10.6%
lower Brier score than GPT-5.4
2×
better calibration than GPT-5.4
8.4×
larger Brier Skill Score lift than GPT-5.4
We trained a compact forecaster for Numinous-style military-strikes questions using public news, resolved outcomes, and Foresight Learning. On a held-out set of military-strikes forecasts, it beat GPT-5.4 on Brier score, calibration, and Brier Skill Score.
What we did
- Generated military-strikes forecasting examples from five public-news search queries.
- Resolved each question from later reporting, with no annotators or proprietary feeds.
- Tuned from GPT-OSS-120B with Foresight Learning and benchmarked against GPT-5.4 and the base model.
Read more
Primary artifacts for this case study.