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 · Pollsters

Rating the British pollsters: accuracy, the Labour–Conservative lead and openness

Abstract. We score each pollster out of 100 on its checked final-poll record in England (70 points) and on how much of its working it publishes (30 points). Since 13 October 2026 the record is ranked against every checked final poll, and a short record is pulled towards the typical final poll rather than a poor one. Most British pollsters with a checked final poll now grade B.

Polling Forecast · updated 9 November 2026 · 20 min read · pollster ratings, house effects, shrinkage, sample skew, disclosure, England

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 Britain the first run of the new scale put Survation on B (73), Ipsos on B (71), Opinium and YouGov on B (62) and JL Partners on C (51). Pollsters with no checked final poll, such as More in Common and Find Out Now, are listed as new until one is added.

The question and the scope of the answer

Every forecast that averages opinion polls makes an implicit judgement about pollsters. Equal weighting says they are interchangeable; excluding one says it is not to be trusted. This paper makes that judgement explicit. It sets out how the site rates each British pollster out of 100, what evidence the rating uses, why its first version was deliberately made hard to score well on, and how it was rescaled in October 2026.

The rating answers a narrow question: on the evidence we can check, how much should a reader trust this pollster's published figure, and how easily can they see how it was made? It does not say which pollster is right today. A pollster that shows Labour well ahead of the others may be the only one that is correct, and we find out only on polling day. Nor is the rating used to weight the polls in our forecast. Our projection reads each pollster's raw answers and applies a correction estimated across the whole industry (Correcting the raw answers: the method behind our UK projection); the ratings sit alongside that, on the pollster ratings page, as a guide for readers.

The rating is about England, where polling error is measured cleanly without the SNP and Plaid Cymru (Measuring British polling error in England, 2015–2024), and about standard voting-intention polls rather than MRP projections or party-commissioned polls. It has three parts: accuracy (50 points), the Labour–Conservative lead (20 points), both measured from past final polls and shrunk towards the record of a poor pollster, and openness (30 points), from a review of what each pollster publishes.

The evidence: checked final polls

The accuracy parts of the rating use only the primary benchmark: each pollster's final England poll at the general elections of 2015, 2017, 2019 and 2024, where the England figure has been taken from the pollster's own tables and is on the same basis as its published headline. The final poll is the one whose fieldwork finished latest before polls opened, provided it was published before voting began. Polls whose England figure had to be estimated from incomplete bases, or which were on a different basis from the headline, are kept for sensitivity analysis but do not count here. The rules and their reasons are set out in Measuring British polling error in England, 2015–2024 and The UK data pipeline: collecting, reading and checking British pollsters' tables.

At the time of writing the primary benchmark contains 17 final polls from eight pollsters. Several of today's largest pollsters have none: either they did not publish a standard final poll that could be checked in 2024, or they did not exist in their current form. That fact drives much of what follows.

For pollster \(j\) at election \(t\), let \(e_{j,t,p} = \hat{v}_{j,t,p} - v_{t,p}\) be the poll's England share for party \(p\) minus the party's actual England share, in percentage points, so that a positive error means the party was overstated. UKIP and the Brexit Party count as Reform UK for this purpose. Two summary quantities are taken from each poll. The miss is the mean absolute error across the five parties with a full history,

\[ m_{j,t} = \frac{1}{5} \sum_{p \in P} \left| e_{j,t,p} \right|, \qquad P = \{\text{Con}, \text{Lab}, \text{LD}, \text{Ref}, \text{Grn}\}, \tag{1} \]

and the lead error is the error on Labour's lead over the Conservatives,

\[ \ell_{j,t} = e_{j,t,\text{Lab}} - e_{j,t,\text{Con}}. \tag{2} \]

A positive \(\ell\) means the poll overstated Labour's lead (or understated the Conservative lead). A pollster's own record over its \(n_j\) checked elections is then the average miss \(\bar{m}_j = n_j^{-1} \sum_t m_{j,t}\) and the average absolute lead error \(\bar{\ell}_j = n_j^{-1} \sum_t |\ell_{j,t}|\).

