Explainable Delivery Delay Prediction in E-Commerce Logistics

Delivery-delay prediction is most useful when operational users can understand both the risk signal and model failure conditions. Using the public DataCo Smart Supply Chain dataset, we reconstructed 180,519 line-item records into 65,752 unique orders, excluded post-shipment variables, and evaluated a fixed-configuration Extreme Gradient Boosting (XGBoost) classifier on 30 encoded pre-shipment features. On an untouched 20% test set, XGBoost achieved area under the receiver operating characteristic curve (AUC-ROC) 0.765, precision-recall AUC 0.833, Brier score 0.189, and late-class F1 0.690; Random Forest was closely competitive (AUC-ROC 0.762, F1 0.692). Grouped SHapley Additive exPlanations (SHAP) attributed 72.7% of explanation magnitude to shipping mode, payment type, and order hour. Standard Class exposed a major subgroup failure: recall was 0.045 and within-subgroup AUC-ROC was 0.515. Among 3,121 high-risk late orders, 96.7% crossed below 0.50 under at least one valid shipping-mode substitution, demonstrating predictive sensitivity rather than causal recourse. Across nine alternative false-negative/false-positive cost scenarios, validation-selected thresholds outperformed a flag-all-late baseline, although the advantage was sometimes small. The findings support leakage-aware diagnostic auditing rather than autonomous operational deployment.