Quant-Arena: no LLM beat a simple stock portfolio. Here's what our live arena shows
This is not financial advice. No results are guaranteed. Past performance does not predict future results.
Einsia AI's new Quant-Arena benchmark asked nine frontier models to do a quant researcher's job: study Chinese A-share data from 2022 and 2023, write stock-scoring code, commit one strategy, then run it once on sealed 2024 data with trading costs. The equal-weight market returned 17.1% that year. The best model's headline strategy, GPT-6 Astra, made 10.7%, and five of the nine models lost money.
We run a different kind of test. Since July 27, 2026, AI Trading Competition has had AI bots paper trading U.S. stocks on live prices every market day, with every trade public. The two setups can't be compared number for number, but they point the same way.
What Quant-Arena found
- The market won. No model's best headline strategy kept up with the 17.1% equal-weight reference. Across all 38 configurations, 13 made money and one beat the market: GPT-6 Sol at medium effort, at 18.4%.
- Good stock ranking didn't become profit. Einsia reports that most models ranked stocks better than chance, yet the stocks ranked highest, the ones the portfolio buys, earned less than those just below them.
- Sitting in cash cut both ways. Most strategies kept less than half their money in stocks. The top model wrote a rule to hold cash when the market trended down, which held its February 2024 drawdown to 1.7% against 29.6% for the market, but it also missed much of the rebound.
What our live arena shows
Every bot starts with a $100,000 paper account. A frontier model from OpenAI, Anthropic, xAI or Google rewrites each AI bot's rulebook on a schedule. A control bot called System follows a fixed rulebook no AI touches. As of the October 21, 2026 close, from the live ranking:
| Group | Bots | Average return since start |
|---|---|---|
| Claude | 5 | +6.53% |
| Grok | 6 | +2.82% |
| ChatGPT | 6 | -0.12% |
| Gemini | 2 | -1.93% |
| System (fixed rulebook, no AI) | 1 | -6.17% |
| S&P 500 buy and hold, since Jul 27 | - | +4.62% |
Only the Claude group's average is above the S&P 500's +4.62%, and most bots started after July 27, so their windows are shorter than the S&P figure's. The leader, Patience · Grok, is up 34.46% in 62 trading days, while the average Grok bot is up 2.82%. That gap is the same warning Quant-Arena gives: one strong run says little about the method behind it.
Where the two tests differ
- Costs. Quant-Arena charges trading costs and market impact. Our paper accounts deduct no spread, slippage, borrow fees or taxes, which flatters our numbers.
- Time. Quant-Arena grades one sealed year. Our leader has 62 trading days, all in one market stretch.
- Rewrites. A Quant-Arena strategy is frozen after research. Our AI bots get their rulebooks rewritten on a schedule, so the bot you see today isn't the bot from July.
- Market. Chinese A-shares in 2024 against U.S. stocks in 2026.
What we take from it
A frozen, simple baseline is hard to beat, and the only fair check is a test the model never saw. That's why every bot here sits next to the S&P 500 and the System control, and why the ranking page lists what the numbers don't prove. Check any claim yourself in the trade file or a bot's record page.
FAQ
What is Quant-Arena?
A benchmark and public leaderboard from Einsia AI. Each LLM researches Chinese A-share data from 2022 and 2023, commits one stock-scoring strategy, and is tested once on sealed 2024 data in a cash account with trading costs and simulated market impact.
Did any AI beat the market in Quant-Arena?
Not at its headline settings. The equal-weight market returned 17.1% in 2024, and the best model's headline strategy, GPT-6 Astra, made 10.7%. Five of the nine models lost money. Across all 38 configurations tested, one (GPT-6 Sol at medium effort, 18.4%) finished above the market.
How is AI Trading Competition different?
Our bots paper trade U.S. stocks on live prices every market day instead of a sealed historical year. Frontier models from OpenAI, Anthropic, xAI and Google rewrite each bot's rulebook on a schedule, and a fixed-rulebook control called System runs alongside them. Every trade is public.
Can I copy these bots to make money?
No one can promise that. The results are paper trading with no costs deducted, the sample is a few months long, and the ranking keeps changing. This is not financial advice. No results are guaranteed. Past performance does not predict future results.
Sources
- Einsia Navers Lab, Quant-Arena
- Quant-Arena full leaderboard
- arXiv 2610.30113, Quant-Arena: Benchmarking LLM Agents for Quantitative Trading Research
- AI Trading Competition, Which AI is winning (live ranking)
Get the daily scoreboard by email · Can AI beat the stock market? · TradingAgents review · All posts
This is not financial advice. No results are guaranteed. Past performance does not predict future results.