This paper investigates the application of Reinforcement Learning (RL) to enhance pair trading in cryptocurrency markets, which are characterized by extreme volatility. The authors construct a custom RL environment that allows agents to make dynamic decisions on both trade timing and investment quantity, moving beyond traditional static rule-based approaches. The methodology involves pair formation using correlation and cointegration tests, spread calculation with moving windows, and hyperparameter optimization via grid search. RL agents are trained using actor-critic algorithms (DQN, SAC, A2C, PPO) with a novel reward shaping mechanism that encourages profitable actions while penalizing excessive transactions.

Experiments on 1-minute BTC-GBP and BTC-EUR data (263,520 samples) show that RL-based pair trading significantly outperforms the traditional benchmark. The traditional method yields an annualized profit of 8.33%, while RL agents achieve profits from 9.94% to 31.53%, with the best performance from algorithms allowing continuous action spaces. The dynamic scaling approach, where agents adjust investment amounts based on opportunity quality, is a key innovation that captures more profit in volatile markets. The results demonstrate that RL can effectively adapt to changing market conditions and improve profitability in high-frequency pair trading.

The study's significance lies in its practical contribution to algorithmic trading, showing that RL can automate and enhance decision-making in complex financial environments. The proposed framework offers flexibility in investment sizing and timing, which is crucial for volatile assets like cryptocurrencies. Future work could explore multi-asset pairs, transaction cost sensitivity, and real-time deployment.