What we don't know yet: limitations of the French forecast
Abstract. The French forecast rests on five presidential and three legislative elections and 99 final polls, enough to see a pattern in round-two polls but not to be sure it will hold. The line-ups are unsettled, the last legislative polls date from autumn 2025, the seat model moves every circonscription uniformly from 2024, and round two depends on withdrawals and transfers that changed sharply between 2022 and 2024. This paper sets out each weakness, including the 38 seats overseas and abroad, the scarcity of raw figures and the ban on publishing polls before the vote, and what evidence would change the method.
Summary
The French forecast rests on five presidential elections and three Assembly elections, and on 99 final polls in all. That is enough to see that round-two polls have overstated the Rassemblement national's candidate three times out of three and that round-one polls have no steady lean, but not enough to be sure either pattern will hold in 2027. The line-ups are not settled, the last legislative polls are a year old, the seat model moves every circonscription by the same amount from 2024, and round two in each seat depends on withdrawals and transfers that changed sharply between 2022 and 2024. The 38 seats overseas and abroad are projected like the rest although they vote very differently. Raw figures exist for only two pollsters, and the ban on publishing polls in the last days of a campaign means late movement is never measured. This paper sets out each of these weaknesses, what the evidence says about its size, and what would lead us to change the method. The British counterpart is What we don't know yet: limitations and open questions in our UK forecast.
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.
- Pollster grades rescaled (13 October) (Rating French pollsters: accuracy, the lead and openness).
A small number of elections
What the evidence is
Every correction on the site is decided by the same test: predict each past election with the correction worked out from the other elections only, and keep the correction if, on average, it is closer than the polls as published (How we read French polls: what we correct, what we don't, and why). In France that test runs on:
- Presidential round one: 40 final polls at five elections, 2002 to 2022.
- Presidential round two: 37 final polls at the same five elections.
- Legislative round one: 22 final polls at three elections, 2017, 2022 and 2024.
Five and three are very small samples for a test of this kind. A single new election can overturn a verdict, and the verdicts are closer than their headlines suggest.
How fragile each verdict is
Round two. The correction was closer at 4 of five runoffs, and the case for it rests on the three runoffs the RN reached: in 2002, 2017 and 2022 the final polls put its candidate too high by 3.2, 4.5 and 3.3 points of the two-candidate vote. The one runoff where correcting did worse was 2007, which had no RN candidate. So the correction of −1.8 points is, in substance, a bet that three cases make a pattern. Three misses in the same direction by similar amounts is more than chance would usually produce, but the causes need not have been the same each time, and if they were different there is no reason for the pattern to repeat.
| Election | Polls' average miss | Corrected, leave one out | Verdict |
|---|---|---|---|
| 2002 round two | 3.21 | 1.62 | closer |
| 2007 round two | 0.73 | 1.22 | further out |
| 2012 round two | 1.11 | 0.78 | closer |
| 2017 round two | 4.48 | 3.45 | closer |
| 2022 round two | 3.32 | 2.24 | closer |
| Average | 2.57 | 1.86 |
Legislative round one. Correcting was closer at all three elections, but in 2022 by a hundredth of a point (1.14 against 1.13). The case really rests on 2017 and 2024, the two elections where the RN was overstated by several points. A fourth election in which the RN was not overstated would shrink the correction and could tip the test.
Presidential round one. Correcting was further out at 4 of five elections, so we take round-one polls as they are. This verdict is the most solid of the three, because it rests on misses that changed direction: the far right was understated by 3.9 points in 2002 and overstated by 3.3 in 2007, and Mélenchon's family was overstated by 2.5 in 2012 and understated by 4.8 in 2022. But "no stable lean" is a statement about the past five elections, not a guarantee.
Whether the older record helps
The long view back to 1965 (Measuring French polling error, 1965–2024) adds six elections, but their polls are few (one pollster in 1965 and in 1981) and taken from secondary sources. It shows larger round-one misses in the 1960s than since, and round-two misses under a point at every runoff from 1965 to 1995, before the RN first reached round two. It says nothing about whether the RN overstatement will recur.
Candidacies are not settled
The election is on 18 April and 2 May 2027, and candidacies close in March 2027. Until then we forecast line-ups the pollsters test, not a known field.
We follow two. The main line-up has Marine Le Pen for the RN, Édouard Philippe for the presidential camp, Jean-Luc Mélenchon for France Insoumise and Raphaël Glucksmann for the Socialists; the second is the same with Jordan Bardella standing for the RN. For each poll we take the scenario closest to the line-up. That rule is mechanical and has costs. Gabriel Attal and Édouard Philippe are both listed as declared candidates from the presidential camp, and polls test each. A poll that tests Attal and not Philippe still counts towards the other candidates' averages, but Philippe's average then rests on fewer polls, and Attal's own share is not carried into the forecast at all. Where pollsters test a line-up rarely, as with the Bardella variant, its average rests on fewer and older polls.
