If we're wrong: replaying past polling misses on today's French polls
Abstract. For a National Assembly election, the site replays the round-one polling misses of 2017, 2022 and 2024 on today's legislative polls and works out all 577 seats again under each, counting a seat as at risk when at least one past miss would change its winner. For the presidency it replays the round-one misses of 2002 to 2022 on today's candidates. The lenses are scenarios shown beside the projection, not averaged into it, and they hold round-two behaviour fixed, which is where the largest French seat misses have come from.
Summary
A lens is one reading of today's polls: the polls as published, our projection, or the polls with a past election's final-poll miss taken off. In France the site keeps two sets. For a National Assembly election it replays the round-one misses of the legislative elections of 2017, 2022 and 2024 on today's legislative polls and works out all 577 seats again under each, including qualification, withdrawals and the round-two duels. For the presidential election it replays the round-one misses of 2002 to 2022 on today's candidate shares and shows who would reach round two. The lenses are scenarios, not forecasts: none of them is averaged into the projection, and none carries a probability. Their main use is to show how much of the Assembly projection depends on the polls being right. Today 91 of the 577 circonscriptions change hands under at least one of the three past legislative misses. This paper defines the lenses, sets out the misses they replay, explains how "seats at risk" is counted and shown on the map and the seat pages, and says what the lenses cannot capture: above all, the round-two behaviour that has produced the largest French seat misses. The British version of the same idea is Lenses: reading the same British polls several ways.
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.
- The lenses on the map (November). The replays are also on the map as "The polls miss again", beside "You set the vote".
- When the polls missed (2 November): Thirteen per cent in the polls, second place on the night: Le Pen and the 21 April 2002 polling miss and Right on the votes, wrong on the seats: the front républicain and France's 2024 projection miss.
Why a French forecast needs lenses
Any forecast built on polls makes an assumption about how wrong the polls are. Our rule for France, set out in How we read French polls: what we correct, what we don't, and why, is the British one: take off half the average past miss, and only where doing so would have made past forecasts better when each election is predicted without knowing its result. That rule gives one best estimate per contest. It does not say how far the answer could move if this election's miss looks like one of the past ones rather than like their average.
French polling misses are not small and they are not all alike. At the last three Assembly elections the final round-one polls overstated the Rassemblement national (RN) and its allies each time, but by 4.3 points in 2017, 0.5 in 2022 and 2.9 in 2024. The presidential round-one record is less orderly still: the far right was understated by 3.9 points in 2002 and overstated by 3.3 in 2007; Jean-Luc Mélenchon was overstated by 2.5 in 2012 and understated by 4.8 in 2022. A reader who wants to know "what if the polls are as wrong as last time?" deserves an answer for each "last time", worked out with the same seat engine as the headline.
That is what a lens provides. Each lens changes only the national round-one shares. Everything downstream (the 2024 baseline in each circonscription, the qualification rules, withdrawals, the duel model) is identical, so any difference between two lenses comes from the assumed polling error and nothing else.
What a lens is
The replay
Let \(v_p\) be today's polling average for bloc \(p\) (for the Assembly) or candidate family \(p\) (for the presidency), in percent. Let \(e_{p,t}\) be the error of the final polls at past election \(t\): the average of each pollster's final poll minus the result, so positive means the polls had the party too high. A replay lens for election \(t\) subtracts that error in full from every party it measured, with a floor of 0.1 points:
and then rescales so that the shares keep the total they had before:
This is LensMath.Replay in Elections.Core/World/WorldElectionDay.cs. A party counts as measured at election \(t\) if it appears in the result and in at least half of that election's final polls. Parties that did not exist, or that the pollsters did not report separately, are left alone apart from the rescale. So a replay of 2017 does not touch Reconquête, which first stood in 2022, and a replay of 2002 does not touch Édouard Philippe's family, which first stood as Ensemble in 2017.
The same sign convention runs through the whole site: the error is poll minus result, and the lens subtracts it. A bloc the polls overstated at election \(t\) goes down in the replay of \(t\).
