Bitcoin Transaction Strategy Construction Based on Deep Reinforcement Learning

This study proposes a framework for automatic high-frequency transactions based on a deep reinforcement learning model - proximal policy optimization. The framework creatively regards the transaction process as actions and returns as awards to align with the idea of reinforcement learning. It compares price prediction policies including long short-term memory (LSTM) and multi-layer perception (MLP) by applying them to the real-time bitcoin price. Then an automatically generating transaction strategy is constructed building on PPO with LSTM as the static policy. The approach is able to trade bitcoins in a simulating environment with synchronous data and achieves []% higher benefit than that of the average market. It is demonstrated that the approach can earn excess returns and be expanded to other financial products.