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

Correcting the raw answers: the method behind our UK projection

Abstract. Our projection starts from what poll respondents actually said, before each pollster's turnout and don't-know adjustments, and removes half of the average amount those raw answers have missed by in England since 2015. Tested on four elections, each predicted without its own result, it beat or matched the pollsters' published figures every time, but by modest margins, and it did not fix the large overstatement of Labour in 2024.

Polling Forecast · updated 9 November 2026 · 21 min read · raw answers, polling error, shrinkage, leave-one-out, England, projection

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). What this paper calls our reading is shown on the site as our adjusted average, and the headline figures as our projection.
  • Our vote shares beside the polls (26 October). Wherever our adjusted shares differ from the poll average, both are shown on the hub, the poll pages and the leader cards, so the size of the correction is visible.
  • The cutoff (28 September). Only elections from 2000 can shape a forecast. The four elections behind the correction (2010 to 2024) are all after it, so the correction is unchanged.

The question

Every published voting-intention figure is the end of a chain. A pollster recruits a sample, weights it to look like the population (almost always including how people say they voted last time), asks who they would vote for, and then applies its own judgement: it discounts people who say they are unlikely to vote, and it decides what to do with those who don't know. The headline is the output of that chain. Other forecasts take the headline as their input.

We start one step earlier. For every poll whose full tables we can read, we take the answers as they stood before the pollster's turnout and don't-know steps, and we ask a different question of history: not "how wrong were the headlines?" but "how wrong were the raw answers, and was that wrongness stable enough to correct?" This paper sets out the answer we act on. It defines the estimator formally, explains how much of the historical error we remove and why, reports the two out-of-sample tests the method has passed, walks through today's calculation, and lists what we tried and set aside.

The short version: in England, at every general election from 2015 to 2024, the raw answers understated the Conservatives and overstated the UKIP, Brexit Party and Reform family. We take half of the average raw-answer miss off each party, rescale, and average across pollsters. That is our reading of the polls, and it drives the forecast the site leads with (our projection). The companion paper Why we start from British pollsters' raw answers rather than the published headline explains why we begin from the raw answers rather than correcting the pollsters' headlines directly; From answers to headlines: the stage model of a published British poll describes how the stages are read from the pollsters' files; and Measuring British polling error in England, 2015–2024 sets out the historical benchmark.

Data and notation

What "raw" means here

We store each poll's tables as printed, stage by stage (see From answers to headlines: the stage model of a published British poll). The stage we call raw is the pollster's weighted table of everyone asked, don't knows included, which we label S1. We then keep only the party rows and rescale them to sum to 100, so that don't knows, refusals and would-not-votes are set aside rather than allocated. The headline, S3, is the pollster's published figure.

"Raw" is therefore not unweighted. Nearly every pollster's S1 is already weighted to demographic targets and to recalled 2024 vote; only Ipsos (for Great Britain) and Qriously in 2019 print a truly unweighted stage. What S1 excludes are the two judgements that vary most between pollsters and between elections: the turnout model and the treatment of don't knows.1

Symbols

Let \(p\) index parties and \(t\) index general elections. The main parties are \(\mathcal{P} = \{\text{CON}, \text{LAB}, \text{LD}, \text{REF}, \text{GRN}\}\), where REF stands for the UKIP, Brexit Party and Reform family (UKIP in 2015 and 2017, the Brexit Party in 2019). The elections with evidence are \(\mathcal{T} = \{2015, 2017, 2019, 2024\}\). All shares are for England and in percentage points.

For a final poll \(i\) at election \(t\), write \(r_{i,p,t}\) for the raw share of party \(p\) (the S1 table, parties only, rescaled to 100), \(h_{i,p,t}\) for the published headline share (parties only), and \(v_{p,t}\) for the result, party \(p\)'s share of the vote in England.

