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

Forecasting French elections from imperfect polls: an overview

Abstract. The France section forecasts the 2027 presidential election and a possible National Assembly election with the rule used for Great Britain: take half the average past miss off today's polls, but only where that would have improved past forecasts made without knowing the result. Round-one presidential polls fail that test and are taken as published; round-two polls, which have overstated the RN candidate at every runoff the RN has reached, and legislative polls pass it and are corrected. This paper sets out the evidence, the seat model, the sources of uncertainty and what the forecast does not claim, and maps the other France papers.

Polling Forecast · updated 9 November 2026 · 13 min read · France, presidential election 2027, National Assembly, polling error, leave-one-out, Rassemblement national, overview

Summary

The France section forecasts two things: the presidential election of 18 April and 2 May 2027, and a National Assembly election that nobody has called but that could come at short notice. Both are built on the rule used for Great Britain. We measure how far the final polls missed at past elections, take half of the average miss off today's polls, and keep that correction only if it would have brought past forecasts closer when each election is predicted without knowing its own result.

In France the rule gives three different answers. Presidential round-one polls have missed by about a point per candidate family, but not in a steady direction; correcting them would have been further out at 4 of 5 elections, so they are taken as published. Round-two polls have overstated the Rassemblement national (RN) candidate at every runoff the RN has reached; correcting them was closer at 4 of 5, so we correct them, by −1.8 points on the RN candidate's share. Legislative round-one polls overstated the RN and its allies at each of the last three Assembly elections; correcting them was closer at 6 of 6, so the Assembly projection uses the correction.

Seats in the Assembly then come from a model of each of the 577 circonscriptions: uniform swing from 2024, the legal thresholds for round two, withdrawals, and a model of two-candidate duels fitted on 2024. That last step is the weakest part of the forecast. This paper sets out the whole chain in brief and points to the paper that treats each part in full.

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.

What is being forecast

France elects its president directly, for five years, in two rounds a fortnight apart. If no candidate wins over half the valid votes in round one, the top two meet in round two. Emmanuel Macron cannot stand in 2027. So there are two questions, and they can have different answers: who reaches round two, and who wins it. A candidate can lead every head-to-head poll and still fail to qualify.

The National Assembly has 577 deputies, one per circonscription: 539 in metropolitan France, 27 overseas and 11 for French people living abroad. A majority is 289. Each seat is also fought over two rounds. A candidate with over half the votes and at least a quarter of registered voters wins in round one. Otherwise the top two go through, together with anyone who has the votes of at least 12.5% of registered voters, so turnout decides how many candidates reach round two. Qualified candidates may withdraw, and in 2024 many did, mostly to stop the RN. No Assembly election is due, but the President can dissolve the Assembly, as Emmanuel Macron did in June 2024, and a new president may well do so after May 2027.

The site therefore shows three forecasts: the presidential election if it were held today, the same election on polling day, and the Assembly if it were elected now. The first is on the presidential page, the second on the election-day page and the third on the Assembly page. Figures in this paper are computed from the database each time the page is built; the current figures use polls to 19 November 2026.

Why French polls need their own treatment

The British method cannot be carried across unchanged, for five reasons.

Two rounds, two kinds of poll. Round-one polls measure a crowded field; round-two polls measure hypothetical head-to-heads between pairs of candidates who may never meet. Their errors behave differently (see Measuring French polling error, 1965–2024), so they are measured, tested and corrected separately.

The candidates are not settled. Candidacies close in March 2027. Pollsters meanwhile test several line-ups in each survey: with Marine Le Pen or Jordan Bardella for the RN, with one or another candidate for the centre, the Socialists and the right. A forecast has to choose which line-up to follow, and it has to say what happens if the choice is wrong (The 2027 presidential forecast: line-ups, runoffs and simulation).

Parties change names and candidates change parties. Errors cannot be measured by party label across twenty years in which the Front national became the Rassemblement national, the UMP became Les Républicains and a new centre formed around Emmanuel Macron. We measure by candidate family, so that Jean-Marie Le Pen in 2002 and Marine Le Pen in 2022 line up as one series. The families are ours, and Measuring French polling error, 1965–2024 explains how they are drawn.

