An Impulse Control Approach to Market Making in a Hawkes LOB Market

Battle Count: 96

Highly relevant to quantitative trading, particularly high-frequency market making. The paper addresses realistic LOB microstructure (queue dynamics, clustered arrivals, endogenous price impact) using Hawkes processes, formulates the problem within impulse control (accounting for discrete intervention constraints), and demonstrates practical RL solutions achieving Sharpe ratios above 30 in simulation. The sensitivity analysis on transaction costs, inventory penalties, and kernel types provides actionable insights for practitioners. The comparison between model-based (DGM) and model-free (RL) approaches is directly relevant to strategy development.

High implementation complexity due to: (1) multi-dimensional mutually-exciting Hawkes process simulation, (2) 23-dimensional state space with queue dynamics, (3) HJB-QVI formulation with non-local operators, (4) two-network PPO architecture with self-imitation learning, (5) DGM with LSTM networks and automatic differentiation for PDE residuals, (6) careful handling of jump equations and state-intervention operators. The RL approach is more tractable than the DGM approach but still requires significant engineering for the Hawkes simulator and training infrastructure.

The paper provides detailed mathematical formulations, state/action spaces, reward structures, and hyperparameters. However, no code repository is mentioned. The Hawkes LOB simulator is referenced from prior work [27]. Training details (60 episodes, 300-second horizon, 0.1s decision frequency) are specified. The DGM architecture (LSTM, 3-10 layers, 20-50 neurons) is described. Reproducibility is moderate given the complexity of the Hawkes process simulation and the specific simulator from prior work.