Rating French pollsters: accuracy, the lead and openness
Abstract. We rate each French pollster out of 100 on the final round-one polls of the presidential and Assembly elections, its miss on the gap between the top two, and openness. Since 13 October 2026 accuracy and the lead are ranked against France's own final polls and a short record is pulled towards the typical one: five pollsters grade B and the rest C.
Summary
We rate each French pollster out of 100 on the same three parts as the British ratings: accuracy of its final polls (50 points), its miss on the gap between the top two (20 points), and openness, meaning how much of its working it publishes (30 points). Accuracy and the lead are measured on the final round-one polls of the presidential elections of 2002 to 2022 and the legislative elections of 2017 to 2024, 62 polls in all, and each pollster's record is pulled towards that of a poor pollster as if it had two more elections at that level. Openness is scored on six items specific to France, starting with whether the pollster prints its raw figures in the notice it must file with the Commission des sondages. The scale's end points are the British ones, but French final polls have missed by less per party, so the "poor pollster" prior is much kinder than in Britain and most French pollsters score an A or B. We say so plainly below. The ratings are a guide for readers and are not used to weight the polls. The paper also covers each pollster's lean against today's average and the raw figures that OpinionWay and Cluster17 print, which we show but do not build on. The British counterparts are Rating the British pollsters: accuracy, the Labour–Conservative lead and openness and Why we start from British pollsters' raw answers rather than the published headline.
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.
Pollster grades were rescaled on 13 October 2026. The first scale scored accuracy between fixed end points (a miss of 0.75 points or less earned full marks, 2.75 or more nothing) and pulled every short record towards a poor pollster, one that three in four final polls beat. Almost every pollster ended up with a D or an E, which told readers little. The rating now works like this:
- Accuracy (50 points) is ranked against the country's own checked final polls. If a share \(r\) of them missed by less than the pollster's average miss, it scores \(50\,(1-r)^{0.75}\): the typical (median) final poll earns about 30 and one in the best tenth about 46. Ranking within the country means a pollster is not marked down because its country is harder to poll.
- The lead (20 points) is scored the same way on the miss on the gap between the two leaders, \(20\,(1-r)^{0.75}\).
- A short record is pulled towards the typical (median) final poll, as if the pollster had two more elections there, not towards a poor one.
- Openness (30 points) is scored out of the items checked so far; an item not yet checked counts neither for nor against.
- Every pollster with a final poll on record gets a grade. One whose publications we have not reviewed is graded on its record alone (accuracy and lead scaled to 100) and marked "record only".
The grade boundaries are unchanged: A from 75, B 60–74, C 45–59, D 30–44, E below 30. Across the active pollsters of all ten countries the grades went from mostly D and E to A 8, B 68, C 76, D 31 and E 5. Where the sections below give the old end points, the poor-pollster prior or grades from before 13 October, they describe the first version of the scale; the live tables show today's.
In France the new scale gives B to Harris Interactive (70), Cluster17 (67), Ifop (65), Odoxa (65) and Ipsos (63), and C to OpinionWay, Elabe, Verian and YouGov. These are the figures of 9 November 2026; the table on the pollsters page is current.
What the rating is for
Our averages give every pollster equal weight: each pollster's latest poll within 45 days of the newest counts once, so a pollster that publishes weekly does not outweigh one that publishes monthly (How we read French polls: what we correct, what we don't, and why). That is a judgement that pollsters are interchangeable, and it is worth checking. The rating asks a narrower question: on the evidence we can check, how close has this pollster's final poll been to the result, and how easily can a reader see how its figures were made?
It does not say which pollster is right today. A pollster that has one candidate well above the others may be the only one that is right, and we find out on polling day. Nor does it feed the forecast. It sits on the pollsters page beside each pollster's current lean, as information.
The evidence: final polls since 2002
Which polls count
For each election the final poll is the one whose fieldwork ended latest before polling day, provided it was published before voting began. Where a pollster ran two series with the same final fieldwork date (Ifop in 2012, Ipsos in 2022), only one counts. The accuracy parts of the rating use two sets of final polls:
- Presidential round one, 2002–2022: 40 polls at five elections (6, 6, 8, 9 and 11 pollsters).
- Legislative round one, 2017–2024: 22 polls at three Assembly elections (8, 7 and 7 pollsters).
