# Market-Bench leaderboard

Models write executable backtesters from plain-language descriptions of three strategies: scheduled trading on MSFT, KO/PEP pairs trading, and MSFT delta hedging. Outputs are compared with a reference implementation.

## Market-Bench: mean MAE (lower is better, 10 models)

| Rank | Model | Maker | Score | Metric | Setting | Reported by | Source | URL | Date |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | grok-4 | xAI | 443.2 | mean MAE |  | Official board | marketbench.ai leaderboard | https://marketbench.ai | date not stated |
| 2 | gpt-5.2 | OpenAI | 969.4 | mean MAE |  | Official board | marketbench.ai leaderboard | https://marketbench.ai | date not stated |
| 3 | gemini-3-pro-preview | Google | 1744.3 | mean MAE |  | Official board | marketbench.ai leaderboard | https://marketbench.ai | date not stated |
| 4 | gpt-5.1-codex-max | OpenAI | 4242.9 | mean MAE |  | Official board | marketbench.ai leaderboard | https://marketbench.ai | date not stated |
| 5 | deepseek-v3.2 | DeepSeek | 4575.6 | mean MAE |  | Official board | marketbench.ai leaderboard | https://marketbench.ai | date not stated |
| 6 | claude-sonnet-4.5 | Anthropic | 5126.8 | mean MAE |  | Official board | marketbench.ai leaderboard | https://marketbench.ai | date not stated |
| 7 | claude-opus-4.5 | Anthropic | 6039.6 | mean MAE |  | Official board | marketbench.ai leaderboard | https://marketbench.ai | date not stated |
| 8 | command-a | Cohere | 6562.1 | mean MAE |  | Official board | marketbench.ai leaderboard | https://marketbench.ai | date not stated |
| 9 | nova-premier-v1 | Amazon | 7740.3 | mean MAE |  | Official board | marketbench.ai leaderboard | https://marketbench.ai | date not stated |
| 10 | llama-3.1-nemotron-ultra | Meta | 9674.2 | mean MAE |  | Official board | marketbench.ai leaderboard | https://marketbench.ai | date not stated |

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

## About this benchmark

- Maker: AfterQuery
- Year: 2025
- Use case: Financial trading and portfolios
- Size: 3 strategies (easy, medium, hard), 5 runs per model
- Public split: Not stated
- Grading: Mean absolute error of P&L, drawdown and position paths against a reference backtester (lower is better). Executability is reported as pass@1 and pass@3. Failed runs get a baseline MAE.
- Licence: Not stated
- Submission: Not stated
- Products allowed: Not stated
- Latest version: paper v2 (2026-01-20)
- What it does not cover: Coding of introductory strategies only. No live trading, strategy quality or risk judgement.
- Last checked: 1 Nov 2026
- Links: paper https://arxiv.org/abs/2512.12264, site https://marketbench.ai, leaderboard https://marketbench.ai
- Directory page: https://benchmarks.finance/benchmarks/afterquery-market-bench/
- Leaderboard: https://benchmarks.finance/leaderboards/afterquery-market-bench/

Source: https://benchmarks.finance/leaderboards/afterquery-market-bench/

Created and maintained by Moneda.