Abstract <p>Small and Medium-sized Enterprises (SMEs) are vital for economic growth but frequently face significant challenges in securing credit due to perceived high risk and the prevalence of Non-Performing Loans (NPLs) in highly imbalanced datasets. This study proposes a robust credit risk assessment framework utilizing a stacked ensemble machine learning approach to improve prediction accuracy for SMEs. Leveraging a real-world dataset from the Ministry of Industry, Thailand, comprising 14 financial and operational variables, the research systematically investigated the impact of four data balancing techniques: Synthetic Minority Over-sampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN), SMOTE and Edited Nearest Neighbors (SMOTEENN), and SMOTE and Tomek Links (SMOTETomeK). The performance of nine machine learning models, including five base classifiers and four meta-learners within the stacking framework, was comprehensively evaluated using Accuracy, Precision, Recall, F1-score, and Area Under the Receiver Operating Characteristic Curve (AUC). Results consistently demonstrate that stacked ensemble models significantly outperformed individual base classifiers, with the SMOTEENN balancing technique proving most effective in addressing class imbalance. Notably, META-Multi-layer Perceptron (META-MLP) and META-Extreme Gradient Boosting (META-XGB) models achieved outstanding performance (e.g., F1-scores of 0.953 and AUC values of 0.990), showcasing superior generalization capabilities for credit risk classification. Conversely, the application of Sequential Feature Selection yielded mixed results, not consistently translating into significant performance enhancements across all scenarios, suggesting the original features were sufficiently informative. These findings highlight the potential of integrating stacking ensemble models with effective data balancing, particularly SMOTEENN, as powerful decision support tools for more accurate and equitable SME credit risk assessments.</p>