Round-two presidential polls are not counted. That choice has a large effect. The biggest French misses are in round two, where the final polls overstated the Rassemblement national's candidate at every runoff it reached: by 3.2 points in 2002, 4.5 in 2017 and 3.3 in 2022 (Measuring French polling error, 1965–2024). Every one of the pollsters on record made that miss, at all three runoffs. Counting round two would therefore lower every rating by a similar amount, and would weigh most heavily on the pollsters that happened to poll those three runoffs. The rating is built from round one, where the pollsters differ more from one another; round two is handled by the correction in our projection (The 2027 presidential forecast: line-ups, runoffs and simulation).
Pollsters are identified by firm, not by brand. Sofres, TNS Sofres, Kantar Sofres, Kantar Public and Verian (Kantar Public's name since 2023) are one series, as the same firm under successive names, so Verian carries the whole record. Harris Interactive and Toluna Harris Interactive are one pollster; the older Louis Harris France and LH2 are kept separate.
Two measures per poll
Candidates change between elections, so errors are measured by candidate family: Jean-Marie Le Pen, Marine Le Pen and the RN are one family, the Republican candidates another, and so on. Legislative polls are measured by bloc. For pollster \(j\) at election \(t\), let \(e_{j,t,p}\) be its final poll's share for family or bloc \(p\) minus the result, in points. The poll's miss is the mean absolute error across the families measured at that election:
where \(P_t\) is the set of families that appear in the result and in at least half the final polls. It holds between eight and eleven families per election, against five parties in the British rating.
The lead error is the error on the gap between the two families or blocs that actually finished first and second, \(A\) and \(B\):
positive when the poll had the winner too far ahead. In a presidential round one this is the gap that decides who tops the poll; in a legislative round one it is the gap between the two largest blocs. A pollster's own record is the average miss \(\bar{m}_j\) and the average absolute lead error \(\bar{\ell}_j\) over its \(n_j\) final polls.
Scoring and shrinkage
The scales
As in Britain, a miss of 0.75 points or better earns the full 50 accuracy points and 2.75 or worse earns none; a lead error of 1 point or better earns the full 20 and 9 or worse earns none, with straight lines in between:
where \(\tilde{m}_j\) and \(\tilde{\ell}_j\) are the rated values defined next.
Pulling towards a poor pollster
One good final poll is not a record. Each pollster's figures are shrunk towards a prior, with the prior given the weight of two elections:
The prior is deliberately poor: \(m^{*}\) is the 75th percentile of the miss across all 62 French final polls, and \(\ell^{*}\) the 75th percentile of the absolute lead error, so an unknown pollster is assumed to be one that three in four final polls beat (Efron and Morris, 1975, set out the general idea of shrinking towards a common value). The code is WorldPollsterRatings.Build in Elections.Core/World/WorldPollsterRatings.cs, and it computes the percentiles from the data every time.
What the French prior turns out to be
With the current 62 polls the prior is \(m^{*} = 1.45\) points and \(\ell^{*} = 2.61\) points. In Britain the same calculation gives 2.50 and about 8.3 (Rating the British pollsters: accuracy, the Labour–Conservative lead and openness). French final polls have missed by less per party (1.21 points on average across the 62, with the best at 0.28 and the worst at 2.51), and their lead errors have been far smaller (1.85 points on average).
Part of that is real accuracy and part is arithmetic. A French round one has ten or more candidates, and the smaller ones poll a few points and miss by a fraction of a point, which pulls the average miss per family down. The British miss averages over five parties, all of them large enough to miss by several points. The lead error differs for a similar reason: in a British election the lead is between two parties that together take well over half the vote, while in a French round one it is between two candidates who have each taken about a third of the vote or less.
The consequence is that the scale is not punitive in France. The table shows the best score a pollster could reach if every one of its final polls had been perfect, with full openness.
| Final polls checked | Rated miss | Accuracy points | Rated lead error | Lead points | Best possible total |
|---|---|---|---|---|---|
| 0 | 1.45 | 32.5 | 2.61 | 16.0 | 78 (A) |
| 1 | 0.97 | 44.6 | 1.74 | 18.1 | 93 (A) |
| 2 | 0.72 | 50.0 | 1.31 | 19.2 | 99 (A) |
| 3 | 0.58 | 50.0 | 1.05 | 19.9 | 100 (A) |
Table: the ceiling for a pollster with no error at any checked election, using the current French prior and full openness points (30).
