Improving Generalization in Reinforcement Learning–Based Trading by Using a Generative Adversarial Market Model

With the increasing sophistication of artificial intelligence, reinforcement learning (RL) has been widely applied to portfolio management. However, shortcomings remain. Specifically, because the training environment of an RL-based portfolio optimization framework is usually constructed based on historical price data in the literature, the agent potentially 1) violates the definition of a Markov decision process (MDP), 2) ignores their own market impact, or 3) fails to account for causal relationships within interaction processes; these ultimately lead the agent to make poor generalizations. To surmount these problems-specifically, to help the RL-based portfolio agent make better generalizations-we introduce an interactive training environment that leverages a generative model, called the limit order book-generative adversarial model (LOB-GAN), to simulate a financial market. Specifically, the LOB-GAN models market ordering behavior, and LOB-GAN's generator is utilized as a market behavior simulator. A simulated financial market, called Virtual Market, is constructed by the market behavior simulator in conjunction with a realistic security matching system. Virtual Market is then leveraged as an interactive training environment for the RL-based portfolio agent. The experimental results demonstrate that our framework improves out-of-sample portfolio performance by 4%, which is superior to other generalization strategies.