In the context of post-crisis economy, where small and medium-sized enterprises (SMEs) remain central to economic recovery, commercial banks face persistent challenges in accurately and efficiently assessing credit risk. Machine learning models offer strong predictive potential, yet their adoption in real-world lending is hindered by issues such as severe class imbalance, limited interpretability, and difficulties integrating them into existing workflows. This study addresses these challenges by providing empirical evidence that ensemble learning methods, when combined with resampling and cost-sensitive techniques, can improve loan default detection in highly imbalanced settings. Using a unique, proprietary dataset of SME loans collected from the branch of a large systemic Greek bank, we evaluate Random Forests, Extra Trees, AdaBoost, CatBoost, XGBoost and Multi-Layer Perceptron (MLP), within a stratified k-fold cross-validation framework. Our findings show that hybrid configurations, specifically Random Forest with ADASYN and XGBoost with SMOTEENN, deliver superior predictive performance. Complementary SHAP analysis identifies the most influential borrower, loan, and sector characteristics, thus enabling branch managers to make more accurate, transparent, and context
sensitive lending decisions.

Jel Classification: C45, C53, G21