Can LLM agents conduct quantitative research?
A benchmark that evaluates the entire quantitative research process, from feature engineering to out-of-sample trading performance.
Paper
Quant-Arena: Benchmarking LLM Agents for Quantitative Trading Research
Authors
Houde Qian, Zheng Jiang, Wenyi Li, Chenxi Xu, Xinjie Lin, Yuzhao Peng, Ruoyu Han, Youjie Zheng, Kaisen Yang, Qinhuai Na, You Fu, and Calvin Xiao
Date
21 October 2026
Most benchmarks for large language models in finance evaluate individual tasks such as generating formulas, writing code or making trading decisions. Quant-Arena takes a broader approach, assessing whether an LLM agent can complete the full quantitative research process and produce a strategy that performs on unseen market data.
The benchmark uses real Chinese A-share market data. Agents construct features and stock-scoring code, receiving feedback on historical data from 2022 and 2023. They must then freeze their strategy before it is evaluated on a sealed 2024 dataset. A cash-account evaluator incorporates trading constraints and transaction costs, including market impact estimated using a simulated order book.
Nine frontier models were evaluated. Although GPT-6 Astra achieved the highest benchmark score, none outperformed an equal-weight market portfolio in the 2024 assessment.
The results highlight the difficulty of translating LLM-generated research into profitable strategies. The benchmark currently covers a single market and one out-of-sample year, but provides a framework for evaluating agents on research outcomes rather than isolated tasks.
