Navers lab
← Research

Quant-Arena:Benchmarking LLM Agents for Quantitative Trading Research

Can a language model research a trading strategy, commit to it, and have it hold up on a year it has never seen?

LaPlace's Eye · Einsia AI · 2026

Live Arena Leaderboard arXiv 机器之心
Quant-Arena evaluates LLMs on complete quantitative research runs; banners show total scores on the iterative research track.

The task

Each model acts as a quantitative researcher. It researches on 2022–2023 with bounded feedback, commits one strategy, and is assessed once on a sealed 2024 window in a cash account. The score combines the research-window result (Sin) with the sealed 2024 result (Sout), weighted toward the out-of-sample year.

  1. 01 Research

    The model studies 2022–2023 data and receives bounded feedback on the candidates it submits.

  2. 02 Commit

    It commits exactly one strategy. Nothing can be changed after this point.

  3. 03 Sealed assessment

    The committed strategy is assessed once on a sealed 2024 window, trading in a cash account.

Iterative research

The model researches on 2022–2023 with bounded feedback on its candidates, then commits one strategy.

One-shot generation

The model writes one strategy in a single pass, without research-window feedback.

Key findings

No model beats the market

In the sealed 2024 test year, the equal-weight market returned 17.1%. No model's best strategy kept pace: the strongest gained 10.7%, and 5 of the 9 models lost money.

Ranking skill does not reach the portfolio

Most models rank stocks better than chance, yet the stocks they rank highest, the ones the portfolio actually buys, earn less than those ranked just below. Average rank correlation says little about whether a strategy makes money.

Some models learn to time the market

After research feedback showed losses, the top-scoring model wrote a market-timing rule: hold cash whenever the market trends down. It sat out the February 2024 crash, with a drawdown of −1.7% against −29.6% for the market. Because it also missed much of the rebound, it still finished behind the market for the year.

Leaderboard

Each model at its best-scoring configuration in each setting. The 2024 columns describe the committed iterative strategy on the sealed year. The full leaderboard lists every configuration.

Quant Arena leaderboard
# Model Iterative One-shot 2024 ARR Excess (pp) Sharpe MDD
1 GPT-6 Astra 51.29 35.22 10.7% −6.3 0.86 −6.8%
2 GPT-6 Sol 51.14 32.15 8.6% −8.4 1.89 −2.8%
3 GPT-5.6 Terra 47.58 42.14 8.4% −8.7 0.42 −28.8%
4 Claude Opus 5.5 44.28 42.61 4.4% −12.7 0.29 −32.7%
5 Gemini 3.8 Flash 43.04 35.91 −2.9% −19.9 0.06 −33.9%
6 Claude Sonnet 5.5 40.93 40.24 −12.2% −29.2 −0.23 −36.1%
7 GLM-5.3 38.91 38.02 −7.8% −24.9 −0.18 −27.8%
8 Kimi K3 38.31 32.01 −8.5% −25.6 −0.10 −36.4%
9 DeepSeek V4.1 Flash 36.45 29.92 −5.7% −22.7 −0.06 −30.3%
— Kepler-1 Einsia AI 77.56 94.0% +76.9 2.98 −15.9%

The equal-weight market returned 17.1% in 2024. Kepler-1 is a proprietary strategy developed by Einsia AI; its construction is not disclosed.

Inside the strategies

Three views of the strategy each model committed at its best-scoring iterative configuration, all on the sealed 2024 window.

2024 cash-account NAV
0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 Jan 2024Apr 2024Jul 2024Oct 2024 Claude Opus 5.5: 1.04 at year end Claude Sonnet 5.5: 0.88 at year end GPT-6 Astra: 1.11 at year end GPT-6 Sol: 1.09 at year end GPT-5.6 Terra: 1.08 at year end Gemini 3.8 Flash: 0.97 at year end DeepSeek V4.1 Flash: 0.94 at year end Kimi K3: 0.91 at year end GLM-5.3: 0.92 at year end Equal-weight market: 1.17 at year end Market 1.17 GPT-6 Astra 1.11 GPT-6 Sol 1.09 GPT-5.6 Terra 1.08 Opus 5.5 1.04 Gemini 3.8 0.97 DeepSeek V4.1 0.94 GLM-5.3 0.92 Kimi K3 0.91 Sonnet 5.5 0.88

Net asset value through 2024, starting from 1. No model beats the equal-weight market (dashed), and only four finish above their starting capital. GPT-6 Astra's flat start comes from its market-timing rule: it holds cash through the early-2024 decline.