At each election we average over the \(n_t\) final polls for which both stages survive, giving \(\bar r_{p,t}\) and \(\bar h_{p,t}\). The raw error and the headline error are

\[ e^{\,raw}_{p,t} = \bar r_{p,t} - v_{p,t}, \qquad e^{\,head}_{p,t} = \bar h_{p,t} - v_{p,t}. \tag{1} \]

A positive error means the polls overstated the party. The mean raw error across elections is

\[ \bar e^{\,raw}_{p} = \frac{1}{|\mathcal{T}|} \sum_{t \in \mathcal{T}} e^{\,raw}_{p,t}. \tag{2} \]

We score any set of estimates \(\hat v_{p,t}\) by the mean absolute error across the five main parties, and by the error in the Labour lead over the Conservatives:

\[ \text{MAE}_t = \frac{1}{5}\sum_{p \in \mathcal{P}} \left| \hat v_{p,t} - v_{p,t} \right|, \qquad \text{LeadErr}_t = (\hat v_{\text{LAB},t} - \hat v_{\text{CON},t}) - (v_{\text{LAB},t} - v_{\text{CON},t}). \tag{3} \]

The mean absolute error per party is the primary metric set out when the project began; the lead error is a secondary one.

The evidence base

The historical evidence is the final polls of 2015, 2017, 2019 and 2024 whose raw and headline tables both survive in England: 18 polls in all. The table below shows the raw-answer errors, with the headline errors 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.

Two regularities stand out. The raw answers understated the Conservatives at every election, and overstated the Reform family at every election. Labour's raw error is positive on average but moves around a great deal, from a small understatement in 2017 to a large overstatement in 2024. The Liberal Democrats and Greens are small and mixed. This pattern, and in particular its stability for the Conservatives, is the whole basis of the method. It fits the explanation the 2015 inquiry gave for that year's miss, that the samples themselves were unrepresentative (Sturgis et al., 2016), and the finding that poll samples contain too many politically engaged people (Mellon and Prosser, 2017). We do not rely on either explanation being the full story; we rely on the pattern having held.

The estimator

Correcting one poll

Fix a shrinkage factor \(\lambda \in [0, 1]\), the share of the historical raw miss we are willing to remove. The correction for party \(p\) is

\[ c_p = \lambda \, \bar e^{\,raw}_{p} \quad \text{for } p \in \mathcal{P}, \qquad c_p = 0 \text{ otherwise}. \tag{4} \]

Parties outside \(\mathcal{P}\) (Restore Britain, the SNP, Plaid Cymru, Your Party and Others) have no raw-answer history in England and are not corrected. For today's poll \(j\) with raw England shares \(r_{j,p}\) (parties only), we subtract the correction, clamp at zero, and rescale so the parties again sum to 100:

\[ \tilde u_{j,p} = \max\!\left(0,\; r_{j,p} - c_p\right), \qquad \hat u_{j,p} = 100 \times \frac{\tilde u_{j,p}}{\sum_{q} \tilde u_{j,q}}. \tag{5} \]

The sum in the denominator runs over every party in the table, corrected or not. The value we use is \(\lambda = 0.5\), and the resulting corrections are shown below. Because we subtract the error, a party the raw answers have understated (a negative error) is moved up.

ConLabLDReformGreen
Average raw-answer miss−4.0+2.0+0.1+1.3+0.5
Correction (−0.5 × miss)+2.0−1.0−0.1−0.6−0.3
Live The correction applied to today's raw answers, England, points. Computed from the database on 19 November 2026.

Two details matter in practice. First, the clamp at zero only binds for a very small party with a positive correction; it has never bound for the main parties. Second, the rescaling changes little, because the main parties' mean raw errors nearly cancel: the raw answers' excess for Labour, Reform, the Lib Dems and Greens is roughly the Conservatives' shortfall. Rescaling mostly redistributes a few hundredths of a point, and it slightly adjusts the uncorrected parties.

