A machine learning approach for credit risk assessment of SMEs: Evidence from Morocco

Abstract

Assessing the credit risk of small and medium-sized enterprises (SMEs) has become increasingly complex as borrower profiles are diverse and often non-linear. Traditional rating methods, still widely used in practice, struggle to capture this variability, which can limit their reliability in modern financial contexts. The objective of this study is to evaluate whether machine-learning techniques can provide more accurate and operationally useful tools for SME credit-risk assessment. Using a dataset of 124 Moroccan SMEs, containing financial, behavioral, and transactional variables, we applied three supervised classification models: logistic regression, random forest, and XGBoost, to predict contract defaults. The models are assessed through standard performance metrics including accuracy, precision, recall, F1-score, and AUC. Results demonstrate that XGBoost provides the strongest detection of defaults, eliminating false negatives in our test set and making it especially suitable for loss-minimization contexts. By contrast, random forest achieves the highest discrimination between risky and non-risky profiles (AUC = 0.93), offering a balanced solution for operational scoring. Logistic regression, while less accurate, retains value for its interpretability and transparency. Overall, the findings highlight that ensemble methods, particularly XGBoost, can significantly improve the reliability of SME credit-risk evaluation. These results provide practical insights for financial institutions seeking to minimize default risk while also advancing the integration of artificial intelligence into credit-risk management frameworks.