PREDICTIVE MACHINE LEARNING MODELS FOR MITIGATING NON-PERFORMING LOANS IN EMERGING BANKING SYSTEMS

Abstract

Rising volumes of non-performing loans (NPLs) remain one of the most persistent threats to financial stability in emerging banking systems, where credit registries are thin, macroeconomic volatility is elevated, and conventional scorecards rely on linear discriminant or logistic specifications that struggle to capture non-linear interactions among borrower, loan, and macroeconomic risk factors. This study benchmarks four supervised machine learning classifiers — logistic regression, random forest, gradient boosting, and extreme gradient boosting (XGBoost) — in predicting twelve-month-ahead loan delinquency using a loan-level panel drawn from commercial banks operating in an emerging Central Asian banking system over 2019–2025. Models are trained on borrower, loan, and collateral characteristics together with a macroeconomic overlay, validated through five-fold cross-validation with chronological hold-out testing, and compared using AUC-ROC, the Gini coefficient, the Kolmogorov-Smirnov (KS) statistic, F1-score, and Brier score. The results show that tree-based ensembles substantially outperform the logistic baseline, with XGBoost achieving the highest discriminatory power, and that augmenting loan-level features with macroeconomic overlay variables materially improves out-of-sample predictive accuracy. Feature-importance analysis identifies delinquency history, debt-service-to-income ratio, collateral coverage, and sectoral exposure as the dominant predictors. The findings offer commercial banks and supervisory authorities in emerging markets an empirically validated architecture for early-warning credit scoring capable of narrowing the gap between reported and risk-based measures of asset quality.