Every forecast is scored, including the misses
Every week, the forecast is checked against what happened next, by the same rules as its benchmarks. After the election, it is scored against the results. Published scores are never rewritten; a correction is added as a new, dated line.
Scores
The first scores are published on Monday, October 19. Weekly scoring dates: Oct 19, Oct 26, Nov 2; then the results.
What we compare
| Social simulation | The headline forecast: statistics plus voter personas, checked against the evidence every Monday. |
|---|---|
| Statistics only | The same chain with the simulation switched off: the number the simulation has to beat. |
| Prediction markets | Kalshi and Polymarket prices, averaged. |
| Cook Political Report | Cook's ratings, turned into chances. |
How scoring works
Before the electionthere are no results, so each week’s forecast is checked against the polls published in the following week: the error is the average gap in margin points, and moves called is the share of races where the forecast moved the same way the polls later did.
After the election, every forecast is scored against the results. The Brier score is the average squared gap between a forecast chance and what happened: 0 is perfect, and a coin flip on every race scores 0.25. Log loss punishes confident misses harder. Lower is better for both.