GPT-6 Astra Tops Quant-Arena but No Agent Beats A-Share Baseline

Nine frontier language models completed an end-to-end quantitative research loop on real Chinese A-share market data, and none of them beat the equal-weight market baseline in a sealed 2024 holdout, according to Quant-Arena, a benchmark paper from Houde Qian and colleagues posted on Hugging Face.

The setup forces models to work the way a human researcher would. Propose a market hypothesis, implement it as an executable signal, test on historical data, decide what to keep. Models receive feedback only on 2022-2023, then freeze "one strategy for assessment on the sealed year 2024." A cash-account evaluator applies trading constraints and costs, including "market impact estimated from a simulated order book." Scoring is "dominated by realized excess return and complemented by signal quality" alongside the diversity of features the strategy uses.

GPT-6 Astra came out on top of the nine. The abstract names no other model and discloses no per-model returns. The authors write that "no model beats the equal-weight market in 2024." Quant-Arena is "open to submissions from any model or agent harness."