Reproduce “Select and Trade: Towards Unified Pair Trading with Hierarchical Reinforcement Learning” on OpenResearch

About this paper

Weiguang Han, Boyi Zhang, Qianqian Xie, Min Peng, Yanzhao Lai, Jimin Huang

Researchers developed TRIALS, a hierarchical reinforcement learning framework that unifies pair selection and trading for statistical arbitrage in financial markets. The framework consistently achieved higher Sharpe Ratios (e.g., 1.84 for S&P 500, 1.91 for CSI 300) and Annualized Returns while maintaining lower Maximum Drawdowns across two real-world stock market datasets.

Key insights

• Unifying pair selection and trading through hierarchical reinforcement learning enables critical information flow between these interdependent sub-tasks, leading to more effective joint optimization.
• Dynamically learning pair selection criteria based on actual trading performance and profitability yields superior results compared to static, historically-based statistical metrics.
• The most profitable pairs, as identified by the unified learning framework, do not necessarily correspond to those with the lowest historical price divergence, challenging assumptions made by traditional pair selection methods.

Summary by alphaXiv.