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

From today to polling day: drift in French presidential polls

Abstract. Polls taken months before a French presidential election have missed the result by far more than final polls. Unlike in Britain, pulling parties back towards their previous result did not help when tested on 2002–2022, because French party families rise and fall too much between elections, so the polling-day forecast takes today's polls as they stand and adds the uncertainty that the time remaining has carried in the past, fitted as a curve that grows with the logarithm of the days left.

Polling Forecast · updated 9 November 2026 · 11 min read · France, election day, horizon, poll drift, pull-back, presidential polls, uncertainty, timeline of elections

Summary

The presidential forecast on the rest of the site answers a hypothetical question: who would win if the election were held today. Round one is in fact 148 days away, and polls taken this far out have missed the result by far more than final polls do. This paper describes how the site turns today's polls into a forecast for polling day.

In Britain, months or years before an election, the older parties have finished much closer to their previous result than their polls suggested, and the election-day model pulls them part of the way back (From today to polling day: how far British polls move before a general election). We fitted the same curve to French presidential polls from 2002 to 2022. The fitted curve pulls back only slightly, and when each election is predicted with the curve fitted on the others it is no closer than the polls as they stand. So the French polling-day forecast keeps today's polls unchanged and adds only the extra uncertainty that the time remaining has carried in the past: about 4.9 points per candidate at 148 days. The result is that the favourite's chance changes little, while the outsiders' chances rise from nothing to something.

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.

Today is not polling day

A poll records how people say they would vote at the moment they are asked. Months before an election, with no campaign under way and, in France, no settled list of candidates, those answers carry a large component that will have changed by polling day. The literature on the "timeline" of elections finds that polls become steadily more informative as the election approaches, and that early polls contain much short-lived movement (Erikson and Wlezien, 2012; Wlezien et al., 2013). Gelman and King (1993) put the puzzle for American presidential elections: campaign polls swing about a great deal, yet the outcome can be predicted from things known well in advance.

Errors measured on final polls (Measuring French polling error, 1965–2024) describe the last few days only. They say nothing about how opinion moves in the months before. A forecast for polling day therefore needs its own model, and the site keeps it separate from the "election today" figures, on the election-day page.

The model's task is narrow: given today's round-one polls and a known number of days to go, to estimate each candidate's share on polling day and how uncertain that estimate is. It does not try to foresee events, and it does not model who will stand.

The evidence: polls at fixed distances from polling day

Horizon points

For each presidential election from 2002 to 2022 we take every published round-one poll in the database and average them at six distances before polling day: 30, 61, 91, 182, 273 and 365 days. Call the distance \(h\). The average at distance \(h\) uses every poll whose fieldwork ended within 30 days either side of the date \(h\) days before round one. Within that 60-day window each pollster's polls are first averaged, and then the pollsters are averaged with equal weight, so that a pollster that published often does not dominate. Shares are summed by candidate family, as in Measuring French polling error, 1965–2024, so that a family is followed whichever of its candidates a poll tested.

Each resulting record, a horizon point, holds the election, the distance, the family, the poll average \(\bar{y}_{p,e}(h)\), the family's share at the previous presidential election \(v_{p,e-1}\), and its share at this one \(v_{p,e}\). A family is included only if it also stood at the previous election, because the model needs a previous result to pull towards. That leaves 198 horizon points: six or seven families at each of six distances in each of five elections. The previous result for 2002 is the 1995 round one.

The polls come mainly from Wikipedia's poll tables, checked against the final polls on file. Coverage is uneven. At 30 days before round one the window holds between 48 polls (2002) and 171 (2022); at nine months out it holds as few as one (2017) or three (2002 and 2007). The longest distance is one year, because the poll tables thin out beyond that.

The previous result as a forecast

The obvious alternative to a poll taken a year out is to assume nothing has changed since the last election. We call the previous result the prior. In Britain, the prior has been a surprisingly good forecast for the three older parties at long range. In France it is a poor one: the previous result missed the next result by 5.76 points per family on average, against 4.36 for the polls a year out and 1.93 for the polls a month out.

The reason is plain from the results. French party families rise and fall between presidential elections by amounts that British parties rarely match. The Socialists took 28.6% in 2012 and 6.4% in 2017; the mainstream right took 20.0% in 2017 and 4.8% in 2022; the Macron camp did not exist in 2012. A model that pulls such families back towards their previous result pulls them towards a number that has little to do with the next one.

