Sentiment-influenced trading system based on multimodal deep reinforcement learning

Abstract

Owing to advancements in deep learning, studies involving the use of deep learning techniques to solve investment decision-making problems are increasing. However, although numerous aspects of the stock market may affect trends in financial data, previous studies have only considered price fluctuations. Therefore, investors may lose out on profits because of the complicated financial market condition. In this study, a multimodal reinforcement trading system is developed, which makes use of three techniques: reinforcement learning, sentiment analysis, and multimodal learning. The agent considers not only the price fluctuations but also news information when making a trading decision. Multimodal learning which can merge different modalities of data to enhance the performance of the model, and sentiment analysis for understanding the sentiment of news are introduced. In addition, an influence model is proposed to enable our agents to gain special insights on the impact that news has on the market. The influence model considers the relationship between sentiment of news and time. The experimental results show that multimodal agents outperform price-concerned agents by at least 13.26%. Our experimental results also indicate that the proposed influence model has the ability to shape the impact of news on the stock market. The model can aid the multimodal agents in evaluating the status of the market. The proposed multimodal reinforcement trading system is demonstrated to be robust in an experiment involving different sectors and evaluations by using various measures. In addition, because the data used are public, investors seeking to profit can easily implement the results of this paper. Therefore, it can be used in advanced research and financial applications.