A stylised example makes the arithmetic concrete. Suppose a poll's raw England answers were Conservative 20, Labour 28, Lib Dem 11, Reform 24, Green 12, Restore Britain 3 and Others 2, and the corrections were close to their current values (Conservative +2.0, Labour −1.0, Lib Dem −0.1, Reform −0.6, Green −0.3, in the "amount added" sign convention of the table). The corrected figures before rescaling are 22.0, 27.0, 10.9, 23.4 and 11.7, with Restore Britain and Others unchanged at 3 and 2. They sum to 100.0, so the rescaling does nothing visible, and the poll's Labour lead over the Conservatives falls from 8 points to 5. These numbers are illustrative only; the live correction is the table above.

Averaging across pollsters

Let \(s\) be the date of the forecast. The set \(J_s\) contains the latest poll from each pollster whose raw tables we hold, with fieldwork ending in the 28 days up to \(s\). Our reading for England is the unweighted mean of the corrected polls:

\[ \hat o_p = \frac{1}{|J_s|} \sum_{j \in J_s} \hat u_{j,p}. \tag{6} \]

Restore Britain is an exception. Not every pollster prompts for it, and for those that do not, its support sits inside Others. Averaging a zero from those polls would understate it, so its share is averaged over the pollsters that show it:

\[ \hat o_{\text{RB}} = \frac{1}{|J^{\text{RB}}_s|} \sum_{j \in J^{\text{RB}}_s} \hat u_{j,\text{RB}}, \qquad J^{\text{RB}}_s = \{\, j \in J_s : \hat u_{j,\text{RB}} > 0 \,\}. \tag{7} \]

Others is then whatever remains of 100, floored at zero. The rules for the window, one poll per pollster and equal weights, are the same as for the published-headline average the site also shows, so that the two averages differ only in what goes into them, not in how they are combined.2 The implementation is OurReading.AverageLatest in Elections.Core/Method/OurReading.cs.

From a reading to a lens

The seat model, the simulation and every page of the site work from the England headline average plus a lens: a set of amounts subtracted from each party's polled share (see Lenses: reading the same British polls several ways). Our reading enters the same way. Let \(H_p\) be the England average of the pollsters' published figures (the latest primary-grade poll per pollster in 28 days, equal weights). The lens adjustment is

\[ a_p = H_p - \hat o_p \quad \text{for } p \in \{\text{CON}, \text{LAB}, \text{LD}, \text{REF}, \text{GRN}, \text{RB}\}, \tag{8} \]

and applying it gives \(H_p - a_p = \hat o_p\) for those six parties, with the other parties left at their headline values; the result is then rescaled to 100. So the England shares that reach the seat model are our reading, up to that final rescaling, which is why the figure on the home page can differ by a tenth of a point from the plain average of the poll-by-poll readings on our method. This is ForecastService.AddOurReadingLens.

The reason for this indirection is practical. Expressing our reading as a lens means it runs through exactly the same seat model (uniform swing from the 2024 result; see From votes to Commons seats: UK swing models and their test) and the same simulation (Simulating UK polling error: where the chances come from) as every other scenario, so differences in seats between "our projection" and "polls as they stand" come from the vote shares alone.

Scotland and Wales

The lens has an empty adjustment in Scotland and Wales, so both are left as polled. The raw-answer correction was measured in England only, and when we tested correcting the Scottish and Welsh polls by their own past errors, it made them worse out of sample (see Scotland and Wales: sparse polls and no steady lean). Great Britain figures are the three nations combined.

How much to correct

Why not correct in full

The obvious estimator takes off the whole average miss, \(\lambda = 1\). That would be right if the average past miss were a perfect estimate of the next one. It is not. With four elections, \(\bar e^{\,raw}_p\) is itself a noisy estimate, and part of each past error is specific to that election and will not recur. Removing all of it means importing that noise into the forecast.

The standard remedy is to shrink the estimate towards a sensible default, here towards no correction at all. The idea goes back to Stein and was set out for applied work by Efron and Morris (1975): when many quantities are estimated from limited data, pulling each estimate part of the way towards a common value reduces total error, because the variance saved outweighs the bias introduced.

