Integrating Dynamic Graph Representation and Reinforcement Learning for Portfolio Optimization

Abstract

The complex dependency relationships and high-frequency, time-varying characteristics of financial market data pose many challenges to portfolio optimization. To address these problems, this paper proposes a robust portfolio optimization framework—dynamic graph reinforcement learning (DyGRL)—by integrating dynamic graph representation and reinforcement learning. The framework incorporates a dynamic graph construction module, which uses distance correlation coefficients to capture nonlinear dependencies among stocks and applies the triangulated maximally filtered graph method to remove redundant connections. Furthermore, DyGRL incorporates a mask reconstruction mechanism into dynamic graph learning to construct a graph representation learning module, which learns robust stock representations. Finally, a soft actor-critic-based decision network is introduced for portfolio weight optimization. Experimental results based on 5-min high-frequency data from the NDX and CSI300 markets from September 2016 to August 2025 show that, compared with the best-performing baseline method, DyGRL improves annualized returns by 3.21% and 1.43%, respectively. Moreover, the moving block bootstrap test confirms that the improvement in Sharpe ratio is statistically significant. The comprehensive results indicate that DyGRL can effectively improve portfolio performance and demonstrate the effectiveness of the proposed scheme.