Forecast-Dojo (Research Tools + Memory): leaderboard
Forecast-Dojo (2026) replays 230 resolved Polymarket events as dated forecast steps; here the research agent also carries its own belief notebook from one forecast date to the next. Multiclass Brier score on a 0 to 2 scale over four rollouts, with unusable forecasts counted as uniform; lower is better.
Metric: Multiclass Brier score (0-2 scale, lower is better; mean over 797 scheduled forecast dates of 230 held-out Polymarket events resolved March-June 2026, four rollouts each; unusable forecasts scored as a uniform distribution; the same research tools plus the agent's own belief notebook carried from the previous forecast date). Source: arxiv.org. Saturation forecast: Around 2029. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Max) | 0.55 | #29 (GPT-5.6 Sol) |
| 2 | GPT-5.5 (xHigh) | 0.57 | #48 (GPT-5.5) |
| 3 | GPT-5.4 (xHigh) | 0.58 | #121 (GPT-5.4) |
| 4 | Claude Opus 4.6 (Max) | 0.6 | #100 (Claude Opus 4.6) |
| 5 | Claude Opus 4.8 (Max) | 0.61 | #56 (Claude Opus 4.8) |
| 6 | MiniMax-M2.5 | 0.64 | #590 |
| 7 | Qwen 3.5 397B A17B | 0.64 | #242 |
| 8 | Kimi K2.5 (High) | 0.66 | #252 (Kimi K2.5) |
| 9 | DeepSeek V3.2 (High) | 0.67 | #403 (DeepSeek V3.2) |
| 10 | GPT-OSS-120B (High) | 0.7 | #699 (GPT-OSS-120B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=forecast-dojo-research-tools-plus-memory · How It Works · Data refreshed daily, snapshot 2026-11-12.