Learning Market Making with Closing Auctions

Battle Count: 61

Highly relevant to quantitative trading. The paper directly addresses market making execution with closing auctions, a critical component of modern equity markets. It provides practical RL-based strategies that outperform classical benchmarks (Avellaneda-Stoikov, TWAP) by 72-125% in synthetic settings and 66-123% in historical settings. The framework is applicable to end-of-day liquidation, auction participation, and inventory management for high-frequency market makers. The use of rough Heston model and S&P 500 historical data adds practical relevance.

High implementation complexity due to: (1) the multi-phase MDP with structurally different state/action spaces for CLOB and auction phases; (2) the generative stochastic market simulator with multiple interacting components (Poisson arrivals, Pareto volumes, Beta-distributed LOB, Bernoulli auction events); (3) the clearing price computation and hypothetical price projection algorithm; (4) training multiple RL agents (DQN, DDPG, TD3, SAC) with phase-specific networks; (5) the rough Heston simulation requiring fractional Brownian motion; (6) the large state space requiring careful pruning. The paper provides detailed algorithms and hyperparameters but no code.

The paper provides detailed hyperparameters (Table 1), market model parameters (Table 2), rough Heston parameters (Table 3), and full algorithmic descriptions (Algorithms 1-4). However, no code repository is mentioned. The generative stochastic model is fully specified, and historical data from S&P 500 is publicly available. The rough Heston simulation scheme is referenced from prior work [59]. Reproducibility is moderate due to the complexity of the RL training pipeline and stochastic simulations.