Early identification of unrecoverable loans is a critical requirement for financial institutions to maintain portfolio quality, comply with regulatory provisioning standards, and minimize credit losses. In Nepal, microfinance institutions and banks are mandated to report loan performance using the Nepal Rastra Bank (NRB) N002 monitoring framework, which contains borrower demographics, loan characteristics, delinquency behavior, and provisioning information. Despite the availability of structured regulatory data, most institutions continue to rely on rule-based aging mechanisms that fail to capture complex nonlinear risk patterns. This study proposes a machine learning-based framework for predicting unrecoverable loans using NRB N002-compliant datasets. A supervised classification problem is formulated, where loans are labeled as unrecoverable based on regulatory delinquency thresholds (Days Past Due >180 or Provision ≥50%). Three models—Logistic Regression, Random Forest, and Extreme Gradient Boosting (XGBoost)—are implemented and evaluated using recall, precision, F1-score, and ROC-AUC metrics, with special emphasis on recall to minimize false negatives in high-risk loan identification. Experimental results demonstrate that XGBoost achieves superior performance with near-perfect recall for unrecoverable loans and an ROC-AUC exceeding 0.97, significantly outperforming traditional statistical approaches. Explainability is ensured using SHAP-based feature attribution. highlighting delinquency duration, overdue principal, outstanding exposure, and provisioning ratios as dominant predictors. The findings confirm that machine learning models can substantially enhance early warning credit risk systems within Nepalese financial institutions while maintaining regulatory transparency and operational interpretability.
Bajgai et al. (Thu,) studied this question.