In Britain a pollster with no checked record cannot rise above a D. In France the same pollster, if fully open, would score 78, an A. The British scale was set to be hard because British polls missed badly at three of the last four elections; with the same end points, French final polls look good. We have kept the end points the same so that the two countries' ratings are on one scale, and we report the consequence rather than retune the French scale to produce a spread of grades we would like. A reader should read a French A as "at least as accurate as a typical French final poll, and open", not as the rare mark it is in Britain.
A worked example
Ipsos has eight final polls in the record: five presidential round ones and three legislative. Its misses were 1.23 (2002), 1.15 (2007), 1.07 (2012), 0.56 (2017) and 1.38 (2022) in the presidential round one, and 1.75 (2017), 0.97 (2022) and 0.91 (2024) in the legislative. The average is \(\bar{m} = 1.13\). Its lead errors averaged 1.33 points in absolute value. Then
giving \(A = 50 \times (2.75 - 1.19)/2 \approx 39.0\) and \(L = 20 \times (9 - 1.58)/8 \approx 18.5\). With 20 openness points (below) the total rounds to 78. With eight elections the prior carries only a fifth of the weight, so Ipsos's rating is mostly its own record.
Openness
What French law already requires
France regulates published electoral polls more than Britain does. Under the law of 19 July 1977, as amended in 2016, the first publication of a poll must give the sample size, the fieldwork dates, the full wording of the questions and the margin of error. Every published electoral poll must also file a notice with the Commission des sondages, and the notice must state the proportion of people who did not answer each question and the criteria used to weight ("redresser") the raw results. This gives French readers a floor of disclosure that British readers do not have. It also means that two of our items are met by almost everyone and do little to separate pollsters.
The six items
Each item is recorded as Yes, Partial or No from a review of the notices and the pollsters' own releases, and scored as full points, a third, or nothing. Items we could not check ("not verified" or "unknown") score nothing.
| Item | Points | Why it matters |
|---|---|---|
| Raw figures in the notice | 8 | Shows what respondents said before weighting, and so what the weighting did |
| Weighting variables shown | 6 | Recalled past vote and other targets decide the headline |
| Full report published | 6 | Question wording, sub-groups and every line-up tested |
| Notice filed with the Commission des sondages | 4 | Required by law for published electoral polls |
| Sample size and fieldwork dates | 3 | The basics for judging a poll |
| Undecided share | 3 | Shows how firm the figures are |
The raw item carries the most weight for the same reason as in Britain: raw figures are the only way an outsider can see how far weighting moved the headline. The weighting item matters more in France than its points suggest. Most French pollsters weight on recalled vote at a past presidential or legislative election, and the choice of which election (the 2022 presidential first round, the 2024 legislative first round, the 2024 European election) can change what the headline says about the RN and the left.
The review
The review was compiled in September 2026 from the Commission's notices and the pollsters' releases already read for the poll data (The French data: results, maps, statistics, deputies and polls). The current values are these.
| Pollster | Raw | Weighting | Report | Notice | Sample | Undecided | Points |
|---|---|---|---|---|---|---|---|
| Cluster17 | Yes | Yes | Yes | Yes | Yes | Yes | 30 |
| Ifop | Partial | Yes | Yes | Yes | Yes | Yes | 25 |
| Le Terrain | Yes | Yes | – | Yes | Yes | Yes | 24 |
| Harris Interactive | No | Yes | Yes | Yes | Yes | Yes | 22 |
| OpinionWay | Yes | Yes | – | Yes | Yes | Partial | 22 |
| Elabe | No | Yes | Yes | Yes | Yes | Partial | 20 |
| Ipsos | No | Yes | Yes | Yes | Yes | Partial | 20 |
| Odoxa | No | Yes | Yes | Yes | Yes | No | 19 |
| CSA | Partial | Yes | – | Yes | Yes | Yes | 19 |
| YouGov | Partial | Yes | – | Yes | Yes | Yes | 19 |
| Verian | No | Yes | – | Yes | Yes | Yes | 16 |
| BVA | No | Yes | – | Yes | Yes | Partial | 14 |
| AtlasIntel | – | – | – | Yes | Yes | – | 7 |
| LH2 | – | – | – | Yes | – | – | 4 |
| Louis Harris | – | – | – | Yes | – | – | 4 |
Table: openness review, September 2026. A dash means not verified or unknown, which scores nothing.
