United Kingdom General electionby 15 Aug 2029in 999 days Majority for any party 34% Labour largest party 93% Calendar Track record Research
Research · United Kingdom · The method

Why we start from British pollsters' raw answers rather than the published headline

Abstract. Correcting the pollsters' published figures by their own past miss has never beaten the uncorrected figures when each election is predicted from earlier ones, because the headline's error swings from election to election. The raw answers' error is larger but steadier, so it can be partly removed. Starting from the raw answers gave the best record of the variants we tested, but no variant, ours included, would have foreseen how far the polls overstated Labour in 2024.

Polling Forecast · updated 9 November 2026 · 16 min read · raw answers, headline, polling error, predictability, 2024, Labour, turnout

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.

  • Names on the site (3 and 8 October). "Our reading" is shown as our adjusted average, and what this paper calls the best bet is our projection.
  • Our vote shares beside the polls (26 October). Wherever our projected shares differ from the pollsters' headline average, both are shown, so the effect of starting from the raw answers is visible on the hub, the poll pages and the leader cards.

The challenge

The site's premise is that the polls will be wrong, and that some of the error can be predicted. A fair objection follows from that premise. Pollsters spend a great deal of effort on their turnout models and on deciding what to do with people who don't know. Those adjustments are informed by their own post-mortems of past misses. If our aim is to predict how wrong the published polls will be, why discard the pollsters' adjustments and begin from the answers before them? Surely the natural approach is to take the published headline and correct it by how wrong headlines have been.

This paper answers that objection with the evidence we have. It sets out the two ways of predicting a polling error, shows formally how they relate, explains why the predictability of an error matters more than its size, and reports a head-to-head test of the two starting points on the same final polls. It then states the method's weakest point plainly: in 2024 the polls overstated Labour in England by a wide margin, no variant we tested would have predicted it from the elections before, and on Labour alone our method has sometimes been further out than the pollsters. The estimator itself is set out in Correcting the raw answers: the method behind our UK projection; this paper takes it as given and asks whether its starting point is the right one.

Two ways to predict a polling error

We use the notation of Correcting the raw answers: the method behind our UK projection. For party \(p\) at election \(t\), \(\bar h_{p,t}\) is the average published headline of the final polls in England, \(\bar r_{p,t}\) is the average of the same polls' raw answers (the weighted table of everyone asked, before the pollster's turnout and don't-know steps, parties only), and \(v_{p,t}\) is the result. The two errors are

\[ e^{\,head}_{p,t} = \bar h_{p,t} - v_{p,t}, \qquad e^{\,raw}_{p,t} = \bar r_{p,t} - v_{p,t}. \]

A forecaster who wants to predict the result from today's polls can start from either figure and subtract a share \(\lambda\) of that figure's average past error:

\[ \hat v^{\,head}_p = h_p - \lambda \, \bar e^{\,head}_p, \qquad \hat v^{\,raw}_p = r_p - \lambda \, \bar e^{\,raw}_p. \tag{1} \]

The first predicts the headline's error; the second predicts the raw answers' error. They are not as different as they look. Define the pollster's adjustment as the gap its own steps open between raw answers and headline, \(A_p = h_p - r_p\), with historical average \(\bar A_p\). Since the result is the same for both errors, \(\bar e^{\,raw}_p = \bar e^{\,head}_p - \bar A_p\). Substituting into (1),

\[ \hat v^{\,raw}_p = \underbrace{\left(h_p - \lambda\,\bar e^{\,head}_p\right)}_{\hat v^{\,head}_p} \;-\; \left(A_p - \lambda \bar A_p\right). \tag{2} \]

This identity is the heart of the matter. Starting from the raw answers is the same as starting from the headline, removing the same share of the headline's past miss, and then replacing the pollster's adjustment today, \(A_p\), with \(\lambda\) times the average adjustment of the past. With full correction (\(\lambda = 1\)), the two estimators differ by exactly \(A_p - \bar A_p\): how far today's adjustment departs from the historical norm.

