arXiv 2610.30113
Quant-Arena: Benchmarking LLM Agents for Quantitative Trading Research
AI21 upvotes
Authors: Houde Qian, Zheng Jiang, Wenyi Li, Chenxi Xu, Xinjie Lin, Yuzhao Peng, Ruoyu Han, Youjie Zheng +4
Overview
The authors built a benchmark and public leaderboard for end-to-end quantitative research on real A-share market data. Models write features and stock-scoring code using feedback only from 2022–2023, then freeze one strategy that is evaluated on the sealed 2024 year with a cash-account evaluator that applies trading constraints, costs, and order-book-based market impact.
Introduction
Existing benchmarks score individual formulas, code snippets, or single trading decisions, so it is unclear whether LLMs can run the full research loop and produce strategies that generalize beyond the period they were developed on.
Results
No one of the nine frontier models evaluated beats the equal-weight market in 2024, even though GPT-6 Astra achieves the highest score. Scores are dominated by realized excess return, with signal quality and feature diversity as complements.
Conclusion
This shows that current LLM agents still fail to produce out-of-sample profitable quant strategies, and the open leaderboard gives a standardized way to track progress.