Noise Trading, Market Liquidity, and Efficiency in Adaptive Markets – a Reinforcement Learning Experiment

This study conducts a reinforcement learning (RL) experiment to investigate the behavior of noise traders under adverse selection. The experiment applies the concepts of Lo’s (2004) adaptive market hypothesis to limit order markets and offers a rational, adaptive learning-based explanation for the survival of noise traders despite persistent losses. We show that noise traders engage in learning through trading in a state-dependent manner. As RL converges to steady-state equilibrium, noise traders learn to exit the market state by state, causing liquidity drains in the market. To keep the market alive, we design a maker–taker mechanism that subsidizes limit orders and show that such a rebate program is effective for liquidity provision. Furthermore, the relationship between the level of rebates and market efficiency exhibits a U-shaped pattern, suggesting the existence of an optimal rebate size