Three caveats belong with this table. First, "not verified" is not "No": OpinionWay, CSA and others may well publish full reports that the review did not find, and they lose those points until someone checks. Second, LH2 and Louis Harris no longer poll and their notices predate 2016, when notices went online; they are credited with the notice only because the law required one. Third, membership of Syntec, the market-research trade body, was recorded where possible but is not scored, because its member directory refused automated requests and could not be checked.
Grades and the current table
The score is the rounded sum \(S_j = \operatorname{round}(A_j + L_j + O_j)\), and the grades are the British ones: A from 75, B from 60, C from 45, D from 30, E below 30. A rating is marked provisional when the pollster has fewer than two final polls in the record, since the prior then carries at least half the weight. Only pollsters whose publications have been reviewed get a score.
| Pollster | Grade | Score | Accuracy | Lead | Openness | Elections |
|---|---|---|---|---|---|---|
| Harris Interactive | B | 70 | 35 | 13 | 22 | 7 |
| Le Terrain | B | 67 | 30 | 14 | 24 | 1 |
| Cluster17 | B | 67 | 25 | 13 | 30 | 3 |
| Louis Harris | B | 66 | 30 | 6 | 30 | 1 |
| LH2 | B | 66 | 22 | 14 | 30 | 3 |
| Ifop | B | 65 | 30 | 10 | 25 | 11 |
| Odoxa | B | 65 | 33 | 13 | 19 | 4 |
| Ipsos | B | 63 | 30 | 13 | 20 | 11 |
| OpinionWay | C | 59 | 27 | 11 | 22 | 7 |
| Elabe | C | 56 | 26 | 10 | 20 | 5 |
| Verian | C | 53 | 28 | 9 | 16 | 9 |
| YouGov | C | 51 | 19 | 14 | 19 | 3 |
| CSA | C | 48 | 22 | 7 | 19 | 6 |
| BVA | C | 45 | 23 | 7 | 14 | 9 |
| AtlasIntel | D | 40 | 22 | 11 | 7 | 1 |
The table is computed from the database each time the page is built. Pollsters with no poll in the 365 days before the newest poll are marked "not polling now" on the pollsters page. Several features will hold however the individual numbers move:
- Accuracy spreads the pollsters much less than openness does. Accuracy points run from about 27 to 40 of the 50; openness runs from 4 to 30.
- The lead part barely separates anyone. With a lead prior of 2.6 points, almost every pollster earns 16 to 19 of the 20 lead points. Louis Harris, whose single final poll (2002) had the Republican family (Jacques Chirac and Alain Madelin) 5.1 points too far ahead of Jean-Marie Le Pen, is the one clear exception.
- YouGov has the lowest accuracy points among pollsters with a record: its two final polls (2022 presidential and 2022 legislative) missed by 1.58 and 2.22 points per family.
- Cluster17 tops the table on openness. It is the only pollster to score on every item, and that, more than its three final polls, puts it first.
Who each pollster leans towards now
Beside the ratings, the pollsters page shows each active pollster's lean: its latest poll of the main line-up, taken within the 45-day averaging window, minus the average of all pollsters' latest polls, for each candidate:
The pollster pages also show its past lean: the average error of its final polls for each family at past elections.
A lean is not an error. The average is made of the pollsters themselves, so the leans roughly cancel by construction, and the pollster furthest from the pack may be the one closest to the result. The comparison of today's lean with past lean is the useful one. A pollster that is low on the RN candidate today and was also low on the RN at past round ones is behaving as before; one that is low today but was high in the past has changed something, perhaps its weighting. With at most five presidential elections per pollster, and misses that change direction from one election to the next, the past lean is a weak guide, and the page says so.
Raw figures: OpinionWay and Cluster17
What they print
French notices must state the weighting criteria, but only some pollsters print the unweighted ("brut") vote intention next to the published figure. In the 2027 cycle two do so regularly. OpinionWay prints a "Brut" column among respondents certain to vote, the same base as its published figure. Cluster17 prints "Brut" and "Redressé" (weighted) figures; its published figure is the weighted one rounded to the nearest half point. We hold raw figures from ten 2027 round-one polls with fieldwork to mid-September 2026: seven from OpinionWay and three from Cluster17. The polls page lists them.