Why score the lead separately when its two ingredients already appear in the miss? Because a poll can be close on every party and still get the government wrong. Suppose a poll overstates Labour by 2 points and understates the Conservatives by 2, with the smaller parties exact. Its miss is only 0.8 points, which is better than the industry average, but its lead is out by 4 points, enough to turn a hung parliament into a comfortable majority. The lead is the number that decides who governs, and the 2017 and 2024 elections were both, in large part, lead misses. The lead component therefore places extra weight on the two largest parties, deliberately.1

Scoring accuracy and the lead

Each record is converted to points by a clamped linear scale. Accuracy is worth 50 points: a miss of 0.75 points or better scores the full 50, a miss of 2.75 or worse scores nothing, and points fall linearly in between. The lead is worth 20 points: a lead error of 1 point or better scores 20, and 9 points or worse scores nothing. Writing \(\tilde{m}_j\) and \(\tilde{\ell}_j\) for the rated miss and lead error (defined in the next section),

\[ A_j = 50 \cdot \operatorname{clip}\!\left( \frac{2.75 - \tilde{m}_j}{2.75 - 0.75},\ 0,\ 1 \right), \tag{3} \]
\[ L_j = 20 \cdot \operatorname{clip}\!\left( \frac{9 - \tilde{\ell}_j}{9 - 1},\ 0,\ 1 \right), \tag{4} \]

where \(\operatorname{clip}(x, 0, 1)\) limits \(x\) to the interval from 0 to 1.

The end points are judgements, not estimates, and it is better to say so plainly. They were chosen with the observed spread of the benchmark in view. The best single final polls in the benchmark missed by a little over half a point per party (Ipsos and Opinium in 2019), so 0.75 is a demanding but reachable standard for a pollster's average. The worst missed by more than 3.5 points (Savanta in 2024). On the lead, the best polls were within about a point, and the worst were out by 9 to 11 points in 2024. A pollster at the industry average miss (about 1.88 points) would score 22 of the 50 accuracy points before any shrinkage, which places the average pollster a little below the middle of the scale. That is intended: at three of the last four elections the polls as a whole missed badly, so the average is not a standard worth half marks or more.

Shrinkage towards a poor pollster

A pollster with one good final poll has not shown that it is accurate. It may have been lucky. The difference between one election and four is the difference between an anecdote and a record, and the rating has to reflect it. The standard statistical answer is shrinkage: pull each pollster's observed record towards a common value, by an amount that falls as its evidence grows (Efron and Morris, 1975). The rating uses the simplest form of this, a weighted average of the pollster's own record and a prior value, where the prior is given the weight of \(k\) pseudo-elections:

\[ \tilde{m}_j = \frac{n_j \, \bar{m}_j + k \, m^{*}}{n_j + k}, \qquad \tilde{\ell}_j = \frac{n_j \, \bar{\ell}_j + k \, \ell^{*}}{n_j + k}, \qquad k = 2. \tag{5} \]

The unusual choice is the prior. In conventional shrinkage the common value is the mean of the population, so that an unknown unit is assumed to be typical. Here it is not. The prior \(m^{*}\) is the 75th percentile of the miss across every final poll in the primary benchmark, and \(\ell^{*}\) is the 75th percentile of the absolute lead error. In other words, every pollster is pulled towards a poor pollster: one that three in four checked final polls beat. The percentile is computed by linear interpolation between order statistics, and with the current 17 polls it gives \(m^{*} \approx 2.50\) points and \(\ell^{*} \approx 8.3\) points.2 A pollster with no checked final poll has \(n_j = 0\), so its rated miss and lead error are exactly those of the poor pollster.

This is punitive by design, and the reasons are specific to the problem. First, the absence of a record is not neutral evidence. Several pollsters have no checked final poll because they published no standard final poll we could verify, or because their tables were not available; a rating that assumed them average would reward that. Second, the cost of misplaced trust is asymmetric. A reader who trusts an unproven pollster that turns out to be poor is worse off than one who distrusts an unproven pollster that turns out to be good, because in the second case the good pollster's figures are still published and still in the average. Third, the recent history of British polling does not support a presumption of accuracy. The industry missed in 2015, 2017 and 2024 (Sturgis et al., 2016; Prosser and Mellon, 2018), and while polling error does not appear to be getting worse over time across countries (Jennings and Wlezien, 2018), neither is it small enough to assume away.