So the question "why throw away the pollsters' adjustments?" has a precise form. We do not throw them away; we replace each pollster's current adjustment with a scaled historical average of such adjustments. That is the better choice if today's adjustment carries less information about the coming error than the historical average does, and the worse choice if the pollsters' adjustments track the changing error from election to election. Which of these is true is an empirical question, and the rest of this paper answers it with data.1

Why predictability matters more than size

A simple model

Suppose each election's error, from whichever starting point, is a persistent part plus a part specific to that election: \(e_{p,t} = \mu_p + \varepsilon_{p,t}\), where \(\varepsilon_{p,t}\) has mean zero and variance \(\sigma_p^2\) and is independent across elections. If we remove \(\lambda\) of the average of \(n\) past errors, the expected squared error at the next election is

\[ \mathbb{E}\!\left[(e_{p,\text{next}} - \lambda\,\bar e_p)^2\right] = (1-\lambda)^2\mu_p^2 + \sigma_p^2\left(1 + \frac{\lambda^2}{n}\right). \tag{3} \]

The first term is the part of the persistent bias left in; a correction can shrink it towards zero. The second cannot be removed by any correction based on the past: it is the election-specific error, slightly inflated by the noise in our estimate of \(\mu_p\). So a large but stable error is a good starting point, because most of it can be taken out, while a small but swinging error is a poor one, because it cannot. Which starting point is better depends less on which one has the smaller average miss than on which one has the smaller \(\sigma_p\).

The distinction between error that is shared across an election and error that varies is familiar in the study of polls. Shirani-Mehr et al. (2018) separate the error of election polls into bias common to an election and variance, and find the election-level component large. Jennings and Wlezien (2018) find that polling error, across many countries and decades, has not been getting worse, but that it has not gone away either. Our concern is narrower: whether the error that remains in England has a part stable enough across elections to correct.

What the errors look like

The first table gives the published final polls' errors in England, from the primary benchmark (every checked final poll, see Measuring British polling error in England, 2015–2024).

ElectionFinal pollsConLabLDReformGreenLab lead
20152−5.7+1.5+1.8+0.8+0.9+7.1
20174+0.6−3.9+0.7+2.1+0.6−4.5
20195−1.6+0.9−0.6+0.8+0.1+2.5
20246−3.5+4.3−0.9+2.0−0.6+7.8
Mean−2.6+0.7+0.3+1.4+0.2+3.3
Root mean square3.53.01.11.50.6
Live Published final polls minus the result, England, points (positive = the polls overstated the party). UKIP and the Brexit Party count as Reform. Computed from the database on 19 November 2026.

The second gives the raw answers' errors for the final polls whose raw tables survive, with the published headline of the same polls in brackets.

ElectionPolls with tablesConLabLDReformGreen
20152−6.6 (−5.7)+1.6 (+1.5)+2.0 (+1.8)+0.9 (+0.8)+1.1 (+0.9)
20174−2.1 (+0.6)−1.5 (−3.9)+0.7 (+0.7)+2.1 (+2.1)+1.0 (+0.6)
20196−3.3 (−1.6)+1.8 (+1.0)−0.5 (−0.5)+1.1 (+0.8)+0.4 (0.0)
20246−4.0 (−3.0)+6.1 (+4.8)−1.7 (−1.3)+1.0 (+1.3)−0.4 (−0.6)
Mean−4.0 (−2.4)+2.0 (+0.9)+0.1 (+0.2)+1.3 (+1.3)+0.5 (+0.2)
Live Raw answers (weighted, everyone asked, don't knows removed) minus the result, England, points; in brackets the published headline of the same polls. Computed from the database on 19 November 2026.

Read the Conservative column of each. The raw answers understated the Conservatives at every election. The published headlines understated them at three and overstated them in 2017, when the pollsters' turnout models moved their figures towards the Conservatives by more than the raw answers' shortfall. Labour's headline error changes sign twice, from an overstatement in 2015 to a large understatement in 2017 and back to a large overstatement in 2024. The Reform family was overstated at every election on both measures.

A summary of the same eighteen polls makes the point numerically. The table below gives the mean and the standard deviation across the four elections of each party's error, from each starting point. These are historical figures, computed once from the polls in the raw-answer table.

Party Raw: mean Raw: s.d. Headline: mean Headline: s.d.
Conservative −4.0 1.9 −2.4 2.6
Labour +2.0 3.1 +0.9 3.6
Liberal Democrat +0.1 1.6 +0.2 1.4
Reform family +1.3 0.6 +1.3 0.6
Green +0.5 0.7 +0.2 0.7

Table: Mean and standard deviation across 2015–2024 of the England error of the same final polls, raw answers against published headline. Points.

For the two largest parties, the raw answers miss by more on average and vary less from election to election. In the terms of (3), they have a larger \(\mu_p\) and a smaller \(\sigma_p\): exactly the combination a correction can exploit. For the Reform family and the Greens there is little to choose; for the Liberal Democrats the headline is slightly steadier. Four elections give only three degrees of freedom for each standard deviation, so these figures are indicative, not precise. But the direction is consistent with the tests below.

