France Presidential election, round one18 Apr 2027in 148 days Le Pen wins 99% Philippe wins 1% Calendar Track record Research
Research · France
EN FR

How we read French polls: what we correct, what we don't, and why

Abstract. French presidential round-one polls have no stable lean and are taken as they are; round-two polls have overstated the RN candidate at every runoff the RN has reached, and legislative polls have overstated the RN at each of the last three Assembly elections, so both are corrected. Assembly seats come from a circonscription model of qualification, withdrawal and duels fitted on 2024.

Polling Forecast · updated 9 November 2026 · 8 min read · France, presidential election, National Assembly, two-round system, RN, polling error, runoff, Commission des sondages

Summary

This paper sets out how Polling Forecast reads French opinion polls for the 2027 presidential election and for a National Assembly election, and why the answer differs between contests. We apply the rule used for Great Britain: measure how far each party's final-poll average missed at past elections, subtract half of the average miss from today's polls, and keep that correction only if it would have improved every past forecast on average when each election is predicted without knowing its result. In France the rule gives three different verdicts. Presidential round-one polls have no stable lean, and correcting them would have made the forecast worse; we take them as they are. Presidential round-two polls have overstated the Rassemblement national (RN) candidate at every runoff the RN has reached, and correcting them helps; we correct them. Legislative round-one polls have overstated the RN at each of the last three Assembly elections, and correcting them helps; we correct them. Seats in the Assembly then come from a circonscription-level model of qualification, withdrawal and two-candidate duels, fitted on the 2024 election.

What has changed since this paper was written

Updated 9 November 2026. The rest of the paper is as first published, except where a sentence below says it was corrected.

  • Only elections from 2000 shape the forecast (28 September). Earlier elections stay in the long view and on the history pages but no longer feed any test, correction or error size. In France the presidential tests now run on the five elections from 2002 to 2022: the round-one correction stays off and the round-two correction stays on.
  • Our figures beside the polls (26 October). Where our projection differs from the poll average, both are shown. The presidential first-round correction is off, so there our shares are the polls.

What is being forecast

France elects its president in two rounds, a fortnight apart. If no candidate has an absolute majority of valid votes in round one, the two leading candidates meet in round two. The next election is set for 18 April and 2 May 2027. The National Assembly has 577 members, each elected in a single-member circonscription, also in two rounds. A candidate wins in round one with over half the valid votes and at least a quarter of registered voters. Otherwise the top two go through, together with any other candidate who received votes from at least 12.5% of registered voters, so turnout decides how many candidates reach round two. A candidate who qualifies may withdraw, and in 2024 many did: 225 qualified candidates withdrew, and of the 312 seats where three or more qualified only 91 still had three candidates in round two. No Assembly election is due, but the President may dissolve the Assembly, and a new president may do so after May 2027.

The data

Results come from the ministère de l'Intérieur through data.gouv.fr under the Licence Ouverte: national presidential results for 2002–2022, and results by circonscription for the legislative elections of 2022 and 2024 and the 2022 presidential election. Poll figures come from each pollster's release or from its notice at the Commission des sondages. Under the law of 19 July 1977, every published electoral poll must file a notice, and the notice must set out how the sample was weighted. Where no primary copy could be found, figures come from Wikipedia's poll tables and are marked as such on the site.

The historical record holds each pollster's final poll before each round: 40 round-one polls and 37 round-two polls across the five presidential elections from 2002 to 2022, and 22 legislative round-one polls across 2017, 2022 and 2024. Where a pollster ran two series with the same final fieldwork date (Ifop in 2012, Ipsos in 2022), only one counts.

Candidates change between elections, so errors are measured by candidate family: Jean-Marie Le Pen, Marine Le Pen and the RN form one family, Nicolas Sarkozy, François Fillon and Valérie Pécresse another, and so on. Legislative results are grouped into blocs: the united left (NUPES in 2022, the Nouveau Front populaire in 2024), the presidential camp, Les Républicains and allies, the RN and its allies, and smaller groups. The groupings are ours; the ministry's own candidate labels (nuances) are kept on each seat page.

The rule

For party or family \(p\) at past election \(e\), let \(\bar{s}_{p,e}\) be the average of the final polls and \(r_{p,e}\) the result. The error is

\[ d_{p,e} = \bar{s}_{p,e} - r_{p,e}, \]

positive when the polls overstated the party. Over the set \(E_p\) of past elections at which the party was measured, the mean error is \(\bar{d}_p\), and today's average \(s_p\) becomes

\[ \hat{s}_p = s_p - \lambda\,\bar{d}_p, \qquad \lambda = \tfrac{1}{2}, \]

rescaled so the shares keep their total. A party needs at least two measured elections before it is corrected. The choice of one half is explained for Britain in Correcting the raw answers: the method behind our UK projection: it hedges between "past misses repeat" and "past misses were noise".

The test is leave-one-out. Each past election is predicted with the correction worked out from the other elections only, and its mean absolute error across parties is compared with that of the uncorrected polls. The correction is kept for a contest only if, averaged over the elections, it is closer than the polls.

