Hybrid Deep Learning and Reinforcement Learning for Adaptive Cryptocurrency Trading Using Attention-LSTM and Soft Actor-Critic

This study proposes a hybrid artificial intelligence-based trading framework that integrates deep learning and reinforcement learning to develop an adaptive Bitcoin trading strategy. The proposed approach combines an Attention Long Short-Term Memory model to estimate the probability of future price movements with a Soft Actor-Critic agent to optimize trading decisions in a continuous action space. The probabilistic output of the prediction model is incorporated as a guidance signal within the reinforcement learning environment, enabling the agent to balance data driven prediction with policy-based decision making. Experiments are conducted on historical Bitcoin data using a comprehensive set of technical indicators, and performance is evaluated using financial metrics including return, Sharpe ratio, and maximum drawdown. The results show that although the prediction model demonstrates limited standalone classification performance, its integration significantly improves trading outcomes. The hybrid model outperforms reinforcement learning without guidance and rule-based strategies, achieving higher risk adjusted returns and more stable portfolio growth. These findings indicate that even weak predictive signals can enhance decision making when effectively integrated with reinforcement learning. The proposed framework provides a robust and adaptive solution for trading in highly volatile financial markets.