Article: Predicting SME loan default with machine learning: evidence from bank-branch-level data Journal: International Journal of Banking, Accounting and Finance (IJBAAF) 2026 Vol.16 No.1 pp.135 - 165 Abstract: In the context of post-crisis economy, where small and medium-sized enterprises (SMEs) remain pivotal to recovery, commercial banks still face challenges in the accurate assessment of credit risk. Machine learning models offer strong potential, yet adoption is hindered by class imbalance, limited interpretability, and integration challenges. We try to address these issues by showing that ensemble methods, combined with resampling and cost-sensitive techniques, improve loan default detection in imbalanced settings. We use a proprietary dataset of SME loans from a branch of a large EU systemic bank, and we evaluate random forests, extra trees, AdaBoost, CatBoost, XGBoost and a multi-layer perceptron (MLP) with stratified k-fold cross-validation approach. Our findings show that random forest with random oversampling, CatBoost and XGBoost with ADASYN, exhibit superior predictive performance. Complementary SHAP analysis identifies the most influential borrower, loan, and sector characteristics, thus making low-level credit-granting decisions more accurate, transparent, and context-sensitive. Inderscience Publishers - linking academia, business and industry through research

Title: Predicting SME loan default with machine learning: evidence from bank-branch-level data

Authors: Augustinos I. Dimitras; Konstantinos Gkillas; George Peppas; Maria Tantoula

Addresses: School of Social Sciences, Hellenic Open University, Patras, Greece ' Department of Management Science and Technology, University of Patras, Greece; Prague University of Economics and Business, Czech Republic ' School of Social Sciences, Hellenic Open University, Patras, Greece ' School of Social Sciences, Hellenic Open University, Patras, Greece

Abstract: In the context of post-crisis economy, where small and medium-sized enterprises (SMEs) remain pivotal to recovery, commercial banks still face challenges in the accurate assessment of credit risk. Machine learning models offer strong potential, yet adoption is hindered by class imbalance, limited interpretability, and integration challenges. We try to address these issues by showing that ensemble methods, combined with resampling and cost-sensitive techniques, improve loan default detection in imbalanced settings. We use a proprietary dataset of SME loans from a branch of a large EU systemic bank, and we evaluate random forests, extra trees, AdaBoost, CatBoost, XGBoost and a multi-layer perceptron (MLP) with stratified k-fold cross-validation approach. Our findings show that random forest with random oversampling, CatBoost and XGBoost with ADASYN, exhibit superior predictive performance. Complementary SHAP analysis identifies the most influential borrower, loan, and sector characteristics, thus making low-level credit-granting decisions more accurate, transparent, and context-sensitive.

Keywords: loan default; machine learning; risk management; SME financing; SHAP analysis.

DOI: 10.1504/IJBAAF.2026.155774

International Journal of Banking, Accounting and Finance, 2026 Vol.16 No.1, pp.135 - 165

Received: 11 Oct 2025
Accepted: 19 Mar 2026
Published online: 13 Aug 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: Predicting SME loan default with machine learning: evidence from bank-branch-level data

Keep up-to-date