QuantArena: 9 LLMs trade A-shares — only 4 profit, none beat the 17.1% benchmark

机器之心 · wechat · 2026-10-22

EinsiaAI's NaversLab released QuantArena, a benchmark putting 9 models from 6 vendors through a full quantitative trading research loop on real A-share data: 3042 stocks, a 2022-2023 research phase (max 5 candidate evaluations), then one frozen evaluation on sealed 2024 data.

Only GPT-6 Astra (10.7% annualized), GPT-6 Sol (8.6%), GPT-5.6 Terra (8.4%) and Claude Opus 5.5 (4.4%) profited; Sonnet 5.5, Gemini 3.8 Flash, DeepSeek V4.1 Flash, Kimi K3 and GLM-5.3 all lost money. None beat the 17.1% market benchmark.
Better signals ≠ profits: Gemini and GLM had positive RankIC yet lost; 7/9 strategies' top-decile returns fell below the sample mean ("hump" pattern).
GPT-6 Astra self-wrote a timing switch (only score stocks when both 5-day and 20-day market-wide return medians are positive), dodging February's crash with a -6.8% max drawdown vs -29.6% for the market — but still trailed the benchmark by 6.3 points.
Feedback-driven iteration mostly fixed early blunders: total scores rose from 36.47 to 43.55 on average, but only 4/9 models improved out-of-sample returns.
A house strategy, Kepler-1 (score 77.56, 2024 NAV 1.93), sits on the open leaderboard as the boss to beat vs Astra's 51.29/1.11.

Original post →

QuantArena: 9 LLMs trade A-shares — only 4 profit, none beat the 17.1% benchmark

More from Venture