Open-ended deep reinforcement learning for a bitcoin trading bot

Abstract

Cryptocurrencies have a high volatility but also provide a high potential of profit. This paper discusses the use of deep reinforcement learning in bitcoin trading aiming to optimize profit. Various parameter configurations and sampling modes were explored such as such as different batch sizes (64, 128, 256) and training frequencies (every 64, 128, 256 steps), focusing on hybrid sampling, which combines random and recent experiences from the replay memory. The results demonstrated that hybrid sampling consistently outperformed recent sampling and, under optimal parameters (batch size 256, training every 64 steps), showed significant promise across both the rising market period (01.01.2024 to 15.05.2024) and the declining market period (30.11.2017 to 30.11.2018) which were used as validation sets. However, performance declines were noted over longer periods, likely due to catastrophic forgetting. These findings suggest that while the hybrid sampling approach can enhance a trading bot’s performance, particularly in the short term, its effectiveness diminishes over time.