<![CDATA[Delivery delay in supply chain logistics is a critical problem that negatively impacts customer satisfaction and generates significant financial penalty costs for logistics companies due to inefficient delivery performance and uncertainty in global distribution networks. This study aims to develop a predictive model for delivery delay risk using an optimized machine learning approach that not only improves classification performance but also provides measurable economic benefits through cost reduction analysis. The proposed method employs a Bayesian-Optimized Stacking Ensemble framework combining XGBoost, LightGBM, and Random Forest as base learners, with Lasso Regression as the meta-learner. Feature selection is performed using Recursive Feature Elimination with Cross-Validation (RFE-CV), while hyperparameter optimization is conducted using the Tree-structured Parzen Estimator (TPE) algorithm implemented through Optuna. The model is trained and evaluated using the DataCo Smart Supply Chain dataset consisting of 180,519 transaction records, with performance measured using accuracy, F1-score, ROC-AUC, and confusion matrix analysis. The experimental results show that the proposed model achieves an accuracy of 90.12%, F1-score of 0.9013, and ROC-AUC of 0.9602, indicating strong predictive capability. Furthermore, the business impact analysis demonstrates a penalty cost reduction of USD 523,062 or 52.85% compared to the baseline scenario without prediction, confirming that the proposed approach provides both high predictive performance and significant economic value in supply chain operations.]]>
