Interpretable and leakage-aware prediction of shipment delays in global supply chains: a Random Forest–SHAP workflow with multi-model benchmarking on the DataCo dataset

Purpose This study develops a leakage-aware evaluation workflow for predicting shipment dispatch delays at the time of order, auditing when each predictor becomes available and quantifying how much apparent accuracy on the DataCo dataset derives from post-outcome information. Design/methodology/approach On 172,765 completed DataCo shipments, a point-in-time audit identifies 21 order-time features and excludes post-outcome fields, including Late_delivery_risk. Five identically tuned models are compared over 15 order-grouped resamples with Nadeau–Bengio corrected t-tests (Holm-adjusted), cluster-bootstrap intervals, a chronological split, and exact SHapley Additive exPlanations (SHAP). Findings The leakage-free Random Forest attains R2 = 0.2719 (95% confidence interval 0.2569–0.2864). Restoring post-outcome fields raises R2 to 0.7589: a leakage inflation of ΔR2 = +0.4870 (same model, with versus without post-outcome inputs). A distinct diagnostic, the leaky model’s margin over a flag-threshold rule (ΔR2 = +0.0951), is meaningful only conditional on the leak. All five honest models lie within 0.0084 R2; SHAP indicates that order-time delay risk is structural, driven by mode, lane, and category base rates. Research limitations/implications Findings derive from one public dataset whose target is dispatch (not delivery) delay; generalisability claims are correspondingly modest. Practical implications Delay largely reflects infeasible service-level promises; the managerial levers are promise recalibration and base-rate triage, with measured detection rates (precision 0.864, recall 0.553) separated from assumed costs. Originality/value A feature-level point-in-time audit, a quantification of leakage inflation on a widely used benchmark, and a replicable leakage-audit protocol including a target-permutation canary.