Updated
Updated · arxiv.org · Aug 28
How Proper Scoring Rules Shape LLM Forecasting
Updated
Updated · arxiv.org · Aug 28

How Proper Scoring Rules Shape LLM Forecasting

1 articles · Updated · arxiv.org · Aug 28

Summary

  • A new study has compared five proper scoring rules as training objectives for large language model (LLM) forecasters on real-world binary events.
  • The research found that reward choice affects not just overall forecasting accuracy, but also calibration, discrimination, and the structure of forecasting errors.
  • These findings suggest that different scoring rules can shape LLM forecasting profiles in distinct ways, with implications for model selection and ensemble design.