Presidential round one: no stable lean

The final-poll averages have missed by about a point per family, with some large single misses. The far right was understated by 3.9 points in 2002, when Jean-Marie Le Pen reached round two against every poll's expectation, and by 2.0 in 2012, but overstated by 3.3 in 2007. The mainstream right was understated by 2.8 in 2007 and overstated by 3.1 in 2022. Jean-Luc Mélenchon was overstated by 2.5 in 2012 and understated by 4.8 in 2022.

Election Polls' average miss Corrected (leave one out)
2002 1.02 1.07
2007 1.10 1.15
2012 0.91 1.21
2017 0.60 0.57
2022 1.26 1.31

Correcting was further out at four of the five elections, so the round-one projection uses the polls as they are. The round-one forecast still allows for error: each family's typical miss sets the spread of the simulation described below.

Presidential round two: the RN candidate overstated

The RN's candidate reached round two in 2002, 2017 and 2022. Each time the final polls put that candidate too high: by 3.2, 4.5 and 3.3 points of the two-candidate vote. In the two runoffs without an RN candidate (2007 and 2012) the misses were about a point and went in opposite directions.

Election Polls' average miss Corrected (leave one out)
2002 3.21 1.62
2007 0.73 1.22
2012 1.11 0.78
2017 4.48 3.45
2022 3.32 2.24

Correcting was closer at four of the five runoffs and worse only in 2007, the one with no RN candidate and the smallest miss. The projection therefore takes about 1.8 points off the RN candidate's share in every head-to-head poll and gives it to the opponent (averaged, where the opponent's family also has round-two history, with that family's own correction). With three cases, the evidence is thin, and the correction is halved to reflect that.

Legislative polls and seat projections

At the three most recent Assembly elections, final round-one polls overstated the RN and its allies each time (by 4.3, 0.5 and 2.9 points) and understated the non-aligned right each time. Correcting was closer at all three elections (1.57 to 1.23 in 2017, 1.14 to 1.13 in 2022, 0.93 to 0.55 in 2024), so the Assembly projection uses the correction.

Pollsters also publish seat ranges between the two rounds. These have missed by far more than the vote shares. The bloc that led round one won fewer seats than every pollster's range at each election: the presidential camp in 2017 (projected about 437–467, won 350), the NUPES left in 2022 (about 161–206, won 131) and the RN in 2024 (about 188–221, won 142). The misses come from round two: who withdraws, and how the voters of eliminated candidates split.

From polls to chances

Presidential. French polls test several candidate line-ups in each survey. For each poll we take the scenario closest to a reference line-up, with Marine Le Pen for the RN, Édouard Philippe for the presidential camp and Raphaël Glucksmann for the Socialists, and separately the scenario with Jordan Bardella instead of Le Pen. Each pollster's latest poll within 45 days of the newest counts once. The simulation draws 20,000 first rounds, moving each candidate by a random error with the standard deviation of that family's final-poll error (at least 1.5 points). The top two meet in round two. The winner is drawn from their corrected head-to-head average plus a random error of about 4.5 points: the largest typical round-two miss, with an allowance for movement. Pairs never polled head to head count as even. These are chances for an election held now; seven months out, polls move more than their final-week error.

Assembly. Each circonscription starts from its 2024 round-one result by bloc, moved by the national change in each bloc (uniform swing). Qualification follows the law. Where the RN and two or more of its opponents qualify, the weaker opponents withdraw, as most did in 2024; in the simulation each withdraws with the 2024 rate. Each duel is decided by a linear model of the round-two share fitted on the 403 duels of 2024:

\[ s^{(2)}_{A} - \tfrac{1}{2} = \beta_0 + \beta_1\,(s_A - s_B) + \sum_{k \ne A,B} \gamma_k\, s_k , \]

where \(s\) are round-one shares and \(k\) runs over the other blocs, so the model learns how each bloc's voters divide between the finalists. Given the actual 2024 round one, it picks 370 of the 403 winners. Its weakness is plain in the data. Rates fitted on 2022 pick 64.5% of the 2024 duel winners (the round-one leader alone picks 61.0%), and rates fitted on 2024 pick 74.8% of 2022's (the leader, 78.0%). Voters' round-two behaviour changed a great deal between the two elections, and the projection assumes it works as it did in 2024. A thousand simulations add national and local random errors.

What we don't use yet

In Britain the projection starts from raw answers, before the pollsters' turnout and don't-know adjustments (Why we start from British pollsters' raw answers rather than the published headline). French notices make that possible in principle, but only OpinionWay and Cluster17 print raw vote shares, the two pull in opposite directions on the RN, and there is almost no raw history to test against. They are shown on the polls page, not used.

Limitations

  • Three or five past elections per contest is little evidence; one new election can overturn a pattern.
  • The round-two correction rests on three RN runoffs.
  • The Assembly projection uses the last legislative voting-intention polls, taken in autumn 2025.
  • Uniform swing ignores local candidates and incumbency.
  • The 38 overseas and abroad circonscriptions (27 overseas, 11 for French people abroad) are projected like the rest, and their results swing differently.

References

Jennings, W. and Wlezien, C. (2018) "Election polling errors across time and space", Nature Human Behaviour, 2, 276–283.

Loi n° 77-808 du 19 juillet 1977 relative à la publication et à la diffusion de certains sondages d'opinion.

How to cite

Polling Forecast (2026). How we read French polls: what we correct, what we don't, and why. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/france-method. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like