Tracing working capital pressure in concentrated supply chains with temporal graph transformers

Concentrated trading relationships can improve coordination, yet they can also amplify operational stress when inventories, delivery commitments, and short-term obligations become misaligned. Conventional concentration ratios summarize dependence at the firm level but reveal little about how this pressure develops over time or travels through multi-tier supplier–buyer relations. This study presents SC-GraphTransformer-Press, a temporal graph framework that combines upstream and downstream concentration, product-specific transaction structure, recurrent firm states, and attention-based propagation to forecast working capital pressure. Pressure is operationalized as an inventory-backed stress proxy derived from unmet inventory-related obligations and available inventory, with firm-scale and network controls used to distinguish firms that have similar relative pressure but different operating capacities. The forecasting experiments use three SupplySim-derived benchmarks in which the required inventory and backlog states are observable. Under chronological multi-horizon evaluation, the proposed model achieves the strongest predictive performance among econometric, machine-learning, static graph, and temporal graph baselines, with its advantage increasing at longer horizons and under supply disruption. Network diagnostics further associate concentrated dependencies with persistent downstream pressure, while model-based scenarios indicate that supplier exits and supply shocks produce larger predicted responses than customer concentration and that product substitution can moderate these responses. Chronological audits and robustness tests across concentration assignments, shock locations, and simulator settings support the stability of these findings. The framework therefore offers a practical route from static concentration measurement to forward-looking diagnosis of working capital stress in dynamic supply-chain networks.