A novel stock trading strategy based on double deep Q-network with sentiment integration

Abstract

Reinforcement learning (RL) has gained significant attention in stock trading strategies. However, existing RL models still show shortcomings. On the one hand, they fail to adequately account for the complex factors in real-world markets; on the other hand, they struggle to accurately capture the dynamic nature of financial markets, resulting in limited drawdown control and suboptimal returns. To address these challenges, we propose a novel stock trading strategy based on a Double Deep Q-Network (DDQN) with sentiment integration. First, sentiment features extracted from social media are combined with technical indicators to enhance the model’s understanding of market dynamics. Subsequently, trading decisions are made using the DDQN framework, which learns optimal policies through interaction with the market environment. To enhance performance, we adopt a Convolutional Neural Network − Bidirectional Gated Recurrent Unit (CNN–BiGRU) architecture as the Q-network, where CNN extracts local price patterns for short-term fluctuations, while BiGRU models temporal dependencies to capture long-term trends. Finally, trading signals from the RL process serve as labels to train multiple supervised classifiers. Experiments show that the proposed framework surpasses baseline models in major performance metrics including return, payoff ratio, and Sharpe ratio. This approach aims to provide accurate trading decision support for investors.