A parallel multi-module deep reinforcement learning algorithm for stock trading

摘要

In recent years, deep reinforcement learning (DRL) algorithm has been widely used in algorithmic trading. Many fully automated trading systems or strategies have been built using DRL agents, which integrate price prediction and trading signal generation in one system. However, the previous agents extract the current state from the market data without considering the long-term market historical trend when making decisions. Besides, plenty of related and useful information has not been considered. To address these two problems, we propose a novel model named Parallel Multi-Module Deep Reinforcement Learning (PMMRL) algorithm. Here, two parallel modules are used to extract and encode the feature: one module employing Fully Connected (FC) layers is used to learn the current state from the market data of the traded stock and the fundamental data of the issuing company; another module using Long Short-Term Memory (LSTM) layers aims to detect the long-term historical trend of the market. The proposed model can extract features from the whole environment by the above two modules simultaneously, taking the advantages of both LSTM and FC layers. Extensive experiments on China stock market illustrate that the proposed PMMRL algorithm achieves a higher profit and a lower drawdown than several state-of-the-art algorithms.(c) 2021 Elsevier B.V. All rights reserved.