StockBench ranked 13 AI agents trading Dow stocks. The best beat buy-and-hold by about 2 points

Most finance benchmarks ask AI models questions. StockBench makes them trade. The benchmark, from a team at Tsinghua University and Beijing University of Posts and Telecommunications, was first posted on arXiv in October 2025 and revised on March 2, 2026. It is one of the few public rankings where AI agents make daily portfolio decisions on real, recent market data.

How the test works

• Universe: the top 20 stocks of the Dow Jones Industrial Average.
• Window: March 3 to June 30, 2025, or 82 trading days. The authors say the period includes both volatility and trend reversals and falls after the knowledge cutoff of mainstream models.
• Capital: $100,000 in cash per agent, with decisions made daily at the market open.
• Inputs: the agent's own actions over the past seven days, up to five news articles from the previous 48 hours per stock, and fundamentals such as market cap, P/E ratio and dividend yield.
• Actions: increase, decrease or hold each position, converted to shares at the opening price.
• Repeats: each model was run three times with different random seeds, and the average is reported.

The yardstick is a passive equal-weight portfolio of the same 20 stocks, bought on day one and held.

The ranking

Models are ranked on a combined z-score of three metrics: final return, maximum drawdown and the Sortino ratio, which rewards returns relative to downside volatility.

Three things the table says

Open models led. The top four spots went to open-weight models from Chinese labs: Moonshot's Kimi-K2, two versions of Alibaba's Qwen3 and Zhipu's GLM-4.5. OpenAI's o3 was the best closed model at fifth, and GPT-5 finished ninth with a 0.3% return, slightly below the passive portfolio's 0.4%. nContext: n=82 trading days, 3 runs per model, 20 DJIA stocks · Mar 3 to Jun 30, 2025 · Chen et al., arXiv:2510.02209 v2, Table 2 · margin: not disclosed

Drawdowns were smaller than buy-and-hold. Every agent had a smaller maximum drawdown than the passive portfolio's -15.2%. That fits a window with a sharp spring sell-off: agents that trimmed positions lost less at the bottom. nContext: n=82 trading days, 3 runs per model, 20 DJIA stocks · Mar 3 to Jun 30, 2025 · Chen et al., arXiv:2510.02209 v2, Table 2 · margin: not disclosed

Reasoning did not reliably help. Qwen3-235B in its instruction version beat its reasoning version on drawdown (-11.2% against -14.9%). The authors conclude there is still a gap between reasoning ability and decision-making in noisy markets. nContext: n=82 trading days, 3 runs per model, 20 DJIA stocks · Mar 3 to Jun 30, 2025 · Chen et al., arXiv:2510.02209 v2, Table 2 · margin: not disclosed

The paper disagrees with itself

Read the abstract and you learn that "most models struggle to outperform the simple buy-and-hold baseline." Read the results section and you learn the opposite: "LLM agents can trade profitably in real-world markets. Most tested models outperform the passive buy-and-hold baseline."

The table supports the second reading on raw return: 10 of the 13 models finished above 0.4%. But the margins are small. The best final return, 2.5%, is about two points above the passive portfolio over four months, and no confidence intervals are given. With three runs per model and one window, a two-point gap is hard to tell apart from noise. nContext: n=82 trading days, 3 runs per model, 20 DJIA stocks · Mar 3 to Jun 30, 2025 · Chen et al., arXiv:2510.02209 v2, Table 2 · margin: not disclosed

Two other details matter for anyone tempted to read this as a trading result. We found no mention of transaction costs or commissions in the paper, so daily rebalancing appears to be free. And the authors' own scaling test points the wrong way for real use: when the portfolio grew from 5 to 30 Dow stocks, all models showed declining mean returns and rising volatility. Kimi-K2 was the most robust.

Where it fits

StockBench complements the long-horizon picture from the 20-year backtest of LLM strategies and the live record in hundreds of AI trading agents with no edge. It is short, recent and post-cutoff, which removes the memory problem but leaves the sample-size problem. The authors say the environment will be updated with new windows. We will track the next one.