Graph Neural Network–Based Cross-Market Risk Contagion Analysis between U.S. Equity Sectors and Treasury Yields during the 2023 Regional Banking Stress

Abstract

The 2023 regional banking stress that began with the collapse of Silicon Valley Bank on 10 March 2023 raised fresh concerns about how shocks travel across U.S. equity and Treasury markets within days. Standard spillover indices treat asset markets as symmetric correlation panels and miss the directed, time-varying structure of cross-market transmission. We study this episode using daily data on eleven SPDR Select Sector ETFs and four constant-maturity Treasury yields drawn from Yahoo Finance and the Federal Reserve Economic Database, covering 502 trading days from 3 January 2022 to 29 December 2023. We construct rolling-window directed graphs whose edges are pairwise generalized variance-decomposition spillover shares, and train a graph attention layer to reconstruct the next-window pairwise spillover matrix, so that the learned attention coefficients align with the contagion-analysis target rather than with a generic forecasting objective. The total connectedness index rises from 61.4 percent before the stress to 78.2 percent in the acute window and peaks at 81.4 percent on 14 March 2023. The Financial Select Sector SPDR Fund and the two-year Treasury yield emerge as the dominant amplifier nodes, and the directed pair XLF → DGS2 carries 12.4 percent of attention-weighted system spillover during the acute window.