<![CDATA[This paper develops and evaluates an early-warning credit risk classifier using a three-class formulation that distinguishes performing loans (L), delinquent accounts (DP), and non-performing loans (NPL). Early-warning modeling is challenging because deterioration events are relatively infrequent, yielding class imbalance and asymmetric error costs where missed NPL cases are particularly consequential. The study aims to build and compare operationally feasible models that improve NPL identification while maintaining reliable performance across all classes. Experiments use 2,253 loan records with 20 early-warning predictors and compare Random Forest and XGBoost under four imbalance-handling strategies: a no-resampling baseline, class/sample weighting, SMOTENC, and Borderline-SMOTE. Performance is estimated via stratified 5-fold cross-validation, with out-of-fold predictions aggregated for consistent fold-robust evaluation and probability-based diagnostics. Evaluation emphasizes imbalance-aware criteria, including macro F1, balanced accuracy, NPL precision/recall/F1, and ranking metrics for NPL versus non-NPL (PR-AUC and ROC-AUC), complemented by agreement statistics (Cohen’s kappa and Matthews correlation coefficient). Results show that class-sensitive metrics reveal clearer differences among strategies than global measures, with XGBoost combined with SMOTENC achieving the strongest NPL-oriented performance (recall 0.7389; F1 0.8035; PR-AUC 0.8780; ROC-AUC 0.9434). The contribution is a controlled out-of-fold benchmark of imbalance strategies for mixed-type tabular credit staging, providing a reproducible evaluation protocol for financial informatics and imbalanced multiclass learning.]]>