How a replay relates to our projection
Our projection for the Assembly takes half the mean past miss off each bloc measured at two or more of the three elections:
again rescaled. Each replay subtracts one full \(e_{p,t}\). So for any bloc with a full record, the projection sits halfway between the polls as they stand and the average of the three replays, before the rescale. That is a useful way to read the lens table below: the projection is a deliberately partial version of "the polls miss the way they usually have", and each replay is one complete historical miss.
One consequence matters for the count of seats at risk. The replays start from the polls as they stand, not from our projection. Our projection already moves the RN down and the non-aligned right up, so a replay whose own RN error was small, such as 2022, can land closer to the published polls than to our projection. A seat can therefore be at risk because our own correction decides it, not only because some historical miss would. That is intended: in both cases the call depends on how wrong the polls turn out to be.
Blocs, not parties
Legislative polls in France are published by list or party group, and the groups change between elections. We sum them into blocs using the ministry's candidate labels (nuances), as described in The French data: results, maps, statistics, deputies and polls: the united left (NUPES in 2022, the Nouveau Front populaire in 2024, and in 2017 the combined Socialist, France Insoumise, Communist and Green vote, since there was no alliance), the presidential camp, Les Républicains and the UDI, the RN and its allies, Reconquête, and smaller groups (other left, other centre, other right, far left, others). The replay works at that level, and the seat engine then sums the blocs into its seven model blocs. The errors being replayed are shown below.
| Party | 2002 | 2007 | 2012 | 2017 | 2022 | 2024 | Average |
|---|---|---|---|---|---|---|---|
| Others | +1.5 | +0.3 | 0.0 | +1.4 | +2.7 | +1.6 | +1.3 |
| Other right | −0.7 | −1.3 | −0.6 | −2.1 | −2.2 | −2.1 | −1.5 |
| Other left | 0.0 | −0.7 | −1.3 | −1.1 | −0.5 | −0.5 | −0.7 |
| Far left | +0.7 | 0.0 | +0.4 | +0.1 | 0.0 | −0.1 | +0.2 |
| Republicans and UDI | −3.7 | −0.5 | +0.2 | +1.4 | −0.9 | +0.1 | −0.6 |
| New Popular Front (united left) | +0.6 | +0.4 | −0.5 | −0.4 | +0.8 | +0.4 | +0.2 |
| National Rally and allies | +2.1 | +1.0 | +1.8 | +4.3 | +0.5 | +2.9 | +2.1 |
| Other centre | −0.1 | +1.2 | +0.7 | – | – | −0.3 | +0.4 |
| Ensemble (Macron camp) | – | – | – | −1.8 | +1.0 | −0.6 | −0.5 |
| Reconquête | – | – | – | – | +1.6 | +0.7 | +1.1 |
Three features of this table drive the lenses. The RN and its allies were overstated at all three elections, by very different amounts. The "other right" (candidates of the right outside Les Républicains) was understated by a little over two points each time. And the presidential camp's error changed sign: understated in 2017 and 2024, overstated in 2022.
The legislative lenses
What each replay assumes
If the polls miss like 2017. The largest RN overstatement of the three (4.3 points), with the presidential camp and the other right understated. Replaying it takes the most off the RN of any lens and adds to Ensemble and the other right. Because the RN reaches round two in most circonscriptions and many of its duels are close, this is usually the lens that moves the most seats away from it.
If the polls miss like 2022. A small RN overstatement (0.5 points), the presidential camp overstated by a point, Reconquête overstated by 1.6 and the "others" by 2.7, the other right understated. This replay leaves the RN close to its published share and lowers Ensemble.
If the polls miss like 2024. An RN overstatement of 2.9 points and the same pattern on the other right and the others as the two earlier elections. It lies between the other two replays for the RN, and it is the most recent.
