Asymmetric Graph-Based Deep Reinforcement Learning for Portfolio Optimization

Haoyu Sun, Xin Liu, Yuxuan Bian, Peng Zhu, Dawei Cheng, Yuqi Liang

Abstract: In recent years, existing studies have sought to enhance the effectiveness of portfolio optimization by modeling asset relations. However, employing conventional graph neural network methodologies for effective aggregation and final representation learning of intricately complex financial information within real-world markets proves challenging. This necessitates the optimization of graph structures to enhance the accuracy of parsing and leveraging financial information. In this paper, we propose an asymmetric graph-based deep reinforcement learning for portfolio optimization. Specifically, leveraging the excellent evaluative capabilities of large language models, we decipher multi-dimensional asymmetric relationships between stocks in multi-dimensional data, constructing asymmetric stock relationship graphs based on news and sectors. We then design a multi-dimensional relationship attention mechanism to jointly represent asymmetric graph information and employ deep reinforcement learning for end-to-end portfolio optimization. Extensive experiments on real datasets from China and the United States have demonstrated the superiority of our method over existing state-of-the-art methods. In the industrial observation conducted at a leading financial technology company, we validated the applicability of our method in real-world market scenarios.