Comparative Analysis of Advanced Machine Learning Models for SME Credit Risk Classification

Abstract

This study addresses the growing need for accurate and reliable credit risk classification in Small and Medium Enterprises (SMEs), which play a vital role in economic development but remain vulnerable to financial instability and non-performing loans. The purpose of this research is to comparatively evaluate the performance of advanced machine learning models in multi-class SME credit risk classification and to identify the most influential predictors affecting creditworthiness. A quantitative experimental approach was employed using 2,000 SME debtor records from 2020–2024, incorporating 17 financial and behavioral variables. The study implemented five supervised learning algorithms Logistic Regression, Random Forest, Support Vector Machine (RBF), XGBoost, and Artificial Neural Network combined with data preprocessing, feature selection, and 5-fold cross-validation. Model performance was assessed using imbalance-aware metrics, including F1-Macro and ROC-AUC, alongside statistical validation using the Friedman test. The results demonstrate that ensemble-based methods outperform traditional models, with XGBoost achieving the highest predictive performance (F1-Macro = 0.92; ROC-AUC = 0.95) and showing statistically significant superiority. Feature importance analysis reveals that Debt-to-Income Ratio, Credit Tenure, and Internal Credit Score are the most influential predictors, aligning with established financial risk theory. In conclusion, this study confirms the effectiveness of ensemble machine learning models in improving SME credit risk classification and highlights their potential integration into automated decision-support systems to enhance risk management, reduce non-performing loans, and support data-driven financial decision-making.