Market-Bench - Pairs Mean-Reversion (Mean of 5)
Professional Work · 2025-12-13
The medium Market-Bench strategy asks a model to implement a COKE/PEP pairs mean-reversion backtester with shared capital, synthetic order books, VWAP tracking, order latency, and immediate-or-cancel execution. This definition reports mean absolute error averaged over five generated backtester runs.
Top models (lower is better)
| Model | Score |
|---|---|
| Gemini 3 Pro Preview | 52.2 |
| GPT-5.2 | 107 |
| DeepSeek-V3.2 | 133 |
| Nova Premier | 133 |
| Opus 4.5 | 134 |
| GPT-5.1-Codex-Max | 137 |
| Sonnet 4.5 | 193 |
| Mistral Large 3 675B Instruct 2512 | 229 |
| Grok 4 | 309 |
| Llama 4 Maverick | 605 |
| Command A | 1093 |
| Llama 3.1 Nemotron Ultra 253B V1 | 2268 |
| Qwen3-Max (2025-09-23) | 408991461 |