AI-Enhanced Cross-Asset Liquidity Contagion Pathway Identification and Dynamic Hedging Strategy Optimization: Evidence from U.S. Equity, Bond, and Derivatives Markets

Abstract

Cross-asset liquidity risk constitutes a persistent threat to financial stability, as demonstrated by the 2007–2009 Global Financial Crisis, the March 2020 COVID-19 market disruption, and the March 2023 banking stress episode. This study investigates the identification of time-varying liquidity contagion pathways across U.S. equity, fixed-income, and derivatives markets, and the optimization of dynamic hedging strategies informed by such pathways. A temporal graph attention network (T-GAT) is constructed from publicly available data—including FRED macroeconomic indicators, CBOE VIX, and Amihud illiquidity measures derived from Yahoo Finance—to capture directional liquidity spillovers among six representative asset proxies. Transfer entropy serves as the edge-weighting mechanism, and the graph structure is updated on rolling 60-day windows. A Proximal Policy Optimization (PPO)-based deep reinforcement learning agent leverages the T-GAT-derived contagion state to allocate hedging positions across assets. Out-of-sample evaluation over the 2022–2024 period indicates that the T-GAT achieves an F1-score of 0.731 in predicting weekly liquidity stress events, representing a moderate improvement over LSTM and static GCN baselines. The contagion-informed hedging agent yields a Sharpe ratio of 0.847, compared to 0.623 for conventional risk-parity allocation, with reductions in maximum drawdown during the March 2023 stress period.