A Comparative Analysis of Reinforcement Learning Approaches to Cryptocurrency Price Prediction

Nowadays, Machine Learning (ML) is present in a high number of application fields. Among these, there is also automatic trading in the financial sector. The research question underlying our research activities is as follows: can ML techniques provide added value in the prediction task in domains with high volatility such as the cryptocurrency financial market? To answer this question, we analyzed and compared different Reinforcement Learning (RL) algorithms on data publicly available online. Specifically, we tested some value-based and policy-based RL algorithms trained for different time intervals, with diverse hyperparameter values and reward functions. The agent that allowed us to achieve the best results was the Deep Recurrent Q-Network trained using the Sharpe ratio as a reward function.