Return by signal decile
GPT-6 Astra RankIC 0.101 · H−L +38.7 bp
-10 0 80 Decile 1: -2.7 bp Decile 2: 25.4 bp Decile 3: 31.7 bp Decile 4: 28.9 bp Decile 5: 26.3 bp Decile 6: 32.0 bp Decile 7: 26.5 bp Decile 8: 33.3 bp Decile 9: 33.0 bp Decile 10: 36.0 bp 1 10
GPT-6 Sol RankIC 0.053 · H−L −2.2 bp
-10 0 80 Decile 1: 65.8 bp Decile 2: 73.2 bp Decile 3: 67.8 bp Decile 4: 67.9 bp Decile 5: 65.6 bp Decile 6: 66.1 bp Decile 7: 65.1 bp Decile 8: 74.5 bp Decile 9: 0.0 bp Decile 10: 63.6 bp 1 10
GPT-5.6 Terra RankIC 0.026 · H−L +28.6 bp
-10 0 80 Decile 1: -9.8 bp Decile 2: 6.9 bp Decile 3: 16.5 bp Decile 4: 14.5 bp Decile 5: 18.9 bp Decile 6: 18.5 bp Decile 7: 20.3 bp Decile 8: 20.1 bp Decile 9: 17.2 bp Decile 10: 18.8 bp 1 10
Claude Opus 5.5 RankIC -0.002 · H−L +4.7 bp
-10 0 80 Decile 1: 1.5 bp Decile 2: 27.5 bp Decile 3: 23.9 bp Decile 4: 18.8 bp Decile 5: 15.6 bp Decile 6: 17.4 bp Decile 7: 13.6 bp Decile 8: 11.3 bp Decile 9: 7.3 bp Decile 10: 6.2 bp 1 10
Gemini 3.8 Flash RankIC 0.045 · H−L 0.0 bp
-10 0 80 Decile 1: 6.9 bp Decile 2: 21.4 bp Decile 3: 26.6 bp Decile 4: 21.0 bp Decile 5: 18.0 bp Decile 6: 17.8 bp Decile 7: 14.9 bp Decile 8: 12.2 bp Decile 9: 9.2 bp Decile 10: 6.9 bp 1 10
Claude Sonnet 5.5 RankIC 0.045 · H−L −0.4 bp
-10 0 80 Decile 1: 5.8 bp Decile 2: 32.5 bp Decile 3: 29.6 bp Decile 4: 28.7 bp Decile 5: 28.3 bp Decile 6: 23.4 bp Decile 7: 22.2 bp Decile 8: 17.0 bp Decile 9: 13.2 bp Decile 10: 5.4 bp 1 10
GLM-5.3 RankIC 0.046 · H−L −2.8 bp
-10 0 80 Decile 1: 6.8 bp Decile 2: 26.5 bp Decile 3: 27.8 bp Decile 4: 22.7 bp Decile 5: 20.1 bp Decile 6: 18.3 bp Decile 7: 17.3 bp Decile 8: 12.6 bp Decile 9: 12.6 bp Decile 10: 4.0 bp 1 10
Kimi K3 RankIC 0.027 · H−L +1.3 bp
-10 0 80 Decile 1: 0.5 bp Decile 2: 24.5 bp Decile 3: 24.2 bp Decile 4: 23.1 bp Decile 5: 17.3 bp Decile 6: 16.7 bp Decile 7: 12.8 bp Decile 8: 11.9 bp Decile 9: 8.3 bp Decile 10: 1.8 bp 1 10
DeepSeek V4.1 Flash RankIC -0.009 · H−L −10.5 bp
-10 0 80 Decile 1: 13.5 bp Decile 2: 21.5 bp Decile 3: 25.6 bp Decile 4: 18.0 bp Decile 5: 17.9 bp Decile 6: 13.7 bp Decile 7: 12.3 bp Decile 8: 8.3 bp Decile 9: 9.2 bp Decile 10: 2.9 bp 1 10

Mean return of the following period, in basis points, for stocks in each signal decile (1 = lowest signal, 10 = highest; the peak and the top decile are highlighted). For most models returns peak in the second or third decile and fall toward the top decile, where the portfolio buys, so a higher RankIC does not translate into portfolio return. GPT-6 Astra and GPT-5.6 Terra rise across deciles, and GPT-6 Sol is roughly flat.

Investment styles
Style exposures of each model's best iterative configuration
Model SizeValue (B/P)Value (E/P)MomentumVolatilityTurnover
GPT-6 Astra −0.62+0.96+0.11−0.24−1.05−0.74
GPT-6 Sol −0.63+0.65−1.00−0.48−0.29−0.46
GPT-5.6 Terra −0.11+0.05−0.10−0.06−0.04−0.08
Claude Opus 5.5 −0.10−0.24−0.43+0.01+0.54+0.56
Gemini 3.8 Flash −0.60+0.54+0.03−0.21−0.52−0.48
Claude Sonnet 5.5 −0.50+0.54−0.17−0.21−0.64−0.61
GLM-5.3 −0.15+0.49+0.16−0.10−0.63−0.47
Kimi K3 −0.85+0.10−0.52−0.38−0.07−0.35
DeepSeek V4.1 Flash +0.12−0.37−0.11+0.13+0.46+0.57

Average standardised style exposure of each strategy's holdings, where 0 is the market. Most models tilt toward small caps, high book-to-price (B/P) and low volatility; Claude Opus 5.5 and DeepSeek V4.1 Flash lean toward higher volatility and turnover, and GPT-5.6 Terra stays close to the market on every style.

← Back to Research