Measuring British polling error in England, 2015–2024
Abstract. Final polls at the last four general elections understated the Conservatives in England at three of them and overstated UKIP, the Brexit Party and Reform at all four, while Labour's error changed sign. We define the error, show how the evidence was chosen and graded, and explain why the size of past misses, not their average direction, sets our uncertainty bands.
Every forecast built on opinion polls rests on an assumption about how far the polls can be trusted. Most forecasters make that assumption implicitly, by treating the published average as the best estimate and adding a margin of error borrowed from sampling theory. We prefer to measure it. This paper sets out how we measure polling error: which polls count, how the England figure inside each poll is obtained, how the error is defined, and what the record from 2015 to 2024 shows. Later papers build on it: From answers to headlines: the stage model of a published British poll looks inside each poll, Correcting the raw answers: the method behind our UK projection uses the pattern described here to correct today's polls, and Simulating UK polling error: where the chances come from turns it into seat ranges.
The findings are simple to state. At the four general elections from 2015 to 2024, the final polls in England understated the Conservatives at three and overstated the UKIP, Brexit Party and Reform family at all four. Labour was overstated at three but understated in 2017, and so the error in the Labour lead over the Conservatives, which is what decides most seats, swung from large and positive to large and negative and back. Four elections is a small sample, and much of what follows is about what can and cannot be concluded from it.
Why measure error in England
The site forecasts the whole of Great Britain: vote shares for each party and the winner of each of the 632 seats in England, Scotland and Wales. Yet polling error is measured in England alone. There are two reasons.
The first is that England is where the parties compete on the same terms. The Scottish National Party and Plaid Cymru stand only in Scotland and Wales, and their presence changes what a Great Britain figure means. A GB poll that overstates the SNP by two points in Scotland shifts every other party's GB share slightly, and the size of that effect depends on how many Scottish respondents the poll happened to have and how it weighted them. Measuring against England removes that distortion. Each party's England error is a statement about how English voters were polled, not a blend of three party systems.
The second is weight. England holds 543 of the 632 GB seats, so an error in England matters far more to the seat count than the same error elsewhere. Scotland and Wales still need their own error records, and they have them, measured from Scotland-only and Wales-only polls (see Scotland and Wales: sparse polls and no steady lean). But the core evidence, and the pattern that the rest of the method leans on, comes from England.
The GB forecast is then assembled from the three nations. England uses England figures with the England error record; Scotland and Wales use their own polls and records; Great Britain is the three combined. Northern Ireland is outside the forecast.
Definitions
The error
Let \(v_{p,t}\) be party \(p\)'s share of the valid vote in England at election \(t\), and let \(\hat{v}_{i,p,t}\) be the share given to that party by poll \(i\), one of the \(n_t\) final polls that meet the evidence rules for that election. The error of a single poll is
The industry error at election \(t\) is the error of the simple average of the qualifying polls,
Because the average is linear, \(e_{p,t}\) is also the mean of the individual poll errors. Each pollster counts once, however large its sample, so the industry error describes what a reader who averaged the final polls would have seen.
The sign convention is used everywhere on the site without exception: error = poll minus result. A positive error means the polls overstated the party; a negative error means they understated it. When we say the Conservatives were understated by 3.5 points, the Conservative error was \(-3.5\).
The Labour lead error
Most English seats are decided between Labour and the Conservatives, so the error in the gap between them is a useful single number. Define the lead as Labour minus Conservative. Its error is
A positive lead error means the polls put Labour too far ahead of the Conservatives (or not far enough behind). The lead error is roughly twice as volatile as either party's error on its own, because the two errors usually have opposite signs: when one main party is understated the other tends to be overstated.
Summary measures
For each party we summarise its \(T\) election-level errors in two ways. The mean error measures direction:
The root mean square (RMS) error measures size, whatever the direction:
When a single number is needed to compare methods across parties, we use the mean absolute error over the five main parties in England (Conservative, Labour, Liberal Democrat, Reform and Green), \(\text{MAE}_t = \tfrac{1}{5}\sum_p |e_{p,t}|\). That was agreed as the primary metric at the start of the project, with the Labour–Conservative lead error, the total absolute error and the RMS as secondary measures.