The value \(k = 2\) controls how quickly a pollster can escape the prior. It means that two further elections at the poor level are imagined alongside whatever the pollster has actually done. With \(k = 2\), a pollster with four checked elections still has a third of its rated record set by the prior. The consequence is easiest to see by asking what a perfect pollster could score: one whose every final poll had zero error on every party.

Checked elections Rated miss Accuracy points Rated lead error Lead points Best possible total (with full openness)
0 2.50 6.3 8.34 1.6 38 (D)
1 1.66 27.1 5.56 8.6 66 (B)
2 1.25 37.5 4.17 12.1 80 (A)
3 1.00 43.8 3.34 14.2 88 (A)
4 0.83 47.9 2.78 15.5 93 (A)

Table: The ceiling on the score for a hypothetical pollster with no error at any checked election, using the current prior (miss 2.50, lead error 8.34) and full openness points.

The table makes the design visible. No amount of openness lifts an unproven pollster above a D. One flawless election can earn a B at most. An A needs at least two checked elections of very high quality. At the other end, a pollster that simply matched the industry's average record at all four elections (a miss of about 1.88 and a lead error of about 5.6) would earn only about 23 of the 70 record points, and would need nearly full openness to reach a C.

A worked example

Survation has three checked final polls in the benchmark. Its misses were 2.73 points in 2015, 0.68 in 2017 and 1.04 in 2019, an average of \(\bar{m} = 1.48\). Its lead errors were \(+8.0\), \(+1.2\) and \(+1.1\), an average absolute lead error of \(\bar{\ell} = 3.43\). Applying equation (5),

\[ \tilde{m} = \frac{3 \times 1.48 + 2 \times 2.50}{5} = 1.89, \qquad \tilde{\ell} = \frac{3 \times 3.43 + 2 \times 8.34}{5} = 5.40 . \]

Equations (3) and (4) then give \(A = 50 \times (2.75 - 1.89)/2 \approx 21.5\) and \(L = 20 \times (9 - 5.40)/8 \approx 9.0\). Survation's own record is better than the industry average, but after shrinkage its rated miss sits at about the industry average, because three elections are not yet enough to outweigh two imagined poor ones. Its openness points (below) bring it to a total of about 55. These inputs are historical and will change only if the benchmark does; the live table later in the paper is the authoritative figure.

Openness

The remaining 30 points reward a pollster for publishing enough that its headline can be checked. Our own method starts from what respondents actually said, before the pollster's turnout and don't-know steps (Why we start from British pollsters' raw answers rather than the published headline), and the evidence on past polling misses points to those steps and to sample composition as the places where things go wrong (Sturgis et al., 2016; Mellon and Prosser, 2017). A pollster that shows each step lets anyone see how much it moved the numbers. One that publishes only the headline asks to be taken on trust.

Five things are checked. Each is recorded as Yes, Partial or No from a review of the pollster's own published tables.

Raw stage (8 points). Voting intention for everyone asked, weighted, with don't knows included, before any turnout step. It is what people said, and it is what our reading of the polls starts from.

Turnout step shown (6 points). A table after the turnout filter or weight, so the effect of that step can be seen. Turnout adjustments were central to the 2017 miss.

Past-vote table (6 points). Voting intention broken down by 2024 vote, including people who did not vote in 2024. It shows who is switching and whether the sample's political composition looks plausible.

Likelihood-to-vote table (4 points). How likely each party's supporters say they are to vote. Needed to judge a turnout step.

England figure (6 points). An England column, or English regions with printed weighted bases, so England can be measured exactly rather than estimated.

Writing \(w_i\) for the points attached to item \(i\) and \(s_{j,i}\) for pollster \(j\)'s value on it,

\[ O_j = \sum_{i=1}^{5} w_i \, s_{j,i}, \qquad s_{j,i} = \begin{cases} 1 & \text{Yes} \\ 1/3 & \text{Partial} \\ 0 & \text{No,} \end{cases} \qquad w = (8, 6, 6, 4, 6). \tag{6} \]

A partial disclosure scores a third rather than a half because a table with gaps usually cannot be used to check the headline. A past-vote table with no column for 2024 non-voters, for example, cannot show whether a sample has too few of them, which is exactly the failure identified after 2015 (Mellon and Prosser, 2017). For the England figure, Partial means regions are printed but without weighted bases, or only estimated bases, so England can be approximated but not measured.

