Late deliveries and supply chain disruptions impact operational and financial performance in global logistics networks. We propose a three-stage machine learning (ML) approach for late delivery risk classification and delay magnitude prediction applied to a large-scale supply chain data set with model explainability techniques. Ensemble classifiers trained in the first stage using feature engineering and calibrated for group-specific thresholds performed well. A shipping mode-specific regression model in the second stage outperformed the model trained globally, and for the first time the underlying operational rule governing the same day deliveries was discovered. In the last stage, findings from the SHAP and LIME application for both global and local explanations were combined. The global explanations from the classification and regression demonstrated the duration of the scheduled shipment as the most important predictor, while the selected local explanation methods were consistent in identifying the key driver. The proposed framework may act as a practical, transparent, and interpretable decision support tool for supply chain practitioners to manage and shift from reactive to proactive delivery management.

Keywords

Supply chain management, date delivery prediction, machine learning, explainable AI and ensemble learning.