JaxMARL-HFT: GPU-Accelerated Large-Scale Multi-Agent Reinforcement Learning for High-Frequency Trading

Battle Count: 72

Highly relevant to quantitative trading. Directly addresses HFT market making, order execution, and directional trading. Provides a practical, open-source tool for developing and testing trading strategies in realistic LOB environments. The 240x speedup enables research previously infeasible due to computational constraints. However, the learned policies are preliminary and do not yet achieve profitability, limiting immediate practical deployment. The framework is more of a research tool than a production trading system at this stage.

Requires GPU hardware (NVIDIA L40S or similar), JAX proficiency, and understanding of MARL algorithms. The framework is modular and well-documented with GitHub code, but setting up heterogeneous agents with appropriate action/observation spaces and reward functions requires significant domain expertise in market microstructure. The vmap parallelization strategy and memory management for large datasets add complexity. Compatible with existing JaxMARL algorithms reduces some implementation burden.

Code is fully open-source on GitHub (https://github.com/vmohl/JaxMARL-HFT). Built on established open-source frameworks (JAX, JaxMARL, JAX-LOB). Uses publicly available LOBSTER data. Detailed hyperparameters provided in Table 2. Hardware specifications clearly stated (NVIDIA L40S GPUs, AMD EPYC 9554). However, the paper is very recent (Nov 2025) and results are preliminary with simplified action/observation spaces.