The reviews are stored in ref.PollsterDisclosure (script 10), one row per pollster per review date, with a short note against each item explaining the judgement. The table is append-only. A new review is a new row with a later date and the site uses the latest; an old review is never edited, so the history of each judgement is kept. The first review, dated 25 September 2026, covers fourteen pollsters and was compiled from a survey of each pollster's published tables. Whether a pollster's headline is model output is recorded in the same row (it is true for YouGov) but is not scored.

Only pollsters that have been reviewed receive an overall score. Pollsters we have not yet been able to review, because their tables could not be obtained, are shown with their accuracy and lead points only and are listed separately as "not rated yet".

Grades, provisional marks and the current table

The overall score is the rounded sum,

\[ S_j = \operatorname{round}\left( A_j + L_j + O_j \right), \tag{7} \]

and it maps to a letter grade.

Grade Score Meaning shown on the site
A 75 and above Proven accurate and fully open
B 60–74 A good record, and checkable
C 45–59 Some evidence of accuracy, or very open
D 30–44 Little evidence it gets elections right
E below 30 Unproven and hard to check

Table: Grade boundaries.

A rating is marked provisional (an asterisk on the site) when the pollster has fewer than two checked final polls. For such a pollster the prior carries at least two-thirds of the weight in equation (5), so the rating says more about the absence of evidence than about the pollster. The main table lists pollsters that have published a poll in the last 365 days; pollsters with a checked record but no recent polls appear separately as past pollsters.

PollsterGradeScoreElections checkedOwn missAccuracy ptsLead ptsOpenness pts
SurvationB7331.48351425
IpsosB7131.73341226
OpiniumB6231.59341314
YouGovB6222.03301220
JL Partners*C5112.10291210
BMG Research*––0–3012–
Convergent Opinion*––0–301212
Deltapoll*––0–30124
Find Out Now*––0–301221
Focaldata*––0–301216
Freshwater Strategy*––0–301218
Good Growth Foundation*––0–301220
Lord Ashcroft Polls*––0–301225
More in Common*––0–301217
Techne*––0–3012–
Trajectory Partnership*––0–3012–
Verian*––0–30126
Live Pollster ratings out of 100: accuracy (50), lead (20) and openness (30). * fewer than 2 checked elections. Computed from the database on 19 November 2026.

The table is computed from the database each time the ratings are refreshed. Two features will hold however the individual numbers move. The provisional pollsters with no checked elections all share the same accuracy and lead points, since for them equation (5) returns the prior exactly; they are separated only by openness. And the pollsters with the best records are not necessarily the most open, and vice versa, which is why the two are scored separately rather than combined into a single impression.

House effects and sample skew

Two further measures appear beside the rating on the pollster pages. Neither is part of the score. Both describe how a pollster differs from the others now, which is a different question from whether it has been right in the past.

House effects: today against the average

A pollster's house effect for party \(p\) is its latest England figure minus the England polling average:

\[ h_{j,p} = s_{j,p} - \bar{s}_{p}, \tag{8} \]

where \(s_{j,p}\) is the pollster's share for party \(p\) in its latest poll in the current average and \(\bar{s}_p\) is the England average itself (the latest primary-grade poll from each pollster with fieldwork ending in the last 28 days, equally weighted). It is computed only for pollsters currently in the average. Restore Britain is compared only where the pollster prompts for it, since pollsters that do not prompt leave its support inside Others. The ratings table's "lean now" column summarises the house effects as a single number, the pollster's Labour–Conservative lead minus the average lead, \(h_{j,\text{Lab}} - h_{j,\text{Con}}\).

