Dual Core Portfolio Strategy: A Deep RL & Multi-Agent Portfolio Strategy

Reinforcement learning gains increasing popularity in portfolio management. However, in a complex stock trading circumstance, agent-based algorithms often face challenges such as slow convergence rates and inadequate cooperation between agents. These lead to learning inefficiencies, increased risk, and higher transaction costs. Finally, the generalizability of the trading strategy is reduced. To address these, we propose a novel multi-agent algorithm called the Dual Core Portfolio Strategy which integrates both deterministic and stochastic policies to capitalize on their complementary strengths. In this strategy, the Deep Deterministic Policy Gradient agent is proficient in deterministic policy learning, while the Soft Actor-Critic agent enhances exploration and generalization through stochastic policy. Multiple agents collaborate by making decisions and interacting with the environment, sharing a centralized critic network and their interaction trajectories. This approach strengthens the robustness and adaptability of the portfolio strategy, improving its generalizability. Experiments demonstrate that the Dual Core Portfolio Strategy model consistently outperforms traditional deep reinforcement learning models. The effectiveness is evaluated using data from 2018 to 2020 and from 2020 to 2022 for all constituent stocks in the DJIA. The DC-PS model achieves state-of-the-art results, with a minimum increase of 15.7% (from 0.213 to 0.247) in accumulated returns in 2021 and 2023, underlining its generalizability in the out-of-sample environment.