Each lens is then run through the deterministic version of the seat engine described in From votes to seats in the National Assembly: every circonscription starts from its 2024 round-one result, moves by the national change in each bloc, applies the qualification rules (the top two, plus anyone with 12.5% of registered voters, with an outright win for over half the votes and a quarter of the registered), withdraws the weaker opponents where the RN and two or more of them qualify, and decides each duel with the model fitted on the 403 duels of 2024. There is no random element in a lens: withdrawals happen wherever the rule says they should, and each duel goes to whichever finalist the model puts ahead.
Seats under each reading
| Reading | Left | Ensemble | LR | RN | Reconquête |
|---|---|---|---|---|---|
| The polls as they stand | 286 | 74 | 83 | 118 | 0 |
| Our projection | 291 | 72 | 90 | 108 | 0 |
| If the polls miss like 2002 | 278 | 81 | 104 | 99 | 0 |
| If the polls miss like 2007 | 291 | 68 | 94 | 108 | 0 |
| If the polls miss like 2012 | 319 | 58 | 87 | 97 | 0 |
| If the polls miss like 2017 | 318 | 83 | 88 | 72 | 0 |
| If the polls miss like 2022 | 287 | 67 | 95 | 113 | 0 |
| If the polls miss like 2024 | 300 | 79 | 95 | 88 | 0 |
The table shows the four largest blocs and Reconquête; the remaining seats go to the others and regionalists, whose seats move little between lenses. The first two rows differ only by our correction. The last three differ from the first only by one historical miss each.
Two patterns follow from the arithmetic rather than from today's particular polls, so they should hold as the polls change. First, the spread across the three replays is much wider than the gap between the polls and our projection. Our correction is half an average; each replay is a full, single miss, and the RN's three misses differ by almost four points. Second, the seat counts react to round-one shares far more than in proportion. A two-point change in the RN's round-one share alters which bloc leads round one in many circonscriptions, who qualifies in some, and, through the duel model, who wins the close duels. This is the same property that made pollsters' published seat ranges so unreliable at past elections, discussed below.
Each row is a central projection. For our projection the site also runs 1,000 simulations with national and local random errors and reports an 80% range for each bloc and the chance of a majority; today no bloc reaches 289 seats in 78% of those simulations. The lenses are not simulated, because the point of the comparison is to isolate the effect of one assumption.
Seats at risk
The count
For each circonscription \(i\) let \(w^{L}_i\) be the winner under lens \(L\), and let \(B\) be our projection. With \(\mathcal{P}\) the three past-miss lenses, the flip count is
A seat is at risk from polling error if \(F_i \ge 1\): replaying at least one past legislative miss on today's polls gives it to a different bloc from our projection. The code is LensMath.Flips, called from FranceLegislative.BuildLenses. Today 91 seats are at risk.
Only the three replays count. The "polls as they stand" lens does not, because the difference between it and our projection is our correction, not a historical error. There is no average-error lens and no pollster-choice lens in France, so nothing else could count.
What the count is not
\(F_i\) counts historical scenarios. It is not a probability. A seat that flips under one replay of three is not "a third likely" to change hands; the chance of each bloc winning each seat comes from the simulation and is shown separately. The two measures answer different questions. The simulation asks how often a seat changes hands if national and local errors of a typical size are drawn at random. The flip count asks whether the seat survives each of three specific, real misses. A seat can have a high simulated chance for its projected winner and still flip under the 2017 replay, because that replay is larger than a typical draw for the RN.
The three replays are also not independent of each other or of our projection. The projection is built from the same three errors, halved and averaged. A seat that flips under two or three replays is one where the outcome turns on the size of the RN's overstatement, which has varied between half a point and four points.
How the site shows it
The count and the per-seat flips appear in five places.
The map. The France map (hex layout by default, or the real boundaries) has a Risk from polling error colouring. A seat that keeps its winner under every past miss is pale grey; one that changes under one past miss is orange; one that changes under two or three is red. The tooltip names the seat, our projection's winner, and the years whose replay would change it ("A different winner if the polls miss like 2017, 2024"). The winner colouring can also be switched to any lens, so a reader can see the whole map under, say, the 2017 miss.
