Modular Intelligent Stock Trading Framework with Meta-Strategy

To address the challenges of training agents for multi-stock trading in automated reinforcement learning systems, this paper (Modular Intelligent Stock Trading Framework with Meta-Strategy) proposes a modular deep reinforcement learning (DRL) framework that integrates multiple trading agents—Advantage Actor-Critic (A2C), Proximal Policy Optimization (PPO), and Recurrent PPO (R-PPO)—with a meta-strategy decision layer.

For the system’s trading agent, we design a cross-stock training environment with a 100-dimensional state feature input and a novel reward function to enhance learning performance. We also introduce an adaptive agent selection mechanism and trading constraints to improve stability, generalization, and profitability in non-stationary financial environments. The framework, trained and validated on over 15 years of real-world NASDAQ data, achieves superior cumulative returns and Sharpe ratios compared with single DRL agents, buy-and-hold strategies, and traditional benchmarks such as the NASDAQ Composite Index.