When AI Trading Agents Compete: Adverse Selection of Meta-Orders by Reinforcement Learning-Based Market Making

Battle Count: 68

Highly relevant to quantitative trading. The paper directly addresses adverse selection in market making, a core challenge in algorithmic trading. It provides insights into how RL-based market makers interact with meta-order execution, quantifies slippage impacts, and explores the dynamics between HFT and MFT agents. The findings on asymmetric learning and the counterintuitive result that increased MM profits don't necessarily increase MFT slippage are particularly valuable for practitioners designing execution strategies and market making algorithms.

High complexity due to: (1) compound Hawkes process LOB simulation requiring careful parameterization, (2) impulse control RL framework approximating HJB-QVI, (3) dual-network architecture (decision + action networks), (4) PPO with self-imitation learning, (5) multi-agent adversarial training setup, (6) TWAP agent with urgent state logic and window-based execution. Requires expertise in stochastic processes, RL, and market microstructure.

Code is publicly available on GitHub (https://github.com/konqr/lobSimulations/tree/master/HawkesRLTrading). The paper provides detailed descriptions of the TWAP agent algorithm, RL training setup, state spaces, and simulation parameters. However, some training hyperparameters and specific Hawkes process parameters are referenced from prior work (Jain et al. 2024a, 2025) rather than fully specified here.