A synchronous deep reinforcement learning model for automated multi-stock trading

Abstract

Automated trading is one of the research areas that has benefited from the recent success of deep reinforcement learning (DRL) in solving complex decision-making problems. Despite the large number of researches done, casting the stock trading problem in a DRL framework still remains an open research area due to many reasons, including dynamic extraction of financial data features instead of handcrafted features, applying a scalable DRL technique that can benefit from the huge historical trading data available within a reasonable time frame and adopting an efficient trading strategy. In this paper, a novel multi-stock trading model is presented, based on free-model synchronous multi-agent deep reinforcement learning, which is able to interact with the trading market and to capture the financial market dynamics. The model can be executed on a standard personal computer with multiple core CPU or a GPU in a convenient time frame. The superiority of the proposed model is verified on datasets of different characteristics from the American stock market with huge historical trading data.