A house effect is not an error. Because the average is built from the pollsters themselves, house effects sum to roughly zero across them by construction, and the pollster furthest from the pack may be the one closest to the result. That is why the site shows, next to each pollster's current lean, its average error per party at past elections. A pollster that leans the same way today as it erred at past elections gives a reader something to think about. A pollster that leans one way today but has erred in both directions before gives much less. The pollster pages summarise this with a lean pattern: whether the pollster overstated Labour's lead at every checked election, understated it at every one, or missed in different directions.

Sample skew: who a pollster reaches before weighting

Sample skew asks a different question: before the pollster weights its data, does its sample contain too many or too few of each party's 2024 voters? Pollsters that print the unweighted size of each column in their raw-stage table make this measurable. For poll \(i\), let \(u_{i,p}\) be the unweighted number of respondents in the column for people who voted for party \(p\) in 2024, and let \(g_p\) be party \(p\)'s actual 2024 vote share in Great Britain. Over the set \(Q_i\) of the five main parties for which the poll prints such a column, the skew is the sample's share minus the real share, both taken among the same parties:

\[ k_{i,p} = 100 \, \frac{u_{i,p}}{\sum_{q \in Q_i} u_{i,q}} \; - \; 100 \, \frac{g_p}{\sum_{q \in Q_i} g_q}, \qquad p \in Q_i. \tag{9} \]

A positive value means the pollster reached too many of that party's 2024 voters. A poll is used only if \(Q_i\) contains at least three parties, including both Conservative and Labour. Columns that overlap others (Ipsos prints a single "Other, including Lib Dem and Green" column of 2024 voters) are ignored, so for such pollsters the skew is available only for the parties with their own column. The pollster's skew is the average of \(k_{i,p}\) over its polls with fieldwork ending in the last 120 days that have stored raw tables. Because the measure needs unweighted column sizes, it cannot be computed for pollsters that print only weighted sizes.3

Renormalising within \(Q_i\) matters. Samples include 2024 non-voters, people who cannot remember, and supporters of smaller parties, and pollsters print different sets of these columns. Comparing each party's share of the same set of columns in both the sample and the result removes those differences.

The reason to look at skew is that weighting corrects the number of each group in the sample but not the kind of person reached within it. If a pollster reaches too few 2024 Conservatives, weighting gives each one it does reach more influence. That costs precision: in the standard approximation, unequal weights inflate the variance of an estimate by a design effect of about \(1 + \mathrm{CV}^2\), where CV is the coefficient of variation of the weights, so a heavily weighted sample behaves like a smaller one (Kish, 1965). More seriously, it relies on the Conservatives it reaches being typical of those it does not. The 2015 inquiry found that unrepresentative samples, not late swing, were the main cause of that year's miss (Sturgis et al., 2016), and Mellon and Prosser (2017) showed how a surplus of politically engaged respondents survived weighting. Skew does not measure that deeper problem, but it shows where a pollster is leaning hardest on its weights.

Why skew is shown but not scored

Skew is deliberately left out of the score, for three reasons. The main one is that accuracy already counts the result: a pollster whose skewed sample weights out well will have a good final-poll record, and one whose sample fails despite weighting will have a poor one. Scoring skew as well would count the same failure twice, and would penalise a pollster for a symptom it may have successfully treated. Second, skew can be measured only for pollsters that publish unweighted column sizes. Scoring it would therefore expose the most open pollsters to a penalty that the least open could never incur, which runs against the purpose of the openness component. Third, there is no evidence yet, within our small benchmark, on how much skew predicts error. It is shown as information, with the caveat printed on the page that weighting corrects for too many or too few of a party's voters only if the ones a pollster does reach are typical of the rest.

The rating also leaves out, on purpose, membership of the British Polling Council (the only list of members we found was undated and out of date), how often a pollster polls, who commissions its polls, and which way it leans today. None of these tells a reader whether the pollster will be right.

From a generous scale to a punitive one

The rating was introduced and then substantially revised on the same day, 25 September 2026, and the change is worth recording because it shows what the scale is for.

