Improving Deep Reinforcement Learning for Financial Trading Using Deep Adaptive Group-Based Normalization

Deep Reinforcement Learning methods have provided powerful tools to train profitable agents for financial trading. However, the noisy and non-stationary nature of financial data often requires carefully designed and tuned input normalization schemes, since otherwise the agents are unable to consistently perform profitable trades. To overcome this limitation, in this work we propose a deep adaptive input normalization approach specifically designed to train DRL agents for financial trading directly using the raw price as input, without any additional pre-processing. The proposed method consists of two trainable neural layers that are designed to perform adaptive normalization, i.e., normalize the input observations after (implicitly) identifying the distribution that was used for generating them. Furthermore, instead of normalizing the whole input at once, the proposed approach performs group-based normalization, which allows for better capturing fine variations in the price trends. Despite being simple to implement and apply, the proposed method can lead to enormous improvements over existing the normalization methods, as demonstrated through the experiments conducted on two challenging FOREX currency pairs.