Small and medium-sized enterprises (SMEs) play a vital role in driving economic growth, yet many still encounter barriers when seeking credit. Traditional models such as logistic regression have long been the backbone of credit scoring because they are simple and interpretable. However, their ability to capture the complex and sometimes nonlinear nature of borrower behavior is limited, which often leads to misclassification of risk. This study applies a quantitative, explanatory approach to compare the performance of four models—Logistic Regression, Random Forest, XGBoost, and a Multi-Layer Perceptron (MLP) neural network—in predicting SME credit risk. Model accuracy, precision, recall, F1-score, and the ROC-AUC metric were used for evaluation, while SHAP and LIME were integrated to improve interpretability and transparency. The findings reveal clear differences in performance. Logistic regression produced an accuracy of 79% with a ROC-AUC of 0.58 but identified only 22% of actual defaulters, highlighting its weakness in imbalanced datasets. Random Forest increased recall to 68%, demonstrating better sensitivity to defaults, but its overall accuracy dropped to 65%, reflecting trade-offs between identifying defaulters and misclassifying non-defaulters. XGBoost achieved 86% accuracy and a ROC-AUC of 0.74, but its recall for defaulters was extremely low at just 2.4%, showing a bias toward majority (non-default) cases. The strongest results came from the MLP neural network, which reached 95% accuracy, balanced precision and recall of 0.95, and a ROC-AUC of 0.98, confirming its ability to capture hidden patterns in borrower data.This study adds to credit risk research by illustrating how advanced machine learning models, when combined with interpretability tools, can provide both accuracy and accountability. For SMEs, the practical implication is fairer and more reliable access to credit, while for lenders and regulators, these models offer a pathway to strengthen credit assessment without sacrificing transparency.
Oyeyemi et al. (Wed,) studied this question.