Party families and "Other"
UKIP in 2015 and 2017, the Brexit Party in 2019 and Reform UK in 2024 are treated as one party family for measuring error: their errors are pooled into a single series labelled Reform. The raw data keep the three parties separate, and in the first phase of the project both a strict version (each party learns only from its own history) and a family version were tested. The family version is used because the three parties occupied the same place in English politics, drew on overlapping voters, and posed the same problem for pollsters: a challenger on the right whose supporters are harder than average to reach and to model. It is a modelling choice, not a fact about the parties, and the evidence that it adds predictive value is thin (see the limits below).
Restore Britain, which appears in polls from 2026, is not part of that family. It has no election history, so it has no measured error and no calibrated band.
The result used for comparison excludes the Speaker's seat from both the numerator and the denominator, since the main parties do not stand there. "Other" is defined by rule in code, as everything outside the named parties, rather than by hand. The England results come from the UK Parliament's constituency results, and a load check confirms that the constituency totals reproduce the national figures.
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.
- The track record across countries (28 September). Final polls at 133 elections in nine countries are measured the same way at /accuracy/elections: the average absolute error per party of the final-poll average. Britain's own record stays at /accuracy.
- When the polls missed (2 November). Shy Tories and a late swing: how the 1992 polls handed Kinnock a lead he never had, The hung parliament that never came: unrepresentative samples and the 2015 polling miss, Fixing 2015 too hard: turnout models and the majority that vanished in 2017 and Right winner, wrong size: how every poll overstated Labour's lead in 2024 tell the story of each British miss, with what the inquiries found.
- Polls without tables (26 October). A GB poll we do not hold is now taken from Wikipedia's list, graded C and counted in the GB average only. It never enters the England benchmark described here, and it is upgraded in place, with its England figures, when the pollster's own tables arrive. The benchmark in this paper is unchanged.
The evidence: final polls and how they were chosen
What counts as a final poll
A final poll is the poll from a given pollster whose fieldwork finished latest before polls opened, provided it was published before voting began. The publication date only breaks ties. Polls whose results were withheld until after the election, such as Survation's 2015 telephone poll, are stored with a flag (PostElectionDisclosure) and left out of every real-time benchmark, because no reader could have seen them. The primary benchmark uses one final poll per pollster; a second benchmark using every poll from the final week is planned as a cross-check.
In total 58 final polls are recorded for the four elections (12 in 2015, 11 in 2017, 13 in 2019 and 22 in 2024). Most cannot be used. The binding constraint is not the headline, which is easy to find, but a trustworthy England figure on the same basis as the headline.
Grading the source
Every figure carries a source grade. Grade A is the pollster's own published table. Grade B is an archived copy of that table (for example on archive.org), which counts as primary evidence, with both the original and archive addresses recorded. Grade C is a reputable secondary transcription that cites the original, such as Wikipedia. Grade D is a press report with no table. Only A and B enter the primary benchmark. Grade C figures may be used in sensitivity analysis and enter the benchmark only once they have been checked against the original.
The basis of the England figure
A poll's tables often show more than one set of vote-intention figures: everyone asked, only those likely to vote, before or after don't knows are reallocated. The England figure must be on the same basis as the published headline, or the comparison mixes two different quantities. Each poll records its basis (HeadlineEquivalent, PreDKReallocation, PreTurnoutFilter, AllRespondents, LikelyVoters or Unknown), and only HeadlineEquivalent figures enter the primary series. England figures on another basis, such as ICM's and Lord Ashcroft's in 2015, are stored as a basis mismatch and used only for sensitivity.
Direct and reconstructed England figures
Where the pollster prints an England column, the figure is taken directly (EnglandDataType = Direct). Where it prints only regions, England is reconstructed from them (Reconstructed). A reconstruction is accepted as exact only if every English region is present, Wales, Scotland and Northern Ireland are separate, the regional figures are on the same basis as the headline, weighted bases are printed, and the regions are combined using those weighted bases. Coarse regions such as North, Midlands, London and South are acceptable on those terms. A reconstruction that has to rely on all-respondent bases for a filtered table (as with Deltapoll and Number Cruncher) is marked EstimatedBasis and kept out of the primary series. We never scale a UK or GB headline to obtain England: if Northern Ireland is in the sample, it is excluded in the same way as Scotland and Wales.
