Title: Enhancing Cryptocurrency Pair Trading via Deep Reinforcement Learning: A Dynamic Multi-Pair Approach

Abstract

This research investigates the potential of Deep Reinforcement Learning (DRL) as a specialized execution overlay to improve pair trading strategies within the highly volatile landscape of cryptocurrency markets. While traditional implementations of pair trading have shown success in conventional equity markets, they often prove inflexible and prone to significant divergence risks when deployed in high-variance environments. To overcome these limitations, this study introduces innovative methodological concepts. We designed a robust system featuring a hierarchical "Filter-then-Rank" approach for pair selection and a proprietary "Fixed Risk, Adaptive Mean" execution model. At the core of this system is a Proximal Policy Optimization (PPO) agent, augmented with a Long Short-Term Memory (LSTM) layer, which manages execution decisions while adhering to strict deterministic risk management constraints.

The strategy was evaluated using 1-hour interval data from the Binance USD-M Futures market. The results demonstrate that the optimized RL policy significantly outperformed the heuristic baseline in out-of-sample testing. A stationary circular block bootstrap analysis confirmed that the agent’s risk-adjusted outperformance is statistically significant at the 10 percent level. While the findings narrowly missed the stricter 5 percent significance threshold, this outcome underscores the extreme idiosyncratic variance inherent in digital assets. Ultimately, this work contributes to the quantitative finance domain by presenting a hybrid architecture that integrates statistical arbitrage with DRL execution policies. Additionally, it offers a novel framework for safe reinforcement learning through deterministic shielding, demonstrating that anchoring neural policies to statistically robust boundaries effectively mitigates severe divergence risks.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC