Predicting Potential Non-Performing Loans Collectibility in MSMEs using Ensemble Stacking with SHAP-Boosting Algorithm

This research examines the challenges banks encounter when lending to micro, small, and medium enterprises (MSMEs) in Indonesia. A significant concern is the high rate of Non-Performing Loans (NPLs), which reached 5.65% in December 2022. To improve credit assessment efficiency and overcome the limitations of traditional scoring methods, an ensemble stacking algorithm utilizing XGBoost, CatBoost, and LightGBM as base-learners is applied to predict prospective debtors with potential NPL status. A SMOTE-ENN hybrid sampling technique is employed to tackle historical data imbalance. The analysis shows that XGBoost excels in class distinction with a ROC-AUC of 92.8%, while LightGBM effectively identifies the NPL class with an AUC-PR of 75.3%. Both models also achieved a Recall of 86%. The stacking model outperforms the individual models, showing the highest Precision at 92%, an F1-Score of 88%, and the lowest Log Loss of 0.055. Model interpretability using the SHAP algorithm reveals that 'Income Range' is the most influential feature on NPL risk, with low income increasing risk and high income decreasing it. Additionally, 'Interest Rate' significantly impacts the CatBoost and LightGBM models, while the 'Region' feature is more dominant in the XGBoost model. In contrast, features like 'Total Loan' and 'Gender' show less significance in influencing predictions. Future research suggested explore ensemble deep learning and alternative interpretability methods to enhance variable influence understanding.