RouteRisk: Dynamic Delivery Delay Prediction Using Shipment Trajectories and Hub Congestion Signals

Abstract

Delivery delay prediction is commonly performed when a shipment is created, although its risk can change substantially as the shipment moves through the logistics network. This paper proposes RouteRisk, a dynamic prediction model that continuously updates delay probabilities using shipment trajectories, hub processing times, route history, and network-wide congestion signals. Rather than treating each shipment independently, the model incorporates recent congestion statistics from upstream and downstream logistics hubs to capture delay propagation across the transportation network. Experiments are conducted on 3.72 million shipments passing through 84 logistics hubs over a six-month period, of which 412,680 arrive later than their promised delivery windows. RouteRisk achieves an AUROC of 0.913 and AUPRC of 0.684, compared with 0.861 and 0.571 for an XGBoost shipment-level baseline. At a fixed precision of 80%, the proposed model identifies 67.3% of delayed shipments, representing a 14.8-percentage-point improvement in recall. Predictions made 24 hours before the scheduled delivery time retain an AUROC of 0.887. Analysis of delayed shipments further shows that incorporating hub congestion improves detection particularly for cascading delays involving two or more intermediate facilities. The results suggest that network-level operational signals can substantially improve real-time delivery risk estimation.