The first version gave 60 points for accuracy and 40 for openness. Accuracy was scored as \(60 \times (3.0 - \tilde{m})/2.5\), so that a miss of 0.5 earned the full 60 and a miss of 3.0 earned nothing. The record was shrunk towards the industry average miss of 1.88 points, again as if the pollster had two more elections at that level. There was no separate lead component. Openness was scored on the same five items but with weights of 10, 8, 8, 6 and 8, and a partial disclosure earned half. Grades were A for 70 and above, B for 60–69, C for 50–59 and D below 50.

Under that scale an unproven pollster was assumed average. Its accuracy score was \(60 \times (3.0 - 1.88)/2.5 \approx 27\) of 60 points, or 45 per cent of the available record points, before it had shown anything at all. Combined with generous openness points, this pushed the ratings upwards. The first ratings put Survation at B 68, Ipsos at B 65 and Lord Ashcroft Polls, which had no checked final poll, at B 62.

The owner's reaction was that the pollsters all seemed to get a top rating and that the scale needed to be punitive. Part of that impression turned out to have a different cause. The list of polls showed a column of letter grades that were almost all A, and those were source grades: A meant only that the figures had been taken from the pollster's own table, not that the pollster was good. That column now shows the pollster rating instead. But the underlying point stood. A rating in which an unproven pollster is presumed average, and in which most active pollsters earn a B, does not help a reader distinguish between them.

The revision made four changes. The prior moved from the industry average to the 75th percentile, so that the absence of evidence counts against a pollster rather than for it. The accuracy scale was tightened to 50 points between 0.75 and 2.75. A lead component worth 20 points was added. Openness was reduced to 30 points, with partial disclosure worth a third rather than a half, and the grade boundaries were widened to five letters with A requiring 75. The effect on an unproven pollster is stark: its record points fell from about 27 of 60 to about 8 of 70, from 45 per cent of the available record points to about 11 per cent. On the first run of the new scale no pollster reached A or B; the best-rated were in the C band, and the provisional pollsters fell to D or E.

The revised scale is a choice. A reader who prefers to presume competence would reach different grades from the same evidence, which is why the components are published separately.

Limitations

The most serious limitation is the size of the evidence. The benchmark has four elections and 17 checked final polls. No pollster has more than three checked elections, and several of the largest active pollsters have none. The shrinkage in equation (5) is a response to this, not a solution: it prevents the rating from overreacting to one good or bad poll, but it cannot manufacture evidence that does not exist.

Methods change between elections, and a rating built on 2015–2024 final polls may describe a pollster that no longer exists in that form. YouGov's headline, for example, has been projected through its MRP model since 2025, whereas its checked final polls were standard polls. A good 2019 poll does not guarantee a good one now, and the site says so.

The prior itself depends on the benchmark. If more final polls are checked and added, the 75th percentile will move and every pollster's rated miss will move with it. The end points of the scales in equations (3) and (4), the choice of \(k = 2\), the weights on the openness items and the one-third credit for partial disclosure are all judgements. They were made with the observed spread of errors in view, but they were not estimated and have not been validated against future performance, because there is as yet no future to validate against. The first real test of whether higher-rated pollsters are more accurate will come at the next general election.

Finally, the rating is an England rating, built from England figures. A pollster could be good in England and poor in Scotland, or the reverse. Scotland and Wales have their own error records (Scotland and Wales: sparse polls and no steady lean), but they are not part of the pollster rating.

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.

Kish, L. (1965). Survey Sampling. New York: Wiley.

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.

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 two big parties' errors enter the score twice: once through the miss, where each counts for a fifth, and once through the lead. That is intended, but it means a pollster whose errors fall mainly on the smaller parties is treated more gently than one of the same overall miss whose errors fall on Labour and the Conservatives in opposite directions.↩

  2. If the benchmark were empty the code falls back to fixed values of 2.5 and 8.0. With the benchmark loaded these are never used.↩

  3. The decisions log, written when sample skew was introduced, lists Find Out Now among the pollsters for which skew is shown. The code states that Find Out Now prints only weighted column sizes and so cannot be measured, and the pollster page shows no skew for it. The code is authoritative; the log entry is out of date on this point.↩

How to cite

Polling Forecast (2026). Rating the British pollsters: accuracy, the Labour–Conservative lead and openness. Retrieved 20 November 2026 from https://www.pollingforecast.com/research/uk-pollster-ratings. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like