Quant-Arena: no frontier LLM beats equal-weight A-share market

Nine frontier large language models were set loose on Chinese A-share market data, told to design a trading strategy of their own, and graded on a year of sealed trading they had never seen. None of them beat an equal-weight basket of every stock in the market.

That is the headline result of Quant-Arena, a new benchmark and public leaderboard aimed at settling, in a falsifiable way, whether LLMs can run the full quantitative-research loop end to end. Each model constructed features and stock-scoring code, received feedback only on 2022 and 2023 data, and froze one strategy "for assessment on the sealed year 2024." A cash-account evaluator then applied trading constraints and costs, "including market impact estimated from a simulated order book."

GPT-6 Astra scored highest of the nine. But as the authors put it, "no model beats the equal-weight market in 2024." Scores combine realized excess return with signal quality and the diversity of features a strategy uses.

The abstract does not list the other eight models, and it does not say by how much each one trailed the equal-weight baseline. The leaderboard is open to submissions from any model or agent harness.