They pull in opposite directions. In those polls, OpinionWay's raw figure for the RN candidate was above its published figure by 2.6 points on average across 20 line-ups tested (from level to 6 points above), and its raw figure for Jean-Luc Mélenchon was 2.9 points below. Cluster17's raw figure for the RN candidate was 3.9 points below its published figure across 8 line-ups (between 3.0 and 4.4 points below), and its raw Mélenchon figure 1.5 points above. One panel reaches too many RN voters before weighting and the other too few. Both pollsters then weight on recalled vote, and their published figures end up much closer to each other than their raw ones.
What history there is
Raw figures for past final polls survive for only five polls, all from 2017 and 2022, because notices before 2016 are not online and most pollsters' notices give weighting targets rather than raw results. Measured against the result over the families with a raw figure:
| Final poll | Published: miss per family | Raw: miss per family | Raw base |
|---|---|---|---|
| Ifop, 2017 round one | 0.65 | 0.88 | All registered voters expressing a choice |
| OpinionWay, 2017 round one | 0.90 | 1.90 | Certain to vote |
| Kantar, 2022 round one | 1.66 | 1.52 | Not stated in our file |
| Kantar, 2022 round two | 3.55 | 2.95 | Not stated in our file |
| OpinionWay, 2022 round two | 1.55 | 0.45 | Certain to vote |
The raw figure was closer three times and further out twice. In the 2022 runoff both raw figures had Marine Le Pen lower than the published ones (41 against 43 for OpinionWay, 44.4 against 45 for Kantar), and she won 41.45%. In the 2017 round one OpinionWay's raw figure had her at 19 against a published 22 and a result of 21.3, so weighting moved her the right way and overshot. In the 2022 round one Kantar's weighting raised both Éric Zemmour (9.0 raw, 11 published, 7.07 result) and Valérie Pécresse (6.0 raw, 8 published, 4.78 result), and both were overstated.
Why we show them but do not use them
In Britain our projection starts from the raw answers, because across four elections and several pollsters the raw error turned out to be larger than the published error but steadier, and so partly correctable (Why we start from British pollsters' raw answers rather than the published headline). None of the conditions for that holds in France yet.
- Two pollsters, opposite directions. A correction learned from OpinionWay's raw figures would push the RN one way and one learned from Cluster17's the other. There is no industry-wide raw error to estimate, only two house effects.
- Almost no history. Five past final polls, from three pollsters, on at least two different bases, cannot support a test of the kind that decided the British method. Our rule is to correct only when a leave-one-out test says it helps, and there is nothing to run the test on.
- "Raw" means different things. OpinionWay's raw figure is already filtered to those certain to vote; Ifop's in 2017 was among all registered voters. Mixing them would mix a turnout step into some raw figures and not others.
So the raw figures are shown for readers who want to see what the weighting does, and they score openness points for the pollsters that print them, but they do not enter the forecast. If more pollsters print raw figures through 2027, and the final polls of 2027 can be checked raw against the result, the question can be tested properly.
Limitations
The record is short. No pollster has more than eight final polls in it, and the round-one misses change direction from one election to the next, so a pollster's average miss says as much about which elections it polled as about the pollster. Methods also change: a pollster's weighting targets, panel and filter for likely voters can all differ from one election to the next, so a good 2012 poll says little about a 2026 one.
The prior is recomputed from the data and will move as final polls are added. The 2027 presidential election will add a new set of round-one final polls, and with only 62 polls in the record they may shift the 75th percentile noticeably.
The scale's end points, the prior's weight of two elections and the openness weights are judgements carried over from Britain, not estimates. On French data they give generous grades. We have said why we keep them, but a reader who prefers a scale set on French evidence would reach lower grades from the same inputs, which is why the three components are published separately.
Round-two polls, where every pollster made the same large miss three times, are outside the rating. The rating therefore describes the pollsters' round-one work, not the forecast that matters most for who wins.
References
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 (as amended, 25 April 2016).
Commission des sondages. Notices des sondages, https://www.commission-des-sondages.fr/notices/.
How to cite
Polling Forecast (2026). Rating French pollsters: accuracy, the lead and openness. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/france-pollsters. Live figures change as new polls arrive; the date above is when the text was last revised.