Round two has a further gap. The chance of each pair meeting comes from the round-one simulation, but the winner of a pair comes from head-to-head polls, and pairs that no poll has tested are treated as a coin toss. Pollsters mostly test the RN candidate against each opponent, so the pairings without the RN are the least polled. If round-one polls tighten so that a runoff without the RN becomes likely, the forecast for that runoff will rest on nothing until pollsters test it.
And the evidence behind the chances assumes the candidates stay. The election-day page puts it plainly: a withdrawal or a new candidate "changes everything", and five campaigns, including the collapse of François Fillon's in 2017, do not cover every kind of shock.
Stale legislative polls
No Assembly election is due. The President may dissolve the Assembly, and a new president may do so after May 2027. Our Assembly projection uses the last legislative voting-intention polls, the five latest from different pollsters up to 8 October 2025. We found no legislative voting-intention poll with fieldwork in 2026.
That matters in three ways. The polls are a year old, and presidential round-one polls have moved a great deal over a year: at 365 days out, the round-one poll average missed each family by 4.36 points on average across the five elections, against 1.93 at 30 days (From today to polling day: drift in French presidential polls). The left and RN configurations tested in 2025 (a united left or separate lists, the RN with or without Éric Ciotti's allies) may not be the ones that stand. And the legislative correction is applied to polls whose error we cannot yet check.
If an election is called, the campaign lasts 20 to 40 days, and new polls will replace these within days. Until then the Assembly projection is best read as "an Assembly election held on the autumn 2025 polls", which is how the site labels it.
Uniform swing and local effects
Each circonscription starts from its 2024 round-one result and moves by the national change in each bloc. That is the simplest seat model there is, and in France it leaves out a lot.
Candidates and incumbency. French legislative elections are fought by individual candidates, and a sitting deputy, a well-known mayor or a dissident from the same camp can move a seat by more than any national swing. Uniform swing treats every seat's 2024 result as the whole of its local character, as though every deputy stood again and every 2024 candidate's personal vote transferred to the next candidate of the same bloc.
Changes since 2024. Ten seats have held by-elections, and several deputies have changed group. The projection starts from 2024, not from these. By-elections have low turnout and strong local candidates, so the site shows them but does not use them; that is a choice, and it means the model knows nothing of the local shifts they may reveal.
The same error everywhere. The correction and every lens change national shares only. A polling miss that falls unevenly, for instance polls that understate the RN mainly where it is already strong, cannot be expressed. The simulation adds a local random error to each seat's shares, which widens the ranges, but it is noise with no structure: it does not know which seats resemble each other.
No demographics. The INSEE statistics on each seat page are descriptive. The model does not use them to shape the swing, and we have no test yet showing that doing so would help in France.
Withdrawals and transfers
This is the weakest part of the forecast, and the part most likely to decide an Assembly election.
After round one, any candidate with 12.5% of registered voters may stay in. In 2024, 225 qualified candidates withdrew, and only 91 of the 312 seats where three or more qualified still had three candidates in round two, mostly to stop the RN. Our model assumes the same behaviour: where the RN and two or more of its opponents qualify, the weaker opponents withdraw, in every case in the projection and the lenses, and at the 2024 rate in the simulation. A campaign in which the parties do not coordinate, or coordinate against a different bloc, would produce a very different map.
Each duel is then decided by a model of round-two shares fitted on the 403 duels of 2024, which learns how each eliminated bloc's voters split. Given the real 2024 round one it picks 370 of 403 winners. That is an in-sample fit. The out-of-sample test is much less flattering:
| Fitted on | Applied to | Duel winners picked | Round-one leader alone |
|---|---|---|---|
| 2022 | 2024 | 64.5% | 61.0% |
| 2024 | 2022 | 74.8% | 78.0% |
Rates fitted on one election barely beat, or lose to, the naive rule that the round-one leader wins. Voters moved very differently between the two elections, and 2024 followed an unusually wide withdrawal of candidates against the RN. Our projection assumes that any snap election works like 2024.
The pollsters' own record shows how much this matters. Their seat projections between the two rounds, made with round one already known, missed badly at each of the last three elections: the presidential camp in 2017 (about 437 to 467 projected, 350 won), the NUPES left in 2022 (161 to 206 projected, 131 won) and the RN and allies in 2024 (188 to 221 projected, 142 won). If pollsters with round one in hand could not get round two right, a model projecting both rounds from year-old polls should be read with at least as much caution. Our simulated 80% ranges carry national and local round-one errors but not uncertainty in the duel model itself, so they are narrower than they should be.
Overseas and abroad
The Assembly has 38 seats outside metropolitan France: 27 overseas and 11 for French people living abroad. We project them like the rest, and they are the seats where that is least defensible.
Turnout is the first reason. In the 2024 round one, turnout was 68.8% across metropolitan France, 41.9% across the overseas seats and 36.6% across the seats abroad, with some overseas seats under 30%. Since qualification for round two needs 12.5% of registered voters, low turnout means that in these seats a third candidate almost never qualifies, and the withdrawal rules rarely come into play. The model handles that correctly, because it uses each seat's own 2024 registered voters and turnout. But it assumes turnout stays as in 2024.