No raw answers to work from. In Britain our projection starts from the weighted answers of everyone asked, before the pollster's turnout and don't-know adjustments (Why we start from British pollsters' raw answers rather than the published headline). French polls must file a notice with the Commission des sondages under the law of 19 July 1977, and the notices describe the weighting, but only OpinionWay and Cluster17 print raw vote shares, and on the RN they pull in opposite directions. There is almost no raw history to test against. We show the raw figures on the polls page and do not build on them (How we read French polls: what we correct, what we don't, and why).

Seats come from duels, not from pluralities. A British seat goes to whoever is first; a French seat goes to whoever wins round two, after qualification and withdrawals. The same national vote can give very different seat totals depending on who a bloc meets in round two and where the voters of eliminated candidates go (From votes to seats in the National Assembly).

The evidence in one place

The historical record holds each pollster's final poll before each round:

  • 40 final round-one polls and 37 final round-two polls at the five presidential elections from 2002 to 2022;
  • 22 final legislative round-one polls at the Assembly elections of 2017, 2022 and 2024, and the seat ranges pollsters published between the rounds;
  • 198 poll averages at set distances before polling day, for the election-day test;
  • the final polls of every presidential election back to 1965, for the long view.

Results come from the ministère de l'Intérieur through data.gouv.fr, under the Licence Ouverte. Poll figures come from each pollster's release or its notice at the Commission des sondages, or, where no primary copy could be found, from Wikipedia's poll tables, and each is graded by source on the site. The notices before 2016 are not online, so the polls of 2002–2012 all rest on secondary sources. The French data: results, maps, statistics, deputies and polls covers the sources and checks.

The test that decides each verdict is leave-one-out cross-validation (Stone, 1974). Each past election is predicted with the correction worked out from the other elections only, so no election helps to predict itself. Here are the three tests side by side.

ElectionPolls' average missCorrected, leave one outVerdict
20021.021.07further out
20071.101.15further out
20120.911.21further out
20170.600.57closer
20221.261.31further out
Average0.981.06
Live Correcting round-one polls, each election left out in turn Computed from the database, polls to 19 November 2026.
ElectionPolls' average missCorrected, leave one outVerdict
2002 round two3.211.62closer
2007 round two0.731.22further out
2012 round two1.110.78closer
2017 round two4.483.45closer
2022 round two3.322.24closer
Average2.571.86
Live Correcting round-two polls, each election left out in turn Computed from the database, polls to 19 November 2026.
ElectionPolls' average missCorrected, leave one outVerdict
20021.180.99closer
20070.680.40closer
20120.700.62closer
20171.571.35closer
20221.141.04closer
20240.930.73closer
Average1.030.86
Live Correcting legislative polls, each election left out in turn Computed from the database, polls to 19 November 2026.

The round-one misses are real but they change direction: the far right was understated by 3.9 points in 2002 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. A correction built on the average of misses like these adds noise. Round two is different: the RN's candidate was overstated by 3.2 points in 2002, 4.5 in 2017 and 3.3 in 2022. The two runoffs without an RN candidate, 2007 and 2012, missed by about a point, in opposite directions.

The rule, and what it corrects

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

\[ d_{p,e} = \bar{s}_{p,e} - r_{p,e}, \tag{1} \]

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

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

and the shares are rescaled to keep their total. A party needs at least two measured elections before it is corrected. The correction is applied to a contest only if, averaged over the leave-one-out test, it is closer than the polls by more than 0.01 points per party.

The half is the British choice, explained in Correcting the raw answers: the method behind our UK projection: it hedges between "past misses repeat" and "past misses were noise", in the spirit of shrinkage estimators (Efron and Morris, 1975). We have not tuned it for France. With five presidential elections and three legislative ones there is too little evidence to choose a factor, and choosing one on the same data used to test it would flatter the result.

