Liquidity takers versus providers: how cross-asset interactions shape market-making profitability

This study investigates optimal multi-asset market-making in high-frequency environments characterized by cross-asset dependencies. Building on the open-source mbt-gym environment, we extend its market-making framework to the multi-asset setting and use it to study how cross-asset interactions affect optimal liquidity provision. We address the curse of dimensionality inherent in traditional stochastic control methods by employing a reinforcement learning framework within a limit order book simulator. The resulting modular and parallelized simulation environment incorporates rich microstructure features and enables the modeling of multi-asset interactions that traditional closed-form approaches cannot effectively scale. Throughout simulation, we find significant heterogeneity in the impact of cross-asset dynamics: while increased coupling among liquidity takers enhances expected returns, stronger interactions among liquidity providers compress profit margins. Notably, we demonstrate that model misspecification, defined as the failure to incorporate these coupling effects, leads to a substantial deterioration in out-of-sample profitability. Our simulation results suggest that explicitly modeling cross-asset interactions is important for developing robust multi-asset market-making strategies in environments exhibiting strong cross-asset dependencies.