ABIDES-MARL: A Multi-Agent Reinforcement Learning Environment for Optimal Execution with Endogenous Liquidity

Battle Count: 53

Highly relevant to quantitative trading. The paper directly addresses optimal execution, a core problem in algorithmic trading. It demonstrates that classical exogenous-impact models (Almgren-Chriss, Kyle) produce structurally misspecified strategies when market makers adapt strategically. The findings that TWAP outperforms analytical schedules under endogenous liquidity, and that strategies optimized under exogenous impact perform poorly once market makers adapt, have direct practical implications for execution algorithm design. The framework provides a reproducible foundation for testing execution strategies against adaptive market participants.

High complexity due to: (1) need to understand discrete-event simulation (ABIDES) and reconcile it with timestep-based RL; (2) multi-agent synchronization protocol with StopSignalAgent; (3) stochastic game formulation with information filtration constraints; (4) multiple heterogeneous agent types with different objectives and information sets; (5) LOB mechanics, pro-rata allocation, and VWAP clearing; (6) PPO implementation for multiple agents; (7) careful experimental design with linear vs. nonlinear parameterizations; (8) statistical validation through AR(1) and Kyle regression diagnostics. However, the code is publicly available, reducing practical implementation burden.

All code and configurations are publicly released on GitHub. The paper provides detailed mathematical formulations, explicit hyperparameters (N=20, Q=1000, phi values, gamma values, network architecture with 2 hidden layers of 64 units, Tanh activations), and clear experimental protocols. The ABIDES-MARL framework builds on existing open-source tools (ABIDES, ABIDES-Gym). However, MARL training can exhibit seed-dependent variability, and the paper acknowledges no formal convergence guarantees.