The pull-back curve

Form

The British model takes the polling-day share as a weighted average of today's polls and the previous result:

\[ \hat{v}_{p}(h) = v_{p}^{\mathrm{prior}} + w(h)\,\bigl(y_{p} - v_{p}^{\mathrm{prior}}\bigr), \qquad w(h) = f + (1 - f)\,e^{-h/\tau}. \tag{1} \]

The weight \(w(h)\) on today's polls is 1 on polling day and falls towards a floor \(f\) as the distance grows, at a rate set by the time constant \(\tau\) in days. A floor of 1 means the polls are taken as they stand at every distance; a floor of 0.5 means only half the change since the last election counts at long range.

Fit

We fit \(f\) and \(\tau\) by least squares over all 198 horizon points, searching a grid of \(f\) from 0 to 1 in steps of 0.01 and \(\tau\) from 5 to 600 days in steps of 5. The best fit is \(f = 0.63\) and \(\tau = 600\) days. The time constant sits at the top of the grid, which means the data would prefer an even slower decline: in effect, the curve is close to flat and close to 1 over the distances we have. At 30 days it puts a weight of about 0.98 on today's polls, at six months about 0.90 and at a year about 0.83. In words, the best-fitting curve pulls a family back towards its previous result by about a tenth of the gap at six months and a sixth at a year. The British curve counts only a little over half of the change since the last election from about three months out.

The test, and why it fails

A curve fitted to all five elections will always look at least as good as the polls on those same elections. The test that matters is out of sample. We predict each election with the curve fitted on the other four, at each distance, and compare its average miss per family with that of the polls as they stood.

Days before polling dayElectionsPolls' missPrevious result's missPull-back's missRMSE
3051.935.761.962.72
6152.795.762.824.05
9153.245.763.284.51
18253.815.763.814.98
27354.465.764.435.95
36554.365.764.305.50
Live Presidential round one: polls against the result by distance from polling day, each election predicted from the others Computed from the database, polls to 19 November 2026.

The two are level. Averaged over the six distances, the polls as they stood missed by 3.43 points per family and the pulled-back polls by 3.43. The pull-back was slightly closer at 273 and 365 days and slightly further out at 30, 61 and 91 days. The rule is the same as for every correction on the site: the pull-back is used only if, averaged over the test, it is closer than the polls by more than 0.01 points. It is not, so the French model sets \(f = 1\) and takes the polls as they stand at every distance.

This is a result about France, not about the method. The pull-back helps in Britain because British parties' results are anchored: a party that surges in mid-term tends to fall back, and the previous result is a fair guide to where. French families are not anchored in that way, for the reasons above. The same test on German Bundestag polls also fails, with the pull-back further out than the polls at every distance (From today to polling day in Germany).

Drift by distance

Taking the polls as they stand does not mean trusting them as much a year out as a week out. The typical miss grows with distance, and the forecast has to carry that uncertainty.

The last column of the table above gives the root mean square miss of the polls at each distance, pooled over the elections and families. It rises from 2.72 points at 30 days to about 5.5–6 at nine months to a year. We fit it with a logarithmic curve:

\[ \text{RMSE}(h) = r_{30} + k \,\ln\!\left(\frac{h}{30}\right), \qquad r_{30} = 3.01, \quad k = 1.16. \tag{2} \]

That gives about 3.0 points at 30 days, 4.3 at 90 days, 5.1 at 180 days and 5.9 at a year. The logarithm reflects the shape of the evidence: the typical miss grows quickly over the first few months and more slowly after that.

The drift standard deviation used by the simulation is \(\sigma_D(h) = \max(0.5, \text{RMSE}(h))\) for \(h \ge 30\), scaled down in proportion to the days left inside the last month, so that it reaches zero on polling day. At 148 days it is 4.9 points.

The miss at each distance varies a great deal from one election to the next. Here is the polls' average miss per family, election by election:

Days out 2002 2007 2012 2017 2022
30 2.47 1.17 0.91 2.55 2.55
61 3.60 0.79 1.24 4.07 4.25
91 4.26 1.73 1.30 3.80 5.08
182 4.86 2.97 1.66 4.58 4.99
273 5.81 3.55 1.99 6.30 4.67
365 5.67 3.59 2.63 5.22 4.68

2012 was a stable election: polls a year out missed by 2.6 points per family, little more than the final polls of some other years. 2002 and 2017 were not. In 2017 the polls nine months out missed by 6.3 points per family, reflecting a campaign in which the mainstream right's candidate, François Fillon, fell from favourite to third and Emmanuel Macron rose from outside the party system. The drift curve averages over elections like these; it does not know which kind 2027 will be.