The head-to-head test

We compared eight ways of predicting each election's England result from its final polls, all on the same polls, all scored by the mean absolute error across the five main parties. Each is tested two ways. In the leave-one-out test (Stone, 1974), each election is predicted using a correction estimated from the other three. In the forward test, each election from 2017 is predicted from the elections before it only, which is how a forecast is really made. The forward means cover three elections and the leave-one-out means four, so the two columns should be read separately rather than compared with each other.

ApproachLeave-one-outForwardLead error (LOO)Lead error (fwd)Labour 2024 (fwd)
Published figures, no correction (the pollsters)1.671.525.55.0+4.8
Published figures minus ½ their past miss1.571.575.35.6+5.1
Published figures minus all their past miss1.681.985.66.3+5.3
Raw answers, no correction1.991.846.05.2+6.1
Raw answers minus ½ their past miss (ours)1.471.424.64.7+5.7
Raw answers minus ¾ their past miss1.341.434.14.7+5.6
Raw answers minus all their past miss1.461.694.24.6+5.4
Half and half, each minus ½ its past miss1.511.505.05.2+5.4
Live Starting from the published headline against the raw answers. Mean absolute error per party (England, points) in both tests, the mean absolute error in the Labour lead, and Labour's own error at the last election in the forward test. Computed from the database on 19 November 2026.

The rows are worth taking one at a time.

Published figures, no correction. This is what the pollsters publish and what most forecasts use. It is the benchmark.

Published figures minus half their past miss. In the leave-one-out test this improves on the pollsters, but in the forward test it is worse than leaving the headlines alone. Leave-one-out lets the 2024 miss inform the prediction of 2019 and the 2017 miss inform 2024; when only earlier elections are available, the headline's past errors point the wrong way often enough to hurt.

Published figures minus all their past miss. This is the most natural reading of "predict how wrong the polls will be", and it is the weakest serious candidate. In the leave-one-out test it is no better than the uncorrected headlines, and in the forward test it is the worst row in the table. This is (3) in action: the headline's error swings, so correcting it fully mostly adds noise.

Raw answers, no correction. The raw answers on their own are clearly worse than the headlines. This matters for the argument. The pollsters' adjustments are not useless; on average they move the figures in the right direction. What the raw-answer approach does is replace them with something better on average, not ignore what they achieve.

Raw answers minus half their past miss (ours). This has the lowest forward error of any row, a leave-one-out error among the lowest (full correction and three-quarters are level with it or lower there), and it cuts the error in the Labour lead in both tests.

Raw answers minus three-quarters. This has the lowest leave-one-out error in the table and a forward error level with ours. We return to it in the final section.

Raw answers minus all their past miss. Good in the leave-one-out test, but worse than the pollsters in the forward test. Full correction overfits: it is more exposed to the one election, 2017, in which the raw-answer pattern was weakest.

Half and half. A 50/50 blend of the two starting points, each corrected by half its past miss, sits between them in both tests. It is better than any headline-based row in the forward test but not as good as starting from the raw answers alone.

The ordering is the same in both tests on the point that matters: starting from the raw answers and removing a moderate share of their past miss beats every way of correcting the headlines. Correcting the headline by its own past miss never beat the uncorrected headline in the forward test, at any share. The leave-one-out test for our preferred variant is reported election by election in Correcting the raw answers: the method behind our UK projection: an average miss per party of 1.47 points against the pollsters' 1.67, and 1.68 for the headlines corrected by their own past miss.

The forward test and the last election

A natural refinement is to weight recent elections more heavily, on the grounds that polling methods change. The extreme version assumes that the next election's error will repeat the last one's. The forward test compares that with our method, which averages every earlier election equally.

ElectionPollstersOurs: ½ × mean of all earlier½ × previous electionAll of previous election
2017 from 20151.58 (−3.9)1.17 (−2.3)1.17 (−2.3)2.03 (−3.1)
2019 from 2015, 20170.79 (+1.0)0.90 (+1.8)1.15 (+2.6)1.44 (+3.3)
2024 from 2015, 2017, 20192.19 (+4.8)2.18 (+5.7)2.00 (+5.2)1.43 (+4.2)
Mean1.521.421.441.64
Live Forward test: each election predicted from the elections before it only. Mean absolute error per party, England, points; Labour's own error in brackets. Computed from the database on 19 November 2026.

