From Weak Signals to Supply Chain Reliability Risk: A Feedback-Driven Risk Accumulation Framework for Late-Delivery Prediction, Early Warning, and Risk Prioritization

Late-delivery risk can develop through accumulated operational signals rather than a single disruption. This study proposes a feedback-driven risk accumulation framework that converts matured historical outcomes and observable shipment information into leakage-controlled rolling memory, interpretable escalation features, and decision-oriented risk scores. The primary evaluation uses 180,519 DataCo order-line records partitioned chronologically at the Order Id level to prevent entity leakage. Relative to a gradient-boosting baseline, the feedback model improves ROC-AUC from 0.7674 to 0.8013 and F1 from 0.7194 to 0.7527. Ablation results show that rolling matured-outcome memory accounts for most of the predictive improvement, while persistence, intensity, cross-dimensional breadth, and amplification primarily add diagnostic structure. Under a matched FN5 threshold policy, the feedback model reduces illustrative scenario loss from 5,560 to 5,401 under the assumed 5:1 false-negative-to-false-positive penalty. A secondary SMT2020 fab-lot simulation shows that the same outcome-availability-safe logic can be instantiated for semiconductor due-date-miss prediction. Overall, the results support a retrospective early-warning and triage-readiness evaluation that links temporal risk memory with leakage-controlled prediction, matched review-capacity analysis, and scenario-based threshold policies.