Whole-number figures are allowed and flagged (RoundingPrecision = 1pp). Rounding is tested separately, by shifting every rounded input by half a point either way, and is not added to the estimated polling error.
The primary series
Applying these rules leaves 17 final polls: 2 in 2015, 4 in 2017, 5 in 2019 and 6 in 2024. Three of them (YouGov 2015, Qriously 2019 and YouGov 2024) were recovered in a later search of archives and third-party sources, which raised the series from 14 polls. Eleven have a printed England column; six are reconstructed, five from regional figures and one (Qriously) from respondent-level data.
| Election | Pollster (name at the time) | Fieldwork | England figure | England base |
|---|---|---|---|---|
| 2015 | Survation | 4–6 May | Direct | 3,334 |
| 2015 | YouGov | 4–6 May | Direct (whole numbers) | 8,619 |
| 2017 | Ipsos MORI | 6–7 Jun | Direct | 603 |
| 2017 | ComRes | 5–7 Jun | Direct | 1,128 |
| 2017 | Survation | 6–7 Jun | Direct | 1,988 |
| 2017 | Opinium | 4–6 Jun | Direct | 1,707 |
| 2019 | Ipsos MORI | 9–11 Dec | Direct | 1,538 |
| 2019 | Savanta ComRes | 9–10 Dec | Direct | 1,487 |
| 2019 | Survation | 10–11 Dec | Direct | 1,483 |
| 2019 | Opinium | 10–11 Dec | Reconstructed | 1,499 |
| 2019 | Qriously | 5–8 Dec | Reconstructed (respondent-level data) | 1,008 |
| 2024 | Ipsos | 1–3 Jul | Direct | 1,247 |
| 2024 | Savanta | 2–3 Jul | Direct (UK sample) | 1,382 |
| 2024 | Opinium | 1–3 Jul | Reconstructed | 1,957 |
| 2024 | Norstat | 1–3 Jul | Reconstructed | 1,877 |
| 2024 | JL Partners | 2–3 Jul | Reconstructed | 991 |
| 2024 | YouGov | 25–27 Jun | Reconstructed (whole numbers) | 2,762 |
Table: The primary evidence series. All are Grade A and on the headline basis. England base is the weighted number of respondents in England behind the headline.
Four sensitivity series sit alongside it: the primary series plus Grade C 2015 transcriptions believed to be on the headline basis; that plus the 2015 basis-mismatch figures; the primary series plus the estimated-basis reconstructions; and everything together. They exist to show which conclusions depend on the choice of evidence, and one does (see below).
A worked example: 2017
The calculation is easiest to follow on one election. In 2017 four final polls qualify. Their England figures, and the England result, were:
| Con | Lab | LD | UKIP | Green | |
|---|---|---|---|---|---|
| Ipsos MORI | 48.36 | 36.49 | 7.75 | 3.96 | 2.24 |
| ComRes | 46.23 | 35.94 | 9.76 | 5.15 | 2.57 |
| Survation | 44.11 | 41.80 | 8.35 | 2.67 | 2.57 |
| Opinium | 45.66 | 38.22 | 8.38 | 4.87 | 2.46 |
| Average | 46.09 | 38.11 | 8.56 | 4.16 | 2.46 |
| Result (England) | 45.50 | 41.98 | 7.82 | 2.05 | 1.87 |
| Error | +0.59 | −3.87 | +0.74 | +2.11 | +0.59 |
Table: England final polls and result, 2017 general election, percentage points.
The Conservative average is \((48.36 + 46.23 + 44.11 + 45.66)/4 = 46.09\), and the result was 45.50, so by equation (2) the error is \(+0.59\): the polls slightly overstated the Conservatives. Labour's average was 38.11 against a result of 41.98, an error of \(-3.87\). The lead error from equation (3) is \(-3.87 - 0.59 = -4.45\): the polls had Labour about 8.0 points behind, and it finished 3.5 behind.
The mean absolute error across the five parties is \((0.59 + 3.87 + 0.74 + 2.11 + 0.59)/5 = 1.58\) points. The same figure appears as the pollsters' score for 2017 in the backtests of Correcting the raw answers: the method behind our UK projection, which is a useful check that the two calculations agree.
The individual polls tell a second story. Survation was within 0.2 points on Labour; ComRes was 6.0 points short. The Conservative errors ranged from \(-1.4\) to \(+2.9\). So in 2017 the Labour miss was shared by the industry while the Conservative figure was mostly a matter of which pollster one read. We return to this distinction below.