A simple model shows what is at stake. Suppose each election's raw error is a persistent bias plus election-specific noise, \(e^{\,raw}_{p,t} = \mu_p + \varepsilon_{p,t}\), with the noise independent across elections, mean zero and variance \(\sigma^2_p\). Estimating \(\mu_p\) from \(n\) past elections and removing \(\lambda\) of it leaves an expected squared error at the next election of

\[ \mathbb{E}\!\left[(e_{p,\text{next}} - \lambda \bar e_p)^2\right] = (1-\lambda)^2 \mu_p^2 + \sigma_p^2 + \lambda^2 \frac{\sigma^2_p}{n}, \tag{9} \]

which is minimised at

\[ \lambda^{*}_p = \frac{\mu_p^2}{\mu_p^2 + \sigma^2_p / n}. \tag{10} \]

The more stable the error relative to its size, the closer \(\lambda^*\) is to one; the noisier it is, the more we should shrink. As a rough illustration only, plugging the four raw errors into (10) gives a value near 0.9 for the Conservatives, whose raw error has been consistently negative, and nearer 0.4 for Labour, whose raw error has swung. With four observations these plug-in values are themselves very uncertain, and we do not use them. They show why a single value around one half is a reasonable compromise, and why a party-specific shrinkage is a natural extension we have not built.3

The grid and the rule

We tested five values, \(\lambda \in \{0, \tfrac14, \tfrac12, \tfrac34, 1\}\), each by leave-one-out (described in the next section). The rule for choosing was set before the tests were run: a correction must improve on the pollsters' published figures consistently out of sample, across elections, and not merely on average because of one good election. The table shows the results.

ShrinkageMean abs. errorMean abs. lead errorElections no worse than pollsters
0.001.996.01 of 4
0.251.725.22 of 4
0.50 (chosen)1.474.64 of 4
0.751.344.14 of 4
1.001.464.23 of 4
Live Share of the past raw-answer miss taken off (the shrinkage factor), leave-one-out. Computed from the database on 19 November 2026.

With no correction (\(\lambda = 0\)) the raw answers are clearly worse than the headlines: the pollsters' adjustments do real work on average. Full correction (\(\lambda = 1\)) has a good average but is worse than the pollsters at one election, 2017, when the raw-answer pattern was weakest; that is what overfitting looks like in a sample of four. Both \(\lambda = \tfrac12\) and \(\lambda = \tfrac34\) are no worse than the pollsters at all four elections, and \(\tfrac34\) has the lower average. We chose \(\tfrac12\) because moving from \(\tfrac12\) to \(\tfrac34\) does not improve consistently: it helps in three elections but makes 2017 clearly worse (an average miss of about 1.6 points against 1.3), leaving it level with the pollsters there rather than ahead. The same "improve consistently" rule that rejects full correction therefore rejects the step to three-quarters.

We should be candid that this is a judgement, not a theorem. On the average alone, \(\tfrac34\) would win. The choice of \(\tfrac12\) gives up a little average accuracy for a correction that is smaller, easier to defend, and less exposed to the one election in which the pattern nearly failed. The companion paper Why we start from British pollsters' raw answers rather than the published headline returns to \(\tfrac34\) as an open option.

Validation

Leave-one-out

A correction estimated from four elections and scored on the same four would flatter itself. We therefore use leave-one-out cross-validation (Stone, 1974): each election is predicted using a correction estimated only from the other three, exactly as if its result were unknown. For held-out election \(t\),

\[ c^{(-t)}_p = \lambda \cdot \frac{1}{|\mathcal{T}| - 1} \sum_{s \neq t} e^{\,raw}_{p,s}, \qquad \hat v^{(-t)}_{p,t} = \bar r_{p,t} - c^{(-t)}_p, \tag{11} \]

and the prediction is scored against \(v_{p,t}\) with (3).4 The comparison is the pollsters' published headlines of the same polls, \(\bar h_{p,t}\), which need no training. We also score the most obvious alternative: the headlines minus the other elections' average headline miss. Ours counts as better at an election when its mean absolute error is at least 0.05 points lower than the pollsters', and level when the two are within 0.05 of each other.