In practice:

Contest Verdict What changes
Presidential round one not corrected nothing; the family's typical miss sets the spread of the simulation
Presidential round two corrected about 1.8 points off the RN candidate in each head-to-head, adjusted where the opponent's family has its own history
Legislative round one corrected the RN and allies down about 1.3 points, the non-aligned right up about 1.1, smaller changes elsewhere

Today's forecast

President

For round one we average the latest poll from each pollster within 45 days of the newest poll of our main line-up, which has Marine Le Pen for the RN, Édouard Philippe for the presidential camp, Jean-Luc Mélenchon for La France insoumise and Raphaël Glucksmann for the Socialists. Round two uses every head-to-head poll of each pair that could meet, corrected as above. Chances come from 20,000 simulated elections.

CandidateFamilyPollsOur betReaches round twoWins
Marine Le PenNational Rally32.832.8over 99%99%
Jean-Luc MélenchonFrance Unbowed18.318.395%under 1%
Édouard PhilippeEnsemble (Macron camp)12.212.23%1%
Raphaël GlucksmannSocialists11.611.62%under 1%
Bruno RetailleauRepublicans6.06.0under 1%0%
Éric ZemmourReconquête3.53.50%0%
Marine TondelierGreens2.62.60%0%
Fabien RousselCommunist Party2.12.10%0%
Nicolas Dupont-AignanSovereigntists2.12.10%0%
Nathalie ArthaudFar left0.80.80%0%
Live Round one today: the main line-up Computed from the database, polls to 19 November 2026.
Match-upPollsOur betChance it happensFirst-named wins
Le Pen – Mélenchon68.166.295%over 99%
Le Pen – Philippe55.053.13%76%
Le Pen – Glucksmann57.056.12%92%
Le Pen – Retailleau57.055.8under 1%91%
Mélenchon – Philippe34.032.10%under 1%
Live Round-two match-ups that have been polled Computed from the database, polls to 19 November 2026.

On this reading Marine Le Pen is the favourite, with a 99% chance of winning if the election were held today. The full method, including how we choose among the line-ups pollsters test and the variant with Jordan Bardella, is in The 2027 presidential forecast: line-ups, runoffs and simulation.

Assembly

Pollsters have not asked about a legislative election since October 2025, so the projection uses the last legislative polls, corrected.

Bloc2024PollsOur betSeats80% rangeMajority
Left (NFP and other left)29.733.033.6291204–31022%
Ensemble and centre22.015.015.27262–1420%
LR and other right11.015.016.29082–1120%
RN and allies33.231.831.110869–1470%
Reconquête0.82.82.300–00%
Far left1.11.00.900–00%
Others and regionalists2.24.33.71614–170%
Live National Assembly: round one by bloc and projected seats Computed from the database, polls to 19 November 2026.

No bloc has a majority in 78% of the simulated elections. From votes to seats in the National Assembly explains how the seats are reached and how uncertain the answer is.

The seat model in brief

Each circonscription starts from its 2024 round-one result, grouped into seven blocs, and moves by the national change in each bloc since 2024: uniform swing, the oldest and simplest projection (Butler and Stokes, 1969). The legal thresholds then decide who wins outright and who qualifies. Where the RN and two or more of its opponents qualify, the weaker opponents withdraw, as most did in 2024. Each remaining duel is decided by a linear model of the round-two share fitted on the 403 duels of 2024, which uses every bloc's round-one share and so learns how the voters of eliminated blocs split.

Given the actual 2024 round one, the duel model picks 370 of those 403 winners. That is an in-sample figure, and the out-of-sample record is much weaker: fitted on 2022, the same model picks 64.5% of the 2024 duel winners, barely better than naming the round-one leader (61.0%); fitted on 2024, it picks 74.8% of the 2022 duels, worse than the round-one leader (78.0%). Voters' behaviour in round two changed a great deal between those two elections. The projection assumes it works as it did in 2024.

Pollsters' own seat projections, published between the rounds, have done no better. At each of the last three Assembly elections the bloc that led round one won fewer seats than every pollster's range. The misses came from round two, which is exactly the step that is hardest to model.

The uncertainty

For the president, each simulated round one moves every candidate by a random error with the standard deviation of that family's final-poll miss (at least 1.5 points); the top two meet; the winner is drawn from their corrected head-to-head plus a random round-two error of about 4.4 points, the largest typical round-two miss with an allowance for the time still to run. A pair that no pollster has tested counts as a coin toss.

