Validation: how do these methods fare on past elections?
Experimental
Experimental. This page is experimental. It replays three forecasting methods on five past presidential elections (2002 to 2022), giving them only the polls published before the cut-off date, then compares their first-round forecasts with the official result. It shows no figure from the current campaign.
The three methods side by side
All first-round forecasts, all horizons together (180 to 7 days before the vote): 275 candidate-forecast pairs over 25 election-horizon combinations. Errors in percentage points; the lower, the closer the forecast to the result.
Experimental
| Method | MAE (pts) | RMSE (pts) | 80% coverage | 80% width (pts) | CRPS | Brier (comes first) | Brier (top 2) |
|---|---|---|---|---|---|---|---|
| Simple poll average | 2.08 | 3.19 | 83 % | 6.64 | 1.53 | 0.029 | 0.047 |
| Model v2 | 2.20 | 3.31 | 85 % | 10.24 | 1.76 | 0.037 | 0.056 |
| Opinion layer | 2.35 | 3.69 | 35 % | 2.59 | 1.92 | 0.023 | 0.046 |
How to read this table. One row per method. The error (MAE, RMSE) measures the gap between the forecast and actual share; coverage says what share of results fell inside the 80% interval; CRPS and Brier scores judge the whole probabilistic forecast.
Learn more
For coverage, the right figure is not "the highest" but "close to 80%": a very wide interval always covers and says almost nothing. In every other column, lower means closer to the result. Small differences are not significant with only five elections.
Mean error by horizon
MAE in percentage points, from 180 days (left) to 7 days (right) before the first round.
Experimental
- Simple poll average
- Model v2
- Opinion layer
How to read this chart. Each line is a method; the lower the line, the smaller the mean error. The dots are the horizons tested, from 180 days before the vote on the left to 7 days on the right.
Learn more
Far from the vote there are few polls and the field of candidates is uncertain, so the error is naturally larger there. Lines that cross mark methods this sample cannot tell apart.
Do the 80% intervals keep their word?
Share of actual results that fell inside the announced 80% interval, by horizon. The thin horizontal line marks the expected value: 80%.
Experimental
- Simple poll average
- Model v2
- Opinion layer
How to read this chart. A method whose 80% intervals are well set should sit on the thin 80% line. Lower: intervals too narrow, the method is overconfident. Higher: intervals wider than needed.
Learn more
Each point rests on few elections, so it moves a lot from one horizon to the next without meaning much. Only a lasting gap from the 80% line is informative.
What to take from it
The main finding is simple: the plain average of recent polls does no worse than the more elaborate models. Over these five elections its mean error is no larger than that of model v2 or the opinion layer. A sophisticated calculation has not, here, shown that it adds anything.
The opinion layer's intervals are too narrow: the announced 80% interval held the result in only a minority of cases. The method thinks it is more precise than it is. The other two methods cover roughly what they announce, but with clearly wider intervals.
Five elections is few: each has a single result, and the candidates of one election are not independent. A gap between two methods may be down to chance. These figures do not say how the methods will behave in 2027; they show how they behaved.
Assumptions and limits
- Publication dates of polls are not in the data: a 2-day delay between the end of fieldwork and publication is assumed, and a poll is used only if published before the cut-off date.
- Polls and results for 2002 to 2012 are read from Wikipedia pages and have not been cross-checked against the Conseil constitutionnel figures.
- Far from the vote the list of candidates is partly hypothetical: only candidates forecast by all three methods and who actually ran are scored.
Run of 2 Nov 2026 · data fingerprint (SHA-256): dcc768516ed0d9b6ff3bbb4b865afcf0bb27e7fe8912600a5c5ac27dc0147f8c