Assuming a repeat of the previous election's raw-answer miss is worse than believing the pollsters: it would have carried 2017's peculiarities into 2019 and 2019's into 2024. Taking half of the previous election's miss does better, and taking half of the average of all earlier elections does best, though the margin between those two is small. Our method averaged 1.42 points per party in this test, against the pollsters' 1.52.

The figures in brackets are Labour's own error, and they tell a less comfortable story. In the forward test our method was closer to Labour's result than the pollsters in 2017, but further from it in 2019 and in 2024. (In the leave-one-out test the pattern is milder: our method was further out on Labour only in 2024.)

The 2024 Labour problem

In England in 2024, the published final polls overstated Labour by +4.3 points and the raw answers overstated it by +6.1.2 The error in the Labour lead over the Conservatives was the largest in our data. Our correction changes Labour's raw share by −1.0 points; the 2024 miss was several times that.

No variant in the head-to-head table predicts this from the elections before it. The column headed "Labour 2024 (fwd)" shows Labour's error in 2024 when each method is trained on 2015, 2017 and 2019 only, and every correction variant leaves Labour roughly five to six points too high, further out than the uncorrected headline. The reason is simple: 2015 to 2019 gave no warning. Before 2024, the published headlines' Labour error averaged about half a point below the result, pulled down by the large understatement of 2017. A method that learns from past headline errors would have added to Labour in 2024. The raw answers' Labour error before 2024 was positive but small and inconsistent, so a method that learns from them took off only a fraction of a point. The pollsters' headline, which takes off more than that through the pollsters' own adjustments, did best on Labour in 2024 of all the approaches, and still missed by more than four points.

The pollsters' adjustments deserve credit here, and this is the strongest version of the objection this paper answers. In the raw-answer table above, comparing each raw figure with the bracketed headline shows that the pollsters' own turnout and don't-know steps lowered Labour at every one of the four elections: by about 0.1 points in 2015, 2.4 in 2017, 0.8 in 2019 and 1.2 in 2024, on average across the polls. That direction was right at three of the four elections, since Labour was overstated in 2015, 2019 and 2024. It was wrong in 2017, when Labour was already understated before the adjustments and the adjustments made it worse. Our method sets those adjustments aside and replaces them with the historical average, as (2) shows. For Labour, that average is dragged towards zero by 2017. So on Labour specifically, our method gives up the pollsters' tendency to mark Labour down, which was right more often than not, in exchange for the gains on the Conservatives and the Reform family, which were larger and steadier.

We state this openly on the site: on Labour alone, our method has been further out than the pollsters at recent elections, and today's gap between our reading and the published average on Labour is smaller than a 2024-style miss would imply. The table of lenses below puts the question in seats. The row for the 2024 miss applies the published polls' 2024 errors to today's polls (England's error in England, and each nation's own in Scotland and Wales); the row for our projection applies our correction.

LensGroupLabConReformLDSeats changed
Our projectionOur method3171543672–
Polls as they standPolls329143377214
Average missPast misses319167327023
2015 missPast misses279210306060
2017 missPast misses395117137285
2019 missPast misses31815436724
2024 missPast misses273200427070
Pollster recordsPollsters308169327016
Best for LabourPollsters347134347234
Best for the ConservativesPollsters325150367211
Best for Reform UKPollsters318136517325
Live Every lens through the same seat model (uniform swing). Seats changed: seats with a different winner from our projection. Computed from the database on 19 November 2026.

What would catch a 2024-style miss

A correction learned from the size of past errors can only foresee a miss that past errors foreshadowed. The 2024 Labour overstatement was not foreshadowed in 2015–2019. To catch the next one, a model would need a view of why polls overstate a party, measured before the election, rather than a record of how much they did.

Three mechanisms are candidates, and all three have been discussed in the analysis of past British polling misses (Sturgis et al., 2016; Mellon and Prosser, 2017; Prosser and Mellon, 2018).

Differential turnout. If one party's supporters are less likely to vote than they say, polls that take stated likelihood at face value overstate that party. Pollsters' turnout models exist to address this, and the fact that their adjustments lowered Labour at every election suggests they have been partly successful. A model of our own would need turnout data validated against what people actually did, not what they said.

Where don't knows go. Respondents who don't know at the time of the poll do not divide on polling day in the same proportions as those who have decided. We tested one simple version, returning don't knows to the party they voted for at the last election, and it did not improve out of sample (see Correcting the raw answers: the method behind our UK projection). A better model might use the squeeze questions that several pollsters print and that we already store.

