# InvestorBench leaderboard

Backtested single-asset trading decisions by an LLM agent framework (with memory) across stocks, crypto and an ETF, with many LLMs swapped in as the backbone.

## InvestorBench: cumulative return % (higher is better, 10 models)

| Rank | Model | Maker | Score | Metric | Setting | Reported by | Source | URL | Date |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | Qwen2.5-72B-Instruct | Qwen | 46.2 | cumulative return % | stock task, average over 7 stocks, 2020-10-01 to 2021-05-06; buy-and-hold 34.099 | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2412.18174v1 | 24 Dec 2024 |
| 2 | GPT-4 | OpenAI | 43.7 | cumulative return % | stock task, 7-stock average | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2412.18174v1 | 24 Dec 2024 |
| 3 | GPT-4o | OpenAI | 39.0 | cumulative return % | stock task, 7-stock average | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2412.18174v1 | 24 Dec 2024 |
| 4 | Llama-3.1-70B-Instruct | Meta | 38.9 | cumulative return % | stock task, 7-stock average | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2412.18174v1 | 24 Dec 2024 |
| 5 | Yi-1.5-34B-Chat | Not stated | 38.0 | cumulative return % | stock task, 7-stock average | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2412.18174v1 | 24 Dec 2024 |
| 6 | Qwen-2.5-Instruct-7B | Qwen | 29.5 | cumulative return % | stock task, 7-stock average | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2412.18174v1 | 24 Dec 2024 |
| 7 | DeepSeek-V2-Lite (15.7B) | DeepSeek | 28.7 | cumulative return % | stock task, 7-stock average | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2412.18174v1 | 24 Dec 2024 |
| 8 | DeepSeek-67B-Chat | DeepSeek | 26.9 | cumulative return % | stock task, 7-stock average | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2412.18174v1 | 24 Dec 2024 |
| 9 | Llama-3.1-8B-Instruct | Meta | 25.5 | cumulative return % | stock task, 7-stock average | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2412.18174v1 | 24 Dec 2024 |
| 10 | GPT-o1-preview | OpenAI | 25.1 | cumulative return % | stock task, 7-stock average | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2412.18174v1 | 24 Dec 2024 |

Scores are never combined across benchmarks. Each row names its metric, source and date.

## About this benchmark

- Maker: Stevens Institute of Technology; The Fin AI; Columbia University; Harvard University
- Year: 2024
- Use case: Financial trading and portfolios
- Size: 3 tasks: 7 stocks, 2 cryptos (BTC, ETH), 1 ETF (NIFTY-based); 13 LLM backbones
- Public split: Paper says datasets and environments are open source; no repository URL found on any fetched primary page
- Grading: Cumulative return and Sharpe ratio vs buy-and-hold
- Licence: MIT (data, per paper's ethics statement)
- Submission: none
- Products allowed: Not stated
- Latest version: ACL 2025 Long Papers (July 2025); arXiv v1 (2024-12-24)
- What it does not cover: Live or contamination-free periods (test windows are 2020-2023 history) and multi-asset portfolios
- Last checked: 1 Nov 2026
- Links: paper https://aclanthology.org/2025.acl-long.126, site https://suchow.io/investorbench/
- Directory page: https://benchmarks.finance/benchmarks/investorbench/
- Leaderboard: https://benchmarks.finance/leaderboards/investorbench/

Source: https://benchmarks.finance/leaderboards/investorbench/

Created and maintained by Moneda.