’s scale_pos_weight parameter adjusts the loss function to upweight minority-class instances, requiring no synthetic data. Whether this simpler approach matches SMOTE’s performance on credit risk data has not been tested. Both approaches were evaluated, alongside a no-rebalancing baseline, across Logistic Regression, Random Forest, and XGBoost on the Home Credit Default Risk dataset (307,511 applications). XGBoost with scale_pos_weight achieved 64.8% recall, surpassing SMOTE’s 43.3% by 21.5 percentage points (\(\:p&lt;{10}^{-6}\), McNemar’s exact test). This advantage did not transfer to Random Forest, where SMOTE instead outperformed. Five-fold stratified cross-validation confirmed stability (64.1% \(\:\pm\:\) 0.7% recall; ROC-AUC 0.755 \(\:\pm\:\) 0.003). SHAP analysis revealed that aggregated external bureau scores dominated the model’s risk assessment. The optimal imbalance strategy, these results suggest, depends on the classifier—not on the resampling method alone.</p>