We generated political forecasting questions from public news context and trained a compact forecaster from GPT-OSS-120B on the resolved outcomes. Trump-Forecaster beats GPT-5 on Brier score and calibration with news context, and is less overconfident when evaluated without context.
What we did
- Generated political forecasting questions from five public-news search areas.
- Filtered the data with a temporal split to avoid leakage.
- Tuned from GPT-OSS-120B with a Brier-score reward and evaluated with and without news context.
Read more
Primary artifacts for this case study.