Politics is the second. Overseas seats are often won by local parties and personalities whose ministry labels (other left, regionalist, other right) do not map onto national blocs, and a national swing in, say, the RN's share says little about Guadeloupe or French Polynesia. The seats abroad have their own electorate and, in recent elections, internet voting in the days before the in-person vote. We have no polling of either group and no evidence on how their results follow national polls.
The data are thinner too. INSEE's statistics do not cover the seats abroad, income figures are missing for 16 overseas seats, and 18 of the 38 have no official boundaries and are drawn as squares on the map (The French data: results, maps, statistics, deputies and polls).
These seats can matter. Thirty-eight seats can decide whether a bloc reaches the 289 needed for a majority, which no bloc does in 78% of our simulations.
Raw figures
In Britain our projection starts from what respondents said before the pollsters' weighting and turnout steps, because the raw error proved steadier than the published error across four elections (Why we start from British pollsters' raw answers rather than the published headline). In France only OpinionWay and Cluster17 print raw figures regularly, and they pull in opposite directions: before weighting, OpinionWay's panel has too many RN supporters and Cluster17's too few. Raw figures survive for only five past final polls, from three pollsters, on different bases, and were closer to the result at three of them. There is no basis for a raw-stage method, so we show the figures and do not use them (Rating French pollsters: accuracy, the lead and openness).
The deeper limitation is that we cannot see what the weighting does for most pollsters. French notices must state the weighting criteria, and most pollsters weight on recalled vote at a past election. Which past election they use, and how they treat respondents who do not remember or did not vote, changes the RN's and the left's shares. Without raw figures we cannot measure how much.
Publication rules before polling day
French law bans publishing electoral polls from the midnight before polling day (the night of Friday to Saturday) until the polls close on Sunday evening. From 1977 to 2002 the ban covered the whole week before each round. The final polls we measure therefore usually finish their fieldwork two or three days before the vote, and before 2002 about a week before.
This has three consequences for the forecast. First, the final-poll error we correct includes any movement in the last days, so it mixes polling error with late swing, and we cannot separate them. If the RN's runoff overstatement is late swing to its opponent rather than a problem with samples, it could vanish in a campaign that ends differently. Second, the long-view errors before 2002 are not comparable with later ones, since they include a week of campaign. Third, no new polls appear in the last weekend, so our figures freeze with the last polls published.
Uncertainty far from polling day
The polling-day forecast adds drift for the days left: the typical miss of poll averages at each distance before the five past elections, fitted as a curve. Today, 148 days out, that adds about 4.9 points per candidate. We tested whether pulling candidates back towards their family's previous result helps, as it does in Britain, and it did not, so the polls are taken as they stand. The fit rests on five campaigns; runoff polls far out are rare, so we assume round-two polls drift like round one; and the drift is the same size for every family, whether it is a large party or a small one (From today to polling day: drift in French presidential polls).
Choices made by judgement
Several choices are ours and could reasonably be made differently:
- The correction takes off half the average past miss, as in Britain. Half is a hedge between "misses repeat" and "misses are noise", not an estimate.
- A family needs two measured elections before it is corrected.
- The candidate families and legislative blocs are our groupings (The French data: results, maps, statistics, deputies and polls).
- Each pollster's latest poll within 45 days (60 for legislative polls) counts once.
- The simulation's spread for each family is its typical final-poll miss, with a floor of 1.5 points.
- The pollster ratings use the British scale, which is generous on French data (Rating French pollsters: accuracy, the lead and openness).
What would change our mind
The method is written so that the evidence decides, and the test re-runs every time the data change. Some of what follows would change the forecast automatically; some would need new work.
- The 2027 round two. If the RN's candidate reaches round two and the final polls do not overstate them, the average miss falls and the correction shrinks with it. If they are overstated again, the case strengthens. Either way the test decides.
- The 2027 round one. If round-one misses at a sixth election line up with an earlier pattern, the leave-one-out test could start to favour a correction, and it would then be applied. Nothing needs to be changed by hand.
- A legislative election. A fourth set of final polls would test the legislative correction, and a second set of duels fitted on something other than 2024 would let us test the duel model out of sample. If a model fitted on two elections beat the round-one leader rule on a third, we would use it; if not, the seat ranges would need to be widened to reflect it.
- New legislative polls. Any legislative voting-intention poll replaces the autumn 2025 figures at once.
- More raw figures. If more pollsters print raw figures through 2027 and those can be checked against the result, a raw-stage method could be tested as it was in Britain.
- Overseas polling or results. Evidence on how overseas and abroad results follow national polls would let us model those seats separately instead of by uniform swing.
- Evidence on local effects. If adding incumbency or demographics to the swing could be shown to improve the projection of 2024 from 2022, we would add it.
References
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.
How to cite
Polling Forecast (2026). What we don't know yet: limitations of the French forecast. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/france-limitations. Live figures change as new polls arrive; the date above is when the text was last revised.