The seats list. The circonscriptions page has an At risk from polling error tab listing every seat with \(F_i \ge 1\), most at risk first (by flip count, then by our projection's simulated chance, lowest first). Each row shows the winner under each replay, marking only the ones that differ.
Each seat page. A table gives the winner under every lens (polls as they stand, our projection, and each replay), flags any that differ, and ends with "the same winner under every past miss" or "a different winner under \(k\) of 3 past misses".
The regions page. Each region's row counts its seats at risk alongside its projected seats and changes since 2024.
The front and Assembly pages. A chart shows each bloc's seats under the three replays as dots, the polls as they stand as a ring and our projection as a bar, with the at-risk count linked to the seats list.
The what-if tool and the prediction page
The what-if tool lets a reader set their own round-one error for each of the seven model blocs, from −8 to +8 points in half-point steps. The tool subtracts the reader's errors from the polls as they stand, applies the same 0.1-point floor and rescale, and sends the shares to the same seat engine. It reports the seats and lists every circonscription that changes hands compared with our projection. The three past misses are offered as presets. A preset sets each slider to the gap between the published polls and the replay's share for that model bloc, rounded to the nearest half point, so a preset's seat totals can differ by a seat or two from the lens table.1
The prediction page works the other way round: the reader types round-one shares directly, starting from the polls, our projection, any past miss or the 2024 result, and sees every circonscription worked out. It also takes presidential round-one shares and sends the top two to round two using our estimate for that match-up.
The presidential version
Replaying round-one misses on candidates
The presidential page carries a second set of lenses, built in FrancePresidential.BuildLenses. For each line-up the site follows (at present one with Marine Le Pen for the RN and one with Jordan Bardella instead), it starts from round one as polled, grouped by candidate family, and replays the round-one miss of each presidential election from 2002 to 2022 with the formula above. The families line up across years: Jean-Marie Le Pen, Marine Le Pen and the RN form one family, Nicolas Sarkozy, François Fillon and Valérie Pécresse another, and so on (Measuring French polling error, 1965–2024). The misses being replayed are these.
| Party | 1965 | 1969 | 1974 | 1981 | 1988 | 1995 | 2002 | 2007 | 2012 | 2017 | 2022 | Average |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Republicans | −1.6 | −4.0 | +0.9 | +1.5 | +3.6 | +2.4 | +0.2 | −2.8 | −0.7 | −0.7 | +3.1 | −0.2 |
| Socialists | −4.7 | +3.4 | +1.2 | −4.5 | +0.9 | −3.3 | +1.6 | −1.6 | −0.7 | +1.3 | +0.3 | +0.2 |
| Others | +1.9 | −0.3 | −0.8 | +0.3 | −0.1 | −0.1 | 0.0 | 0.0 | – | −0.4 | −0.6 | −0.3 |
| Far left | – | −0.1 | – | +0.2 | −0.9 | −0.1 | +0.6 | +1.3 | +0.1 | +0.3 | +0.3 | +0.5 |
| Centre (UDF, MoDem) | +4.4 | +4.7 | −2.1 | −0.8 | −0.1 | – | −0.6 | −0.7 | +1.1 | – | – | −0.1 |
| Communist Party | – | −3.8 | – | +3.6 | +0.2 | +1.0 | +1.9 | +0.6 | – | – | +0.3 | +0.9 |
| Greens | – | – | – | −0.4 | −1.3 | +0.2 | +0.8 | −0.2 | +0.1 | – | +0.1 | +0.2 |
| National Rally | – | – | – | – | −2.4 | −1.3 | −3.9 | +3.3 | −2.0 | +0.9 | −0.2 | −0.4 |
| Sovereigntists | – | – | – | – | – | +1.3 | – | −0.2 | −0.1 | −0.7 | +0.2 | −0.2 |
| France Unbowed | – | – | – | – | – | – | – | – | +2.5 | −0.4 | −4.8 | −0.9 |
| Hunting and rural party | – | – | – | – | – | – | −0.6 | +0.3 | – | – | – | −0.1 |
| Ensemble (Macron camp) | – | – | – | – | – | – | – | – | – | −0.1 | −1.5 | −0.8 |
| MNR | – | – | – | – | – | – | −0.1 | – | – | – | – | −0.1 |
| Reconquête | – | – | – | – | – | – | – | – | – | – | +2.7 | +2.7 |
Each presidential lens reports the candidates' shares, the top two, and, if that pair has been polled head to head, who wins round two and with what share. The round-two figure is our corrected head-to-head average for the pair, the same one used in the projection (it takes about 1.8 points off the RN candidate, as explained in The 2027 presidential forecast: line-ups, runoffs and simulation). If the pair has never been polled, the lens names the top two but no winner.
