Tsinghua University; Beijing University of Posts and Telecommunications, 2025
StockBench
Whether LLM agents can trade a basket of large US stocks profitably over several months. Each day the agent gets prices, fundamentals and news and chooses buy, sell or hold.
About this benchmark
- Size
- 20 DJIA stocks, 82 trading days (2025-03-03 to 2025-06-30), 13 models plus a passive baseline, 3 seeds each
- Public split
- Code on GitHub; results on the project page
- Grading
- Final return, max drawdown and Sortino ratio, combined into a composite rank by z-score aggregation
- Licence
- Apache-2.0 (code); paper says academic, non-commercial research use only
- Submission
- maker-only
- Products allowed
- Not stated
- Latest version
- arXiv v2 (2026-03-02)
- What it does not cover
- Markets beyond 20 US large caps, and periods after June 2025
Results
11 modelsEach score shows who reported it: the benchmark's authors (quoted from their paper), the benchmark's official board, or Moneda (run with the benchmark's own grader and rules). We do not copy other aggregators.
StockBench arXiv v2 (2026-03-02)
final return %, higher is better. 11 models.
Updated 2 Mar 2026
- Qwen3-235B-Think2.5
- Qwen3-235B-Ins2.4
- GLM-4.52.3
- Claude-4-Sonnet2.2
- Qwen3-30B-Think2.1
- Kimi k21.9
- OpenAI-O31.9
- DeepSeek V3.11.1
- gpt-50.300
- DeepSeek-V30.200
- Qwen3-Coder0.200
StockBench arXiv v2 (2026-03-02)
final return %, higher is better. Top 10 of 11 models.
Updated 2 Mar 2026
- 1Qwen3-235B-Think2.5
- 2Qwen3-235B-Ins2.4
- 3GLM-4.52.3
- 4Claude-4-Sonnet2.2
- 5Qwen3-30B-Think2.1
- 6Kimi k21.9
- 6OpenAI-O31.9
- 8DeepSeek V3.11.1
- 9gpt-50.300
- 10DeepSeek-V30.200
| 1 | Qwen3-235B-Think | Qwen | 2.5 | composite rank 4 | Paper Table 2 Reported by the authors, 2 Mar 2026 |
|---|---|---|---|---|---|
| 2 | Qwen3-235B-Ins | Qwen | 2.4 | composite rank 2 | Paper Table 2 Reported by the authors, 2 Mar 2026 |
| 3 | GLM-4.5 | Zhipu AI | 2.3 | composite rank 3 | Paper Table 2 Reported by the authors, 2 Mar 2026 |
| 4 | Claude-4-Sonnet | Anthropic | 2.2 | composite rank 7 | Paper Table 2 Reported by the authors, 2 Mar 2026 |
| 5 | Qwen3-30B-Think | Qwen | 2.1 | composite rank 6 | Paper Table 2 Reported by the authors, 2 Mar 2026 |
| 6 | Kimi k2 | Moonshot AI | 1.9 | composite rank 1 (z-score of return, drawdown, Sortino); avg of 3 runs | Paper Table 2 (same as project leaderboard) Reported by the authors, 2 Mar 2026 |
| 6 | OpenAI-O3 | OpenAI | 1.9 | composite rank 5 | Paper Table 2 Reported by the authors, 2 Mar 2026 |
| 8 | DeepSeek V3.1 | DeepSeek | 1.1 | composite rank 8 | Paper Table 2 Reported by the authors, 2 Mar 2026 |
| 9 | gpt-5 | OpenAI | 0.300 | composite rank 9 | Paper Table 2 Reported by the authors, 2 Mar 2026 |
| 10 | DeepSeek-V3 | DeepSeek | 0.200 | composite rank 11 (passive baseline is rank 12 at +0.4%) | Paper Table 2 Reported by the authors, 2 Mar 2026 |
| 10 | Qwen3-Coder | Qwen | 0.200 | composite rank 10 | Paper Table 2 Reported by the authors, 2 Mar 2026 |
Papers and links
Last checked 1 Nov 2026. Something out of date or wrong? Send a correction.