ElectionPollstersPollsters minus past missOursLead: pollstersLead: oursOurs vs pollsters
20152.131.751.79+7.1+5.5better
20171.582.541.29−4.5−3.3better
20190.790.620.64+2.7+1.9better
20242.191.822.18+7.8+7.8level
Mean1.671.681.475.54.63 better
Live Leave-one-out test: each election predicted with the correction estimated from the other elections only. Mean absolute error per party, England, points; lead = error in the Labour lead over the Conservatives (mean row: mean absolute). Computed from the database on 19 November 2026.

Our reading averaged 1.47 points per party against the pollsters' 1.67, was better at 3 elections and no worse at 4. The error in the Labour lead was also smaller on average. The margins are modest: a few tenths of a point per party. The 2024 row deserves attention. There our method ties the pollsters, and neither comes close to the result, because both overstated Labour by a wide margin. That miss is the subject of the companion paper.

The middle column is instructive. Correcting the headlines by their own past miss (the obvious fix, and the one most readers would reach for first) gains nothing on average and is much worse in 2017. The headline's error has not been stable enough to correct; the raw answers' error, at least for the Conservatives and the Reform family, has been.

Forward in time

Leave-one-out lets later elections inform the prediction of earlier ones, which no real forecaster could do. The forward test removes that advantage: each election from 2017 onwards is predicted using only the elections before it. It also lets us compare two ways of weighting the past, averaging all earlier elections (our method) or trusting only the most recent one.

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.

Over the three elections that can be tested this way, our method averaged 1.42 points per party against the pollsters' 1.52. The means here cover 2017, 2019 and 2024 only, so they are not comparable with the leave-one-out means, which also include 2015. Two results matter. Taking half of all earlier elections' misses does slightly better than taking half of the previous election's alone. And assuming the next election will be exactly as wrong as the last one ("all of previous election") is worse than simply believing the pollsters. Averaging is doing useful work.

The Labour figures in brackets are less comfortable. In the forward test our reading was further from Labour's result than the pollsters' headlines in 2019 and in 2024; in the leave-one-out test, only in 2024.

What the tests cannot show

Three limitations apply to both tests. The sample is four elections, and the forward test has three; a single unusual election can move every average in the tables. The shrinkage factor was chosen by looking at the leave-one-out results, so the leave-one-out score for \(\lambda = \tfrac12\) is not a clean out-of-sample estimate of its accuracy: the selection itself used all four elections. The forward test fixes \(\lambda\) at the chosen value and so shares that caveat. And the Conservatives were in government at all four elections; a pattern that held in those conditions may not hold under a different government or a different set of challengers.

Today's calculation

The polls

The table below lists the polls that go into today's reading: the latest from each pollster whose full tables are loaded, with fieldwork in the last 28 days. For each, it shows how England was obtained (the pollster's own England column, English regions summed by their weighted bases, or Great Britain minus Scotland and Wales at population shares), and the Labour and Conservative shares at three points: the raw answers, the pollster's published headline, and our reading.

PollsterFieldwork endEngland fromRawPublishedOurs
Find Out Now19 Nov 2026RegionSum27.7 / 20.329.7 / 20.926.7 / 22.2
YouGov16 Nov 2026Column28.0 / 24.027.0 / 24.027.0 / 26.0
More in Common15 Nov 2026PopulationShares30.2 / 23.730.2 / 26.329.2 / 25.7
Survation12 Nov 2026RegionSum31.5 / 21.631.1 / 22.530.5 / 23.5
Opinium6 Nov 2026RegionSum28.5 / 20.828.5 / 20.827.5 / 22.8
Live The polls behind our projection: the latest from each pollster with full tables in the last 28 days, England, Labour / Conservative. Computed from the database on 19 November 2026.

