Stock Trading Strategy Based on Multi-Scale Deep Reinforcement Learning and Price Movement Prediction

Owing to the dynamic and complex properties of the stock market, generating a stable and highly profitable trading strategy is a challenge. Therefore, in this paper, a novel trading strategy is proposed, grounded in multi-scale deep reinforcement learning and stock price trend prediction, with supervised learning and reinforcement learning unified within a unified framework. Our proposed method is separated into three stages. First, a novel network was designed to predict stock price movement (upward, stationary, or downward) with 67.34% accuracy by combining the strengths of Convolutional Neural Network and Long-Short Term Memory. Second, leveraging the trained trend prediction network, states were enhanced with daily and weekly stock information, and multi-scale and backbone network modules were employed for effective feature extraction. Finally, a Double Deep Q-Network algorithm based on the augmented state was adopted to learn robust trading strategies. This study also contributes to the theoretical advancement of stock price prediction models by introducing a hybrid CNN-LSTM architecture for trend prediction and a multi-scale feature extraction framework for improved decision-making. Experimental results across six U.S. financial assets demonstrate that a notable accuracy of 67.34% for price trend prediction is achieved by our Convolutional Neural Network-Long-Short Term Memory network. Furthermore, the proposed trading strategy outperforms other deep reinforcement learning algorithms, yielding an average annualized return of 44.47%. By comparison, the PPO model achieves 37.60%, the TDQN model achieves 13.38%, the DQN-vanilla model achieves 28.71%, and the DQN-pattern model achieves 22.21%. These results validate the efficacy of the approach, showcasing substantial excess returns.