For the Assembly, 1,000 simulated elections move each bloc's national share at random, add local noise in every seat, and make each possible withdrawal happen at the 2024 rate. The 80% ranges are read off those runs. They also sit beside a set of "lenses": today's legislative polls replayed with the miss of 2017, 2022 or 2024, every seat worked out again. 91 seats change hands under at least one of those past misses.

For polling day, the question is how far polls move in the months before an election. In Britain the older parties drift back towards their previous result, and the election-day model pulls them back. We tested the same idea on French presidential polls since 2002 and it did not help, so the polling-day forecast keeps today's polls and adds the drift that polls this far out have shown: about 4.9 points per candidate at 148 days. From today to polling day: drift in French presidential polls gives the evidence.

What we do not claim

We do not claim to know who will stand. The forecast follows the line-ups pollsters test. A withdrawal, a new candidate or a court ruling can change the field, and the forecast will not see it coming.

We do not claim that the round-two correction rests on much. It comes from three runoffs with an RN candidate. The pattern is consistent, and the literature on the 2002 miss (Durand, Blais and Larochelle, 2004) shows that measuring far-right support has long been a problem for French polls, although in 2002's round one the error ran the other way and Jean-Marie Le Pen was understated. But three cases are three cases, and the correction is halved to reflect that.

We do not claim that round-one polls are accurate. They are not corrected because their errors have no stable direction, not because the errors are small. Across many countries and decades, the average size of polling errors has been fairly stable (Jennings and Wlezien, 2018); a stable size does not imply a stable direction. The round-one uncertainty in the simulation is set by the size of past misses for exactly that reason.

We do not claim that the Assembly projection is current. It rests on polls taken in autumn 2025 and on a model of round-two behaviour fitted to one election. Uniform swing ignores local candidates and incumbency, which matter a great deal in French legislative elections.

We do not claim that the overseas and abroad seats are well modelled. They are projected like the rest, but their politics and their swings differ from those of metropolitan France.

A guide to the France papers

Forecasting French elections from imperfect polls: an overview is this paper.

Measuring French polling error, 1965–2024 sets up the evidence: how a final poll is defined, why errors are measured by candidate family, the round-one and round-two record from 2002 to 2022, the long view back to 1965, the legislative record by bloc from 2017 to 2024, the pollsters' seat projections, and the error in the gap between the top two.

How we read French polls: what we correct, what we don't, and why is the short statement of the method: the rule, the three verdicts and why raw figures are not used.

The 2027 presidential forecast: line-ups, runoffs and simulation covers the presidential forecast: choosing the line-up, averaging the polls, the round-two pairs and their correction, the 20,000-run simulation, and the Bardella variant.

From votes to seats in the National Assembly covers the seat model: uniform swing from 2024, qualification, withdrawals, the duel model and its tests, the overseas and abroad seats, and what the 80% ranges mean.

From today to polling day: drift in French presidential polls covers the gap between today and polling day: the evidence at fixed distances, the pull-back curve that failed its test, and the drift that is added instead.

If we're wrong: replaying past polling misses on today's French polls shows the same polls read in different ways, each run through the same seat model.

Rating French pollsters: accuracy, the lead and openness rates French pollsters on their final-poll record, their miss on the gap between the top two, and how much of their working they publish.

The French data: results, maps, statistics, deputies and polls documents the sources, licences and checks.

What we don't know yet: limitations of the French forecast collects the known weaknesses in one place.

References

Butler, D. and Stokes, D. (1969). Political Change in Britain. London: Macmillan.

Durand, C., Blais, A. and Larochelle, M. (2004). The polls in the 2002 French presidential election: an autopsy. Public Opinion Quarterly, 68(4), 602–622.

Efron, B. and Morris, C. (1975). Data analysis using Stein's estimator and its generalizations. Journal of the American Statistical Association, 70(350), 311–319.

Jennings, W. and Wlezien, C. (2018). Election polling errors across time and space. Nature Human Behaviour, 2(4), 276–283.

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

Stone, M. (1974). Cross-validatory choice and assessment of statistical predictions. Journal of the Royal Statistical Society: Series B, 36(2), 111–147.

How to cite

Polling Forecast (2026). Forecasting French elections from imperfect polls: an overview. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/france-overview. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like