The pollsters in today's reading are Find Out Now, YouGov, More in Common, Survation and Opinium. Reading across a row shows how different pollsters' adjustments are. For some, the headline sits well away from the raw answers because the turnout and don't-know steps move a lot of support. Opinium is the limiting case: its raw table and its headline have the same party shares, because the only differences between them are a base filter and the removal of don't knows. For YouGov, the published headline is the output of a multilevel regression model, and the raw stage is the constituency-prompted question the model is built from (see From answers to headlines: the stage model of a published British poll). Our correction is the same for every pollster; what differs is how far each pollster's own adjustments had moved its figures first.

From the pollster average to our bet

The table below moves from the number most forecasts use, the average of all pollsters' published figures, to our bet in five steps, and splits the difference into three parts.

StepConLabReformLDGreen
1. Average of all pollsters' published figures (5 pollsters)22.929.322.110.69.8
2. Only the 5 pollsters with full tables22.929.322.110.69.8
3. Their raw answers22.129.222.010.810.3
4. Correction+2.0−1.0−0.6−0.1−0.3
5. Our bet24.128.221.410.710.1
Pollster mix (2 − 1)0.00.00.00.00.0
Undoing the pollsters' adjustments (3 − 2)−0.8−0.1−0.1+0.2+0.6
Our correction after rescaling (5 − 3)+2.0−1.0−0.6−0.1−0.3
Total (5 − 1)+1.2−1.1−0.7+0.2+0.3
Live From the pollster average to our projection, England, today. Computed from the database on 19 November 2026.

Write \(H_p\) for step 1, the average of all pollsters' published England figures. Step 2, \(H^{J}_p\), is the same average restricted to the pollsters in \(J_s\). Step 3, \(\bar r^{J}_p\), is those pollsters' raw answers. Step 5, \(\hat v_p\), is our bet. The total difference decomposes exactly:

\[ \underbrace{\hat v_p - H_p}_{\text{total}} \;=\; \underbrace{(H^{J}_p - H_p)}_{\text{pollster mix}} \;+\; \underbrace{(\bar r^{J}_p - H^{J}_p)}_{\text{undoing their adjustments}} \;+\; \underbrace{(\hat v_p - \bar r^{J}_p)}_{\text{our correction}}. \tag{12} \]

Pollster mix. Our reading can only use pollsters whose raw tables we read. If those pollsters' headlines already differ from the full average, part of our difference from the pollster average has nothing to do with our method. We show this separately so that it cannot be mistaken for a finding.

Undoing their adjustments. This is the effect of setting aside the pollsters' turnout and don't-know steps. For the Conservatives it was negative at every past election in our data, and for Labour positive: the pollsters' own steps have pushed the Conservatives up and Labour down, and starting from the raw answers reverses that.

Our correction. This is step 4 after rescaling and after the lens is applied to the headline average. It is close to the correction in step 4 but not identical, for the reasons given under (5) and (8).

For the Conservatives, the second and third parts usually pull in opposite directions: undoing the pollsters' adjustments takes the party down, and our correction puts it back up. The net effect is the difference between two larger movements, which is why our calls are modest. Our method does not so much overrule the pollsters as replace each pollster's own adjustment, which varies from pollster to pollster and from election to election, with a single adjustment estimated from how raw answers have actually missed. The companion paper makes that equivalence precise.

From our reading to our projection

When the projection is used

Our reading becomes the headline forecast, labelled "our projection", only when it rests on enough polls. The rule in ForecastService.AddOurReadingLens is that at least two pollsters must have raw tables loaded with fieldwork in the last 28 days, and the historical backtest must have run. If either fails, the site falls back to the published polls as they stand, and says so. One pollster is not enough: a single pollster's house effect could then masquerade as our correction.

Pollsters whose raw tables we cannot read (today, those in the full average but not in the list above) are not in our projection at all. Their headlines count in step 1 of the table above, but not in our reading. This is a cost of the approach: our reading rests on fewer polls than the published average.

