ML vs. Value Sorting for Shipment Prioritization.

In supply chain management, models predicting delay risk are typically judged by their predictive accuracy. However, this research argues that the practical utility for managers, who can only review a limited number of shipments, lies in identifying which shipments to check first. The study investigates whether machine learning (ML) models can surpass a straightforward baseline: prioritizing shipments based solely on their highest monetary value. Using three real-world supply chain datasets—SCMS procurement, DataCo logistics, and Olist e-commerce—the researchers employed leakage-controlled rolling-origin evaluation and bootstrap confidence intervals. They found that ranking by predicted delay severity multiplied by known value (M1) generally outperformed severity-only ranking. However, M1 did not consistently beat the simple value-sorting approach across all datasets. The key differentiator appeared to be the "severity learnability" of the data. DataCo, with a higher R^2 and low calibration bias, showed ML outperforming value sorting. In contrast, SCMS and Olist, with poor R^2 and negative calibration bias, saw ML underperform. The study concludes that value sorting should remain a permanent benchmark, and ML models should only be deployed for prioritization after rigorous auditing of severity learnability and calibration under robust evaluation protocols.