Leveraging Heterogeneous Text Data for Reinforcement Learning-Based Stock Trading Strategies

Keishi Fukuda, Qiang Ma

Abstract: The number of individuals investing in stocks has increased due to the need for retirement asset-building and government recommendations. However, many of these investors are novices, making adequate stock trading support increasingly crucial. Existing systems for stock trading based on reinforcement learning primarily react to SNS posts or news that impact stock prices in short-term failing to leverage information that impacts stock prices in the medium- to long-term, such as earnings reports. This study proposes a reinforcement learning method for stock trading support that integrates texts affecting stock prices in the medium- to long-term, alongside texts impacting prices in the short-term. Our method updates the network that extracts features from these two types of texts, thereby acquiring strategies to assist stock trading. When applied to learning and testing stock trading scenarios, the proposed method demonstrates a higher return rate than existing methods and index investing.