When we say the polls are too high or too low

For each main party, the home page says whether we think the polls overstate it, understate it, or are about right. The call is made on the England difference between our bet and the published average, the "total" row in the table above. A party is called overstated if our bet is at least 0.3 points lower than the pollster average, understated if it is at least 0.3 points higher, and "about right" otherwise (BestBet.AgreeWithin in Elections.Core/Method/BestBet.cs). The threshold is a presentational choice rather than a statistical test. It stops us describing differences of one or two tenths, which are well within the rounding of individual polls, as calls. Each call is shown with its evidence: the published and raw misses at each past election, and, where it is large enough to matter, how much of the difference comes from the pollster mix rather than our correction.

What the seats inherit

The projection's England shares go through the same seat model and simulation as every other lens. Its uncertainty is not reduced by the correction: the ranges shown around it are built from the full spread of past polling errors (Simulating UK polling error: where the chances come from), because a correction that removes half the average miss does nothing about the variation around that average.

What we tried and did not adopt

Correcting the headlines by their average past miss. This is the obvious method and the benchmark we most wanted to beat. It gained nothing on average in the leave-one-out test and was much worse in 2017, and in the forward test it never beat the uncorrected headlines. The companion paper Why we start from British pollsters' raw answers rather than the published headline tests it alongside a blend of the two starting points.

Reallocating don't knows to their 2024 party. Some pollsters assign part of their don't knows to the party those respondents voted for last time. We tested doing this ourselves on the raw tables, returning between a quarter and all of each 2024 party's don't knows to it (\(r = 0.25\) to \(1.0\)). It narrowed the error in the Labour–Conservative lead a little but made the party-by-party error worse, and it no longer beat the pollsters at every election, so it failed the consistency rule. It is not used.

Full correction and three-quarters. As described above, \(\lambda = 1\) overfits, and \(\lambda = \tfrac34\) does not improve consistently on \(\tfrac12\).

Trusting only the last election. The forward test shows that assuming a repeat of the previous election's raw miss is worse than the pollsters' headlines. Averaging across elections is better.

Correcting Scotland and Wales. Correcting each nation's polls by its own past error made them worse out of sample, so both are left as polled.

Correcting Restore Britain. Restore Britain did not exist at any of the four elections, so there is no raw-answer history to learn from, and it is not linked to the UKIP, Brexit Party and Reform family. It receives no correction. Its share in our reading moves only through the rescaling in (5), and it is averaged only over the pollsters that prompt for it.

Limitations

Four elections are very few. The Conservative understatement in the raw answers has held every time, but its size has ranged from about 2 to about 7 points, and the forward test has only three elections to score. We regard the backtest as evidence that the method is reasonable, not as proof that it will beat the pollsters at the next election.

"Raw" is not identical across pollsters or across time. For some historical polls, the widest weighted table printed is already restricted: to people who say they are likely to vote, to people who might vote, or, in one 2024 poll, to the table after the turnout filter. Some raw stages were rebuilt from English regions. The historical set also differs slightly from the primary benchmark in Measuring British polling error in England, 2015–2024: it includes Number Cruncher's 2019 and 2024 final polls, which that benchmark treats as sensitivity-only because their England figures rest on estimated bases, and it lacks YouGov's 2024 final call, whose raw tables we do not hold.5 Each difference is small, but together they mean the "raw error" is an average over somewhat different things.

Almost every pollster weights to recalled 2024 vote before the stage we call raw. That weighting is part of what our correction learns about. If pollsters change their weighting schemes, the raw answers will lean differently and the historical correction may no longer fit.

The correction is estimated in England and applied only there. England holds most of the seats in Great Britain, but errors in Scotland and Wales are left as they are.

Finally, the method has not yet faced a live general election. It leads the site because it is the only approach we have found that beat the published figures out of sample; that is a claim about the past, and the next election will test it.

Conclusion

Our reading starts from the answers poll respondents actually gave, before each pollster's turnout and don't-know adjustments, and removes half of the average amount those answers have missed by in England since 2015. The case for it is empirical and specific. Raw answers have understated the Conservatives and overstated the Reform family at every election in our data, consistently enough to