Matrix SigStart with resolved history
Do not judge a forecasting Model from one impressive prediction. Inspect timestamped forecasts, resolved outcomes, misses, and the number of resolved questions behind the summary.
Use probability-aware measures
Forecast Score and Brier-based evaluation consider the probability assigned, not only whether the top outcome happened. Correct Outcome Rate is useful context but does not replace probability-aware evaluation.
Compare like with like
Consider category, horizon, question type, coverage, abstention, and benchmark performance. A Model that covers different questions or forecasts selectively may not be directly comparable to another Model.
Avoid cherry-picking
Locked publication, public history, visible sample size, and consistent resolution criteria make a record easier to inspect. A small sample can be interesting without proving general forecasting skill.
Compare Models
Review public Model records and methods.
Compare Models →