What the replays do
The round-one replays mostly move the RN candidate and Mélenchon, because those are the families with the largest single misses.
- 2002 raises the RN candidate by about four points: the polls understated Jean-Marie Le Pen by 3.9, which is how he reached round two.
- 2007 does the opposite, lowering the RN candidate by 3.3 and raising the Republican family by 2.8.
- 2012 lowers Mélenchon's family by 2.5 and raises the RN by 2.0.
- 2017 was the most accurate round one in the record (0.60 points per family on average) and moves little.
- 2022 raises Mélenchon's family by 4.8, lowers the Republicans by 3.1 and Reconquête by 2.7, and raises the presidential camp by 1.5.
So the 2022 replay is the one most likely to change who reaches round two when Mélenchon is close behind the second-placed candidate, and the 2002 and 2007 replays are the ones that move the RN candidate furthest. Whether any replay changes the pairing depends on today's gaps, which the presidential page shows.
Why there is no round-two lens and no "at risk" count
We do not replay round-two misses. Round-two polls are already corrected in the projection, because the RN candidate was overstated at every runoff the RN has reached and the correction passed its test at four of five runoffs. A replay of a single runoff would repeat most of that correction at full size.
Nor is there a presidential equivalent of seats at risk. The presidency is one contest, and its uncertainty is already expressed as chances from 20,000 simulated elections in which each family's round-one share moves by a random error the size of its typical final-poll miss. The round-one lenses draw on the same history as that simulation. They add something the chances cannot: a named, reproducible picture of a real past miss applied to today's field.
Why round one is shown but not corrected
For round one the leave-one-out test is plain: correcting by half the average miss was further out at 4 of the five elections. The misses change direction from one election to the next, so their average is close to zero for most families and correcting by it adds noise. The replays show the size of those misses without pretending to know their direction this time. That is why the presidential page says of them: "We don't use these for our projection, because the misses change direction; they show how far things could move."
Why the lenses are not averaged
It would be possible to average the three legislative replays, or to weight them, and report one seat total. We do not, for four reasons.
First, averaging the replays' shares is itself a correction: the full mean past miss, twice the size of the one that passed our test. Our projection already uses half of it. There is no evidence that the full version does better, and taking the whole of a past miss as a forecast is the assumption the British forward test found wanting (Correcting the raw answers: the method behind our UK projection).
Second, averaging seat totals is not the same as the seat total of averaged shares. The French seat engine is strongly non-linear. Small changes in round-one shares change who qualifies, who withdraws and who leads each duel, so the mean of three seat totals can be a figure that no coherent set of round-one shares would produce.
Third, the replays are not draws from a known distribution. With three elections there is no basis for giving each a probability, and the RN's misses (4.3, 0.5, 2.9) do not suggest a stable size.
Fourth, the spread is the information. A reader who sees that one past miss would cut the RN's seats sharply while another would leave them close to where the published polls put them has learned how fragile the projection is. An average would hide exactly that.
Where we want chances, we use the simulation. The lenses do a different job.
What the lenses cannot show
The limitations of the lenses are those of the evidence behind them, plus one that matters more in France than in Britain.