How drift enters the forecast

The polling-day forecast runs the round-one simulation of The 2027 presidential forecast: line-ups, runoffs and simulation again, with 10,000 runs and the drift added. Each candidate's share is drawn with a standard deviation that combines the family's final-poll miss and the drift:

\[ \sigma_c(h) = \sqrt{\max\!\left(1.5,\ \text{RMSE}_{f(c)}\right)^2 + \sigma_D(h)^2}. \tag{3} \]

The round-two draw widens in the same way: \(\sqrt{\sigma_2^2 + \sigma_D(h)^2}\), where \(\sigma_2 \approx 4.4\) is the round-two spread used for the "election today" forecast. We use the same drift for round two as for round one, because head-to-head polls a year ahead are too rare to measure their own drift.

Six months out, for example, the drift is 5.1 points. A candidate whose family's final-poll miss is the floor of 1.5 points then has a round-one spread of about 5.3 points; Marine Le Pen's, with the RN's 2.5, about 5.7; and the round-two spread becomes about 6.8 points. The drift dominates both.

Adding the drift to the final-poll error counts the final-poll error twice, since the drift is measured as the miss against the result, which already contains it. The British model makes the same choice and says so (From today to polling day: how far British polls move before a general election). At long range the double count is small beside the drift itself; close to polling day it is not, and the bands will be somewhat too wide.

The polling-day chances

CandidateTodayPolling dayWins todayWins on polling day
Marine Le Pen32.832.899%91%
Jean-Luc Mélenchon18.318.3under 1%under 1%
Édouard Philippe12.212.21%5%
Raphaël Glucksmann11.611.6under 1%2%
Bruno Retailleau6.06.00%under 1%
Éric Zemmour3.53.50%under 1%
Marine Tondelier2.62.60%under 1%
Live Today against polling day (148 days away) Computed from the database, polls to 19 November 2026.

The shares are the same in both columns, because the polls are taken as they stand. The chances differ. Marine Le Pen wins 99% of simulated elections held today; when one candidate has a large round-one lead, as at the time of writing, that chance changes little on polling day, because the leader still qualifies in almost every run and the round-two picture is the same. The candidates behind gain most. A candidate with no chance of reaching round two today, on today's shares and the final-poll error alone, can reach it in some of the polling-day runs, because five points of drift is enough to close a gap of several points. That is the main thing the polling-day forecast adds: not a different favourite, but a truer sense of how open the race for second place still is.

The Bardella variant is treated the same way, with one difference. Its drift is counted from the date of its own newest poll, which is older than the main line-up's, so it carries more drift.

The Assembly

No legislative election is scheduled, so there is no polling day to forecast. If the President dissolves the Assembly, the constitution requires the election to be held between 20 and 40 days later. The Assembly projection (From votes to seats in the National Assembly) already carries the final-poll uncertainty, and its polls are a year old. An election called at short notice would bring new polls, and those, rather than any model of drift, would be the thing to watch.

Limitations

Five elections. The test has five elections and six or seven families in each. The verdict that pulling back does not help is a verdict on thin evidence, and a different set of elections might have given a different answer. It would take a clear improvement over several French elections before we would pull the polls back.

Families, not candidates. The prior is the family's previous result, not the candidate's. A family whose candidate changes can behave very differently from one election to the next, which is part of why the prior is weak.

Who stands. Candidacies close in March 2027. A withdrawal or a new candidate would change the field more than any drift the model allows for. The model follows the line-ups pollsters test and says nothing about who will be on the ballot.

Events. The evidence covers five campaigns, including 2017, when the leading candidate's campaign collapsed. The next shock may be larger, or come later in the campaign, than any of these.

Independent drift. The drift is drawn independently for each candidate. In reality, one candidate's rise is another's fall, and candidates of the same camp tend to move against each other.

Round two far out. We apply round one's drift to round two because runoff polls a year ahead are rare. Whether head-to-head polls drift more or less than round-one polls is unknown.

References

Erikson, R. S. and Wlezien, C. (2012). The Timeline of Presidential Elections: How Campaigns Do (and Do Not) Matter. Chicago: University of Chicago Press.

Gelman, A. and King, G. (1993). Why are American presidential election campaign polls so variable when votes are so predictable? British Journal of Political Science, 23(4), 409–451.

Wlezien, C., Jennings, W., Fisher, S., Ford, R. and Pickup, M. (2013). Polls and the vote in Britain. Political Studies, 61(S1), 66–91.

Constitution of 4 October 1958, article 12.

How to cite

Polling Forecast (2026). From today to polling day: drift in French presidential polls. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/france-election-day. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like