The record, 2015–2024
The table below applies the same calculation to all four elections. It is computed from the database each time the page is served, so it will change if further final polls are recovered and pass the evidence rules.
| Election | Final polls | Con | Lab | LD | Reform | Green | Lab lead |
|---|---|---|---|---|---|---|---|
| 2015 | 2 | −5.7 | +1.5 | +1.8 | +0.8 | +0.9 | +7.1 |
| 2017 | 4 | +0.6 | −3.9 | +0.7 | +2.1 | +0.6 | −4.5 |
| 2019 | 5 | −1.6 | +0.9 | −0.6 | +0.8 | +0.1 | +2.5 |
| 2024 | 6 | −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 square | 3.5 | 3.0 | 1.1 | 1.5 | 0.6 |
Three patterns stand out.
The Conservatives were understated at three of the four elections. The misses in 2015 and 2024 were large, 2019 smaller, and 2017 the exception with a small overstatement. That pattern holds in every sensitivity series; what changes between series is its size, which depends heavily on 2015.
The UKIP, Brexit Party and Reform family was overstated at all four. The overstatement is modest, one to two points, but it has the same sign every time in the primary series. This finding is less secure than it looks. The 2015 figure rests on two polls; when three Grade C transcriptions for 2015 (ComRes, Opinium and BMG) are added, the family's 2015 error turns negative. At GB level, where all twelve 2015 final polls can be used, UKIP's error was close to zero. The primary series says "overstated every time"; the fuller but less verified 2015 evidence says "overstated since 2017". Either way it has not been understated since 2015.
Labour's error changes sign. Labour was overstated in 2015, 2019 and 2024 and understated in 2017, and the size varies from under one point to over four. The mean is small only because large errors of opposite sign cancel.
The lead error inherits Labour's instability and adds the Conservatives' consistency. It was large and positive in 2015 and 2024, large and negative in 2017, and smaller and positive in 2019. Its mean is positive, meaning the polls have more often flattered Labour relative to the Conservatives than the reverse, but a reader who had corrected the 2017 polls using 2015's lead error would have made the 2017 forecast worse, not better. That is the central difficulty with correcting polls for past error, and it is why the method described in Correcting the raw answers: the method behind our UK projection is tested election by election rather than on the average.
Lib Dem and Green errors are small and change sign. There is no evidence of a consistent bias for either, though both parties were polled at lower shares in 2015–2024 than they are now, which matters for the bands discussed below.
Industry error and house effects
A final-poll error has two sources. Part of it is shared by every pollster: something about the election, or about how polls in general were conducted, that pushed all of them the same way. Part is specific to each pollster: its sampling, weighting, turnout model and question wording. The second part is often called a house effect.
For election \(t\) and party \(p\), the average squared poll error splits exactly into the square of the industry error and the variance of the polls around their own average:
The share of the first term in the total says how much of the error was common to the industry. Computed from the primary series:
| Election | Polls | Con: industry error | Con: share shared | Lab: industry error | Lab: share shared | Range of lead errors |
|---|---|---|---|---|---|---|
| 2015 | 2 | −5.7 | 91% | +1.5 | 75% | +6.2 to +8.0 |
| 2017 | 4 | +0.6 | 13% | −3.9 | 74% | −8.3 to +1.2 |
| 2019 | 5 | −1.6 | 53% | +0.9 | 15% | −0.7 to +8.4 |
| 2024 | 6 | −3.5 | 82% | +4.3 | 89% | +4.0 to +11.0 |
Table: Decomposition of England final-poll error by equation (5), primary series. Errors in points; "share shared" is \(e_{p,t}^2\) divided by the mean squared poll error.
2024 was an industry miss: every qualifying pollster overstated Labour, by between 2.4 and 6.4 points, and most of the squared error on both main parties was shared. 2017 was an industry miss on Labour but not on the Conservatives. 2019 was mixed, with one pollster (Savanta ComRes) responsible for much of the spread in the lead. With only two polls, the 2015 figures say little about the spread between pollsters.
