Comparative Evaluation of Graph Neural Networks for Cross-Market Risk Contagion Path Identification in Multi-Layer Financial Networks

Abstract

Cross-market risk contagion---the propagation of financial distress across asset classes such as equities, bonds, derivatives, and foreign exchange---poses a persistent challenge for systemic risk monitoring. While graph neural networks (GNNs) have demonstrated strong performance in single-market financial risk tasks, their effectiveness for identifying contagion paths across interconnected multi-layer financial networks remains insufficiently explored. This paper presents a comparative empirical evaluation of four GNN architectures---GCN, GAT, GraphSAGE, and EvolveGCN---alongside a spatial-temporal graph convolutional baseline (ASTGCN) and three traditional econometric methods, applied to multi-layer financial networks constructed from U.S. equity, bond, derivatives, and foreign exchange market data spanning January 2018 to December 2023. The evaluation covers contagion path identification accuracy against realized stress episodes, node vulnerability ranking precision, the incremental benefit of multi-layer network construction over single-layer alternatives, sensitivity to edge-density thresholds, and a feature-matched comparison that isolates the contribution of graph-based nonlinear aggregation from that of richer input features. Results indicate that temporal-aware GNN architectures achieve moderate yet consistent gains in contagion detection, with ASTGCN attaining the highest AUC-ROC of 0.841 (95% CI: [0.823, 0.858]) on multi-layer networks. Multi-layer construction yields a 7.9 percentage-point improvement in AUC-ROC over the best single-layer configuration (95% CI: [5.4, 10.3] pp, p < 0.001). Macro-financial factors---particularly the VIX and credit spreads---emerge as the most informative features for vulnerability scoring. A case study on the March 2020 COVID-19 market crash validates the practical relevance of the identified contagion paths.