Who is in the sample. Samples with too many politically engaged people misrepresent the electorate even after weighting (Mellon and Prosser, 2017). The sample skew we measure for each pollster (Rating the British pollsters: accuracy, the Labour–Conservative lead and openness) is a first descriptive step, but it is not yet linked to the forecast.

The natural evidence for all three is the British Election Study, whose internet panel follows the same respondents across elections and which has validated reported turnout for some of its samples against official records. None of this is built. It would be a model of a different kind from the one described here, and it would need to pass the same leave-one-out and forward tests before it replaced anything.

The three-quarters option

One variant in the table would take more off Labour than ours does: raw answers minus three-quarters of their past miss. It has the lowest leave-one-out error of all eight rows, a forward error level with ours, and a smaller error in the Labour lead in the leave-one-out test. On the average alone it is the better choice.

It was not chosen because of the rule, set before the tests, that a correction must improve consistently across elections, not only on average. Moving from a half to three-quarters helps in three elections but makes 2017 noticeably worse, to the point where it is only level with the pollsters there. We should be clear about how much would change if we chose it instead. On Labour in 2024, the forward test shows three-quarters would have left the party only about a tenth of a point closer to the result than a half. It is not a remedy for the 2024 problem; its gains are on the Conservatives. It remains an open choice, and one we will revisit if the evidence base grows, for instance if the 2010 final polls, which the project plans to bring forward, can be read into the same stage model.

Conclusion

The objection that started this paper is a good one, and the answer is not that the pollsters' adjustments are worthless. On their own, the raw answers are worse than the headlines. The answer is that the adjustments vary from election to election in ways that make the headline's error hard to predict, and a polling error that cannot be predicted cannot be corrected. Formally, our method keeps the headline and replaces each pollster's current adjustment with a scaled historical average; the tests show that this replacement has been the better bet. The raw answers have missed by more but more steadily, understating the Conservatives and overstating the Reform family at every election from 2015 to 2024, and removing half of that average miss beat every way of correcting the headlines, both when each election was predicted from the other three and when it was predicted from the elections before it only.

That is the predictable part of polling error, and it is what our projection corrects. It is not the whole of polling error. The 2024 overstatement of Labour was not predictable from the three elections before it, by our method or any other we tested, and the pollsters' own tendency to mark Labour down, which we set aside, was right more often than not. We therefore show the published polls beside our projection, show what a repeat of each past miss would do in seats, and carry the full spread of past errors into the ranges. Catching a 2024-style miss in advance would need a model of why polls go wrong, not just a record of how far. We have not built one yet, and until we have, we say so.

References

British Election Study (Fieldhouse, E. et al.). British Election Study Internet Panel. www.britishelectionstudy.com

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

Mellon, J. and Prosser, C. (2017). Missing nonvoters and misweighted samples: explaining the 2015 great British polling miss. Public Opinion Quarterly, 81(3), 661–687.

Prosser, C. and Mellon, J. (2018). The twilight of the polls? A review of trends in polling accuracy and the causes of polling misses. Government and Opposition, 53(4), 757–790.

Shirani-Mehr, H., Rothschild, D., Goel, S. and Gelman, A. (2018). Disentangling bias and variance in election polls. Journal of the American Statistical Association, 113(522), 607–614.

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

Sturgis, P., Baker, N., Callegaro, M., Fisher, S., Green, J., Jennings, W., Kuha, J., Lauderdale, B. and Smith, P. (2016). Report of the Inquiry into the 2015 British General Election Opinion Polls. London: Market Research Society and British Polling Council.


  1. The identity holds for averages over the same set of polls. Today's pollsters are not the same as the historical ones, so in practice \(A_p\) is the adjustment of today's pollsters and \(\bar A_p\) that of the historical set. For a pollster whose raw answers and headline coincide on the parties (Opinium today), \(A_p = 0\) and our method applies the historical correction in full; for one with large adjustments, it undoes them first. The step-by-step table in Correcting the raw answers: the method behind our UK projection shows both effects separately.↩

  2. The two figures come from slightly different sets of 2024 polls. The published-polls figure is the primary benchmark, which includes YouGov's final call; the raw-answer figure uses the six 2024 polls whose raw tables we hold, which include Number Cruncher (a sensitivity-only poll in the primary benchmark) and not YouGov. On the raw-answer set, the headline of the same polls overstated Labour by a little more than the benchmark figure, as the bracketed values in the raw-answer table show.↩

How to cite

Polling Forecast (2026). Why we start from British pollsters' raw answers rather than the published headline. Retrieved 20 November 2026 from https://www.pollingforecast.com/research/uk-raw-or-headline. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like