This matters for how errors are used. A shared error cannot be removed by averaging more polls; only a change in method, or a correction from outside the polls, can reduce it. Pollster-specific error does shrink when polls are averaged. No pollster has usable England data for more than three of the four elections, so house effects estimated from final polls are descriptive rather than predictive. The pollster ratings use them, with heavy shrinkage towards a poor score, for exactly that reason (see Rating the British pollsters: accuracy, the Labour–Conservative lead and openness).
Bias, variance, and what the literature says
Two kinds of error
Mean error and RMS error answer different questions. The mean asks whether the polls lean in a consistent direction; the RMS asks how far off they typically are. They are related by an identity that follows directly from equation (4):
The first term is systematic error, or bias; the second is the variation of the error from one election to the next. For the Reform family almost all of the RMS is bias: the errors are all positive and similar in size. For Labour almost all of it is variation: the errors are large but their mean is small. The Conservatives have both.
This is the same distinction that Shirani-Mehr et al. (2018) draw for election polls in the United States. Studying a large collection of US election polls, they show that the error of a poll is much larger than its sampling margin would suggest, and that a substantial part of the excess comes from error shared by polls of the same election. That shared part behaves like bias within an election but varies across elections, which is precisely the pattern in Labour's England record. Averaging more polls from the same election does little about it.
Jennings and Wlezien (2018), assembling polls from many countries over several decades, find that polling error has not been growing over time, despite the common impression that polls have become less reliable. Their result is a useful corrective. It does not mean that polling error is small, only that a miss like Britain's in 2015 is not in itself evidence that polling has got worse. Prosser and Mellon (2018) reach a similar conclusion in their review of trends in polling accuracy, which also discusses the causes of recent misses.
The 2015 miss
The 2015 general election is the best documented British polling failure, and it shaped the methods pollsters used in the elections that followed. The inquiry commissioned by the British Polling Council and the Market Research Society (Sturgis et al., 2016) concluded that the main cause was unrepresentative samples: the people the polls reached were not representative of those who voted, in ways that weighting did not repair. It found little support for late swing or for voters misreporting their intentions. Mellon and Prosser (2017), using British Election Study data, identified a specific version of that problem: polls contained too many politically engaged respondents and too few non-voters, so their samples of the whole population were wrong before any turnout adjustment was made.
The England figures here are consistent with that account. In 2015 the Conservatives were understated by several points, and the two primary polls agree on the direction. Pollsters changed their methods after 2015, and today most weight their samples to recalled past vote and apply their own turnout and don't-know steps. Those steps change what a headline figure means, which is one reason why From answers to headlines: the stage model of a published British poll looks beneath the headline at what respondents actually said.
Why bias is not simply subtracted
It would be natural to take the mean error from the table above and subtract it from today's polls. The first phase of this project tested exactly that, with several variants (the plain mean, recency-weighted means, a mean shrunk towards zero, and "the last election's error again"), each predicting every election only from the ones before it. None improved on the uncorrected average at every election, and the unshrunk corrections were worse on average. The reason is visible in the Labour column. An error that changes sign cannot be predicted from its own mean.
The correction that the site does now apply, described in Correcting the raw answers: the method behind our UK projection and Why we start from British pollsters' raw answers rather than the published headline, works differently: it starts from what respondents said before the pollsters' own adjustments, where the errors turn out to be steadier, and takes off only half of their average. That is a separate argument, built on separate evidence. The published-headline errors in this paper remain the basis for measuring uncertainty.
From error to an 80% band
The error record is used to say how far the truth could plausibly be from our estimate. For each party the half-width of the 80% band is
where \(z_{0.9}\) is the 90th percentile of the standard normal distribution. The band is the estimate plus or minus \(h_p\). For example, a party whose past errors had an RMS of 3.0 points would get a band of \(\pm 3.8\) points: if today's estimate is 25%, the 80% range is about 21% to 29%.
Three features of this choice deserve comment.
First, the band uses the RMS, not the standard deviation. By equation (6) the RMS includes the bias term. A party that has been overstated by two points at every election has a small standard deviation but a large RMS, and the band reflects that the estimate could well be two points out. Using the standard deviation would, in effect, assume the bias had been corrected.
Second, the band is centred on the estimate, not shifted by the past mean error. Equation (7) is what one would get by assuming that the next election's error is drawn from a normal distribution with mean zero and variance \(\text{RMS}_p^2\). That is a deliberately cautious reading of four elections: it treats past bias as a reason for wider uncertainty rather than as a reliable correction. Any directional correction is made, and tested, separately.
