Graph neural network-based prediction of cascading supply chain disruptions using real-time logistics and supplier risk data

Abstract

Cascading supply chain disruptions arise when a localized failure propagates through supplier, production, transportation, and distribution dependencies, so that the eventual operational impact is determined by both the condition of individual firms and the structure of the network connecting them. Conventional risk scores and stand-alone time-series models are limited in this setting because they usually treat suppliers as independent observations and therefore fail to represent higher-order exposure. This paper develops a graph neural network-based framework for predicting near-term node-level disruptions by integrating relatively slow-moving supplier risk indicators with high-frequency logistics signals. The supply chain is represented as a directed, weighted, time-indexed graph in which nodes denote firms or facilities, edges encode material or logistics dependencies, node attributes include financial, geopolitical, concentration, quality, lead-time, and resilience indicators, and dynamic observations include delay, congestion, capacity stress, demand volatility, inventory pressure, estimated-time-of-arrival variance, and exogenous shocks. An edge-weighted graph encoder learns structural exposure, while a gated recurrent unit summarizes the temporal evolution of graph states. Because no public benchmark currently combines complete multi-tier supplier relationships with synchronized real-time logistics and verified disruption labels, the empirical section uses a reproducible semi-synthetic network designed to emulate these data-generating conditions without presenting simulated observations as real company data. On a held-out temporal test period, graph-aware models achieved stronger discrimination than non-graph baselines. The temporal GCN-GRU obtained an AUROC of 0.844, AUPRC of 0.546, recall of 0.710, and F1 score of 0.592, while a snapshot GCN achieved the highest AUROC of 0.845 and AUPRC of 0.552. Ablation analysis showed that removing network propagation reduced AUROC to 0.810 and removing real-time logistics signals reduced AUROC to 0.825. The results support a central conclusion: supply chain disruption prediction benefits materially from representing inter-firm dependency as a graph, but temporal complexity should be added only when the data resolution and propagation lags justify it. The paper concludes with deployment guidance covering graph construction, calibration, explanation, data governance, concept drift, and integration with digital supply chain twins. Graph Neural Networks, Supply Chain Disruption, Cascading Failure, Ripple Effect, Supplier Risk, Real-Time Logistics, Temporal Graph Learning, Supply Chain Resilience, Risk Propagation, Digital Supply Chain Twin.