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

Battle Count: 71

Directly addresses the core problem of electronic market making in limit order books. Provides a complete pipeline from analytical benchmarks (GLFT) through RL-based controllers to robust regime-aware fine-tuning. Highly relevant for HFT firms, market-making desks, and quantitative trading researchers working on liquidity provision, inventory management, and order-flow regime detection.

Requires implementing: (1) a queue-faithful event-driven LOB simulator with FIFO priority and per-order tracking; (2) a full Rainbow DQN with C51 distributional head, dueling architecture, PER, NoisyNet, and n-step SMDP targets; (3) Bayesian online change-point detection on the MO clock; (4) a scenario-bandit fine-tuning loop with adaptive sampling; (5) GLFT calibration via censored waiting-time estimation and volatility signature plots. The multi-stage training pipeline and simulator fidelity requirements add significant engineering complexity.

Code, trained models, and paper source are publicly available at https://github.com/felipemoret77/robust-deep-market-making. All hyperparameters are explicitly listed in Table 1. Calibration parameters (lambda, mu, theta_cxl, A, kappa, sigma) are fully specified. Paired-seed evaluation ensures deterministic comparisons. The Santa Fe simulator is well-documented in prior literature.