Third, the normal distribution is an approximation. With four elections the tails cannot be checked. An 80% band depends less on the shape of the tails than a 95% band would, which is some comfort, but the band should be read as a calibrated guide rather than a probability statement that has been verified.
The bands appear on the track record page and in the home page's chart of past misses, where each party's current England share is shown with its 80% range. In the "election today" forecast the vote bands are computed live from the error history by equation (7).1 The SNP's and Plaid Cymru's bands come from Scotland's and Wales's own records. For the election-day forecast, which must also allow for real change in opinion before polling day, the band combines the two sources of uncertainty as \(h_p = 1.2816\sqrt{\text{RMS}_p^2 + d_p^2}\), where \(d_p\) is the standard deviation of drift for the time remaining (see From today to polling day: how far British polls move before a general election). The seat simulation does not use the bands directly: it replays whole past error patterns, so that parties move together as they have in practice (see Simulating UK polling error: where the chances come from).
Restore Britain has no error history and therefore no calibrated band. The Green and Lib Dem bands were calibrated when those parties polled much lower than they do now, so they are probably too narrow; a band proportional to share is a candidate improvement that has not yet been adopted.
Limits
The evidence is thin, and it is better to say so plainly.
Four elections. Every summary statistic here rests on four numbers per party. An RMS computed from four errors is itself very uncertain: a fifth election like 2015 or 2024 could raise it substantially, and one like 2019 could lower it. The patterns described as consistent (Conservatives understated, Reform family overstated) could arise by chance more easily than a reader might assume. If each election's error were equally likely to be positive or negative, four errors of the same sign would still occur one time in eight. The earlier elections of 2010 and 2005 are loaded for constituency results but do not yet have an England poll benchmark; bringing 2010 in is planned.
Few polls per election. The primary series has two polls in 2015 and no more than six in any year. In 2015 especially, the industry error is really the error of two pollsters, and the decomposition of equation (5) says almost nothing about spread. The Grade C evidence for 2015 changes the sign of the Reform family's error, and until those figures are checked against archived originals the finding of an overstatement at every election rests partly on which pollsters happened to publish usable England tables.
Selection by availability. Polls enter the primary series because their tables survived and show England on the headline basis, not because they are representative of the industry. The pollsters with the most detailed tables may differ from the rest. Where the comparison can be made (2015 at GB level) the fuller set told a somewhat different story from the primary subset.
House effects and a changing industry. Pollsters change their methods between elections, sometimes in response to the last miss, so a house effect measured in 2019 may not apply in 2024. Several pollsters in today's average have no checked final poll at all. YouGov's headline has been produced by a multilevel regression model since 2025, whereas its final polls from 2015 to 2024 were standard polls, so its historical record describes a different product from the one it now publishes.
Final polls only. These errors compare the last polls of a campaign with the result. They say nothing directly about polls taken a year or three years before an election, when real opinion has time to move. That is why the "election today" estimate and the election-day forecast are kept separate, and why the latter has its own model of how votes move before polling day (Wlezien et al., 2013; Erikson and Wlezien, 2012).
Rounding. Two of the seventeen primary polls report whole numbers. A half-point shift in every rounded input moves the affected industry errors by well under half a point, which does not change any conclusion, but the effect is not built into the bands.
With those limits stated, the record is still informative. It shows that the published final polls have missed the Labour–Conservative lead in England by several points at most recent elections, that the direction of the Labour miss has not been predictable from the previous one, and that the right-populist challenger has not been understated since 2015. The rest of the method is built around those three facts.
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.
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.
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.
Wlezien, C., Jennings, W., Fisher, S., Ford, R. and Pickup, M. (2013). Polls and the vote in Britain. Political Studies, 61(S1), 66–91.
-
The band tables on the track record and methodology pages show the values stored with the model version (PollModel 1.1) when it was built from the primary series; the forecast recomputes the same quantity from the error history each time it runs. The two agree while the primary series is unchanged.↩
How to cite
Polling Forecast (2026). Measuring British polling error in England, 2015–2024. Retrieved 20 November 2026 from https://www.pollingforecast.com/research/uk-measuring-polling-error. Live figures change as new polls arrive; the date above is when the text was last revised.


