Would our forecasts have been right?
A forecaster earns trust by being right in public. We are new, so we do two things. Before every election we lock our final forecast here and never change it. And for every past election our models learn from, we ask what our method would have said from the final polls, without knowing the result, and score it.
Final forecasts, locked
| Locked | Election | Our final forecast | |
|---|---|---|---|
| 2 Nov 2026 | United States: Midterms: House and Senate | History → |
Next to be locked:
- Spain, General election: locked on 28 November 2026
- France, Presidential election: locked on 17 April 2027
- United States, Presidential election: locked on 6 November 2028
- Canada, General election: locked on 14 October 2029
Locked means stored as it stood on the eve of polling day, from the same numbers the site showed that day, and never edited afterwards. When the result is in, it is scored here.
Every past election, predicted blind
58 elections since 2000 in eight countries (France, Germany, Spain, Italy, the Netherlands, Poland, Canada and Australia), 377 party results. Each election is predicted from its own final polls, corrected only with what the other elections taught, with each party's range set by how far its polls missed at the other elections. Then we check.
When we said X%, did it happen X% of the time?
| We gave | Parties | Average | Came first |
|---|---|---|---|
| 0–5% | 286 | 0% | 1% |
| 5–20% | 19 | 12% | 16% |
| 20–40% | 8 | 33% | 25% |
| 40–60% | 15 | 49% | 47% |
| 60–80% | 6 | 68% | 100% |
| 80–95% | 13 | 89% | 77% |
| 95–100% | 30 | 99% | 90% |
Dots on the dashed line mean the chances meant what they said. Above the line: we were too cautious; below: too confident. Dot size: how many parties.
Country by country
| Country | Elections | In 50% range | In 80% | In 95% | Favourite first | Brier, ours | Brier, polls | Correction |
|---|---|---|---|---|---|---|---|---|
| France | 6 | 69% | 92% | 98% | 5 of 6 | 0.360 | 0.396 | on |
| Germany | 7 | 51% | 80% | 90% | 7 of 7 | 0.032 | 0.032 | off |
| Spain | 8 | 77% | 85% | 92% | 7 of 8 | 0.227 | 0.242 | on |
| Italy | 6 | 48% | 70% | 89% | 5 of 6 | 0.412 | 0.393 | on |
| Netherlands | 8 | 71% | 90% | 95% | 6 of 8 | 0.556 | 0.602 | on |
| Poland | 7 | 58% | 77% | 81% | 6 of 7 | 0.264 | 0.264 | off |
| Canada | 9 | 62% | 89% | 98% | 7 of 9 | 0.166 | 0.148 | on |
| Australia | 7 | 52% | 67% | 96% | 6 of 7 | 0.340 | 0.341 | off |
What this does and doesn't show
- Vote shares, not seats: every seat model is tested separately on the country's own track-record and method pages, against the official results.
- Parties' misses are drawn independently here; in a real election they move together, which our seat simulations do keep. That makes these ranges a fair test of size, not of how parties move together.
- Few elections per country (four to nine since 2000), so the country rows are noisy; the pooled figures are the test.
- The United Kingdom and the United States have their own: the UK backtest and the US races called right.