The digital transformation of the banking industry requires credit scoring systems that are both accurate and adaptable to complex, diverse data. This study aims to develop and evaluate a credit scoring model using ensemble supervised learning to predict credit risk for a consumer loan service (Product X) at Bank XYZ. Ensemble algorithms such as Random Forest, AdaBoost, LightGBM, CatBoost, and XGBoost were compared to a single classification method, Decision Tree. Model performance was assessed using precision, recall, F1-score, and ROC-AUC. The results show that XGBoost outperformed other models, achieving the highest ROC-AUC score of 0.803, indicating strong generalization and low risk of overfitting. SHAP analysis revealed key features influencing the model, including loan tenor, loan amount (plafond), income, and Days Past Due (DPD) history. Compared to the baseline Decision Tree model (ROC-AUC 0.573), XGBoost significantly improved classification accuracy. It also showed the potential to reduce the Non-Performing Loan (NPL) rate from 4% to below 3% and increase the approval rate from 65% to over 70%, aligning with Product X’s KPIs. These findings confirm that ensemble learning models especially XGBoost offer strategic value in enhancing credit portfolio quality and decision-making in digital banking.