<![CDATA[The banking sector faces a major challenge in the form of credit risk due to customers' inability to repay loans, which can threaten financial stability. Traditional methods are less effective at handling complex data and class imbalances, so a machine learning approach is needed to improve classification accuracy. This study compares the performance of the Naïve Bayes, AdaBoost, Random Forest, and XGBoost algorithms on the German Credit dataset (1000 entries, 70% low risk and 30% high risk) with SMOTE techniques to overcome class imbalances as well as SHAP as Explainable AI for model interpretability. Data processing is carried out using Python (Pandas, Scikit-learn, XGBoost, SHAP) in Google Colab, including preprocessing (handling missing values, encoding, scaling), 10-fold cross-validation evaluation, and SHAP analysis. The results showed that Random Forest achieved the best performance with an average accuracy of 87% (Std: 0.036), precision 0.88, recall 0.87, F1-score 0.87, and ROC-AUC 0.93, followed by XGBoost (85%, Std: 0.041). Naïve Bayes and AdaBoost only reached 81%. SHAP analysis revealed Credit Amount and Duration as the most influential features on high risk prediction (positive correlation ~0.62). The ensemble model excels in accuracy and stability, while the integration of SMOTE and SHAP improves minority class recall as well as transparency for banking decision-making. This study outperformed previous studies (71.33% in ANN and 83% in Random Forest) thanks to the combination of these techniques, supporting more accurate, ethical, and regulated credit risk management in financial institutions.]]>