Only three legislative elections. The replays can only reproduce misses that have happened, and three is a small catalogue. The bloc definitions also moved between them: in 2017 there was no left alliance, Reconquête did not stand until 2022, and the RN's allies in 2024 included the Ciotti wing of Les Républicains. Each replay applies its own year's error to today's blocs as if the blocs were the same.
The same error everywhere. A lens changes only national shares, and uniform swing spreads the change evenly across circonscriptions. No lens can express a miss that falls unevenly, for instance polls that overstate the RN mainly where it is already strong.
Round two is held fixed. This is the main gap. Every lens changes round-one shares only. Withdrawals follow the 2024 rule and duels follow the 2024 model in every lens. Yet the largest French forecasting misses of the last decade came from round two, not round one. Between the rounds, pollsters publish seat projections. At each of the last three elections the bloc that led round one won fewer seats than the pollsters' average range, and in 2024 the RN won fewer than every pollster's range:
| Election | Bloc | Projected (average range) | Won | Ranges that held it |
|---|---|---|---|---|
| 2017 | Ensemble (Macron camp) | 437–467 | 350 | 0/3 |
| 2017 | Republicans and UDI | 67–88 | 130 | 0/3 |
| 2017 | Left | 31–52 | 58 | 1/3 |
| 2017 | National Rally and allies | 1–6 | 8 | 0/3 |
| 2017 | Others | 3–8 | 4 | 3/3 |
| 2022 | Ensemble (Macron camp) | 258–298 | 245 | 1/7 |
| 2022 | Left (NUPES) | 161–206 | 131 | 0/7 |
| 2022 | National Rally and allies | 26–48 | 89 | 0/7 |
| 2022 | Republicans and UDI | 54–72 | 64 | 5/7 |
| 2022 | Other left | 6–12 | 22 | 1/6 |
| 2022 | Other right | 3–7 | 11 | 0/5 |
| 2022 | Other centre | 2–4 | 4 | 3/4 |
| 2022 | Others | 5–10 | 1 | 0/7 |
| 2024 | New Popular Front (united left) | 155–185 | 178 | 5/7 |
| 2024 | Ensemble (Macron camp) | 121–151 | 156 | 2/7 |
| 2024 | National Rally and allies | 188–221 | 142 | 0/7 |
| 2024 | Republicans and UDI | 40–58 | 42 | 5/7 |
| 2024 | Other left | 12–17 | 14 | 2/2 |
| 2024 | Other centre | 6–8 | 6 | 1/1 |
| 2024 | Others | 10–19 | 3 | 0/7 |
In 2024 the RN and its allies were projected at 188 to 221 seats on average and won 142, after round-one polls that had overstated them by only 2.9 points. Most of that gap came from withdrawals and from how the voters of eliminated candidates split. The duel model itself shows how much this can change: rates fitted on 2022 pick 64.5% of the 2024 duel winners, against 61.0% for simply taking the round-one leader, while rates fitted on 2024 pick 370 of the 403 2024 duels. A lens that replayed round-two behaviour would need more than two elections of duel data to be worth showing. Until then, readers should treat the lenses as a lower bound on how far the seat count could move if the polls are wrong.
Stale polls. The legislative lenses are applied to the last legislative voting-intention polls, taken in autumn 2025. If an election is called, the lenses will be recomputed from new polls; until then they describe an Assembly election held on those polls. What we don't know yet: limitations of the French forecast discusses this and the other open questions.
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.
Ministère de l'Intérieur. Résultats des élections législatives 2017, 2022 et 2024 et de l'élection présidentielle 2002–2022, data.gouv.fr.
-
The what-if presets are built from the lenses' model-bloc shares, not from the bloc-level errors. The sliders therefore show the net effect of each replay on each model bloc after the rescale, which can differ slightly from the raw historical error: the "LR and other right" slider, for example, combines Les Républicains' error with the other right's.↩
How to cite
Polling Forecast (2026). If we're wrong: replaying past polling misses on today's French polls. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/france-lenses. Live figures change as new polls arrive; the date above is when the text was last revised.



