Agent Trading Arena: A Study on Numerical Understanding in LLM-Based Agents

[arXiv] [PDF]

The Agent Trading Arena is a closed-loop, prior-free human-like trading environment designed to evaluate and advance self-play-capable financial agents.

Preparation

Python Install

git clone https://github.com/MTMQuantAI/Agent-Trading-Arena.git
cd Agent-Trading-Arena
pip install -r requirement.txt

Set OpenAI API Key:

Export your OpenAI API key as an environment variable. Replace "your_OpenAI_API_key" with your actual API key.

cd Stock_Main/content/utils.py
openai_api_key = "your_OpenAI_API_key"

OA_OPENAI_KEY is the OpenAI API key. You can get it from the OpenAI.

Experiment

Run the code

cd Agent-Trading-Arena
sh run.sh

Citation

If you find our work useful, please consider citing us!

@inproceedings{ma2025agent,
  title={Agent Trading Arena: A Study on Numerical Understanding in LLM-Based Agents},
  author={Ma, Tianmi and Du, Jiawei and Huang, Wenxin and Wang, Wenjie and Xie, Liang and Zhong, Xian and Zhou, Joey Tianyi},
  booktitle={Findings of the Association for Computational Linguistics: EMNLP 2025},
  pages={5496--5514},
  year={2025}
}