Deep Learning of Robust Market Making under Regime-Switching Order Flow

RESUMO

Classical market-making strategies based on stochastic control provide closed-form quoting rules, but rest on assumptions that break down at realistic microstructure timescales. We develop a deep reinforcement-learning market maker (RLMM), a Rainbow-style distributional DQN, calibrated and tested in a zero-intelligence limit order book. RLMM outperforms GLFT across the entire observed risk-return frontier and is more robust to flow asymmetry, though it still suffers drawdowns from inventory saturation under persistent directional imbalance.