Beyond Accuracy: Leakage-Audited Machine Learning for Explainable Microcredit Risk Screening at Bank Mandiri

Abstract

Accuracy can misrank models when at-risk borrowers are rare and missed risk is costlier than unnecessary review. This study evaluates a leakage-audited, explainable machine-learning framework for microcredit risk screening at Bank Mandiri and addresses prior work that rarely links temporal feature availability, class imbalance, and asymmetric error costs. A quantitative cross-sectional experiment used 1,718 anonymized active-facility records from four Palembang microbusiness units. An audit reduced 29 candidate fields to 13 operational predictors. Logistic Regression, Decision Tree, and Random Forest were evaluated through a stratified 80:20 split and training-only fivefold cross-validation. A tuned Decision Tree achieved the highest test accuracy (96.22%; 331/344) but detected 15 of 24 at-risk facilities. Random Forest at a training-selected 0.46 threshold achieved 95.35% accuracy and detected 18 of 24. The accuracy difference was not significant (exact McNemar p = .607). Cost analysis showed that thresholded Random Forest becomes preferable when a false negative costs more than twice a false positive. Account liquidity explained 88.72% of Decision Tree impurity reduction. The study contributes a transparent governance rule connecting leakage control, interpretability, and asymmetric screening costs while preserving human oversight.