A Multi-Layer Graph Attention Network for Tail Risk Spillover Path Identification Across U.S. Equity, Credit, and FX Markets

Abstract

Cross-market financial stress propagates through pathways that are simultaneously heterogeneous, directional, and concentrated in the tails of joint return distributions, making contagion difficult to model with standard correlation-based or single-layer network approaches. This paper presents a Multi-Layer Heterogeneous Graph Attention Network (ML-HGAT) for tail risk spillover path identification across U.S. equity, high-yield credit, and foreign exchange markets. The framework couples three components: a quantile-regression CoVaR estimator and Dynamic Conditional Correlation GARCH that supply intra-layer tail-dependence edges, a rolling-window Granger causality network that supplies inter-layer directional edges, and a type-aware heterogeneous graph attention encoder that aggregates evidence across layers and emits both a contagion-event probability and a Top-k spillover path. Empirically, ML-HGAT is trained and evaluated on twenty years of daily observations drawn from FRED macroeconomic and liquidity series, CRSP equity returns, ICE-BofA U.S. High-Yield indices, and Federal Reserve liquidity indicators, spanning the 2008 global financial crisis, the 2011 European sovereign episode, the 2020 COVID-19 dislocation, the 2022 inflation shock, and the 2023 regional banking stress. Against six baselines, ML-HGAT attains an F1 of 0.8413, an AUC of 0.9237, a path hit-rate-at-ten of 0.7682, and a mean average precision of 0.6914, improving on the strongest single-layer GAT baseline by 3.46 absolute points in F1 and 5.58 points in path hit rate. Ablation analysis confirms that the inter-layer directional overlay and the tail-dependence intra-layer edges contribute additively to identification accuracy, and that the type-aware attention head is responsible for the bulk of the improvement during distinct stress regimes.