Enhancing supply chain efficiency with AI: predicting and preventing customer backorders

Abstract We propose a hybrid, interpretable AI module that enables autonomous backorder prevention in cyber-physical supply chains. The module integrates data-level rebalancing (SMOTE–Tomek), ensemble learners (e.g., LightGBM, XGBoost, Extra Trees), and an explanation layer combining intrinsic feature importance with permutation importance and partial-dependence/ALE analyses to infer rare backorder risks at scale while exposing the operational drivers behind each prediction. On a million-record industrial dataset, the approach outperforms competitive baselines on precision, recall, F1 and ROC-AUC, and remains transparent and deployment-ready. Explanations consistently highlight demand signals, inventory positions, and shipment delays as dominant determinants of risk, yielding policy-level levers for autonomous or human-in-the-loop replenishment agents (e.g., reallocation, expedited orders). Coupling predictive accuracy with actionable explanations, the module functions as a decision component within an autonomy loop (sense → predict → explain → act → feedback), advanc- ing learning for autonomy and improving resilience, cost-to-serve, and service reliability in networked operations.