Elections · resolves 3 Nov 2026
Will Democrats hold most governorships after 2026?
55%chance, our model
Every forecast we publish is frozen before the result and scored after it, here, whoever's model made it.
Data as of 13 Oct 2026 · 844 forecasts tracked · 351 resolved · 492 open · Source: this site's forecast ledger
Hit rate and log loss compare across domains; each Brier stays on its own scale.
Hit rate: the share of probability questions where the answer we rated most likely happened (a 50/50 call counts as no pick). Log loss: the average of minus the natural log of the probability we gave what happened; a coin scores 0.693, lower is better.
Brier on a yes/no question runs 0 to 1 (a coin scores 0.25); on a question with several answers it sums over every answer and runs 0 to 2. The two never share a column. Ranges held: the share of published ranges that contained the result, against the level the range was published at.
Up the side: how often it happened. When we say 30%, it should happen about 30% of the time.
Every probability a resolved forecast stated is one point: the yes on a yes/no question, and each answer on a question with several, scored against whether that answer happened. Points are pooled across every domain into ten-point bands; a dot sits at the band’s average stated probability and its observed frequency, and its size tracks how many points the band holds. 776 points so far.
Tap or drag across the chart: each dot is a ten-point band of what we said. The dashed line is perfect calibration.
Same questions, same frozen moment: positive skill means we beat the price.
Only questions where a real market priced the same claim count here. The market’s quoted probabilities are normalised to sum to one before scoring. The comparison is a reference line for judging the model, never betting advice. The sport-by-sport version, against closing prices, is the Ledger.
What we said when the forecast froze, and what happened.
Our forecasts, frozen before the result and scored after it.
Elections · resolves 3 Nov 2026
55%chance, our model
Elections · resolves 3 Nov 2026
88%chance, our model
Polymarket prices it at 91% (the tick). A reference line, not advice.
Elections · resolves 3 Nov 2026
60%chance, our model
Polymarket prices it at 64% (the tick). A reference line, not advice.
Elections · resolves 7 Nov 2028
58%chance, our model
8 more on the forecast page.
The forecasts come from this site’s own models, each publishing the same shape: the question, our probability or range, the moment it freezes and the rule that settles it. A daily job snapshots every open forecast, freezes it at its stated freeze time (never a version issued after it), records the result from the domain’s resolver, and scores it. The record is append-only.
Elections are graded by the election scorer against filed final counts; a provisional count resolves nothing. Games come settled from each sport’s own predictions ledger, scored to the same Brier those pages show. Market prices are Polymarket and the bookmakers’ posted or closing lines, shown as a reference.
Per-sport skill against the closing market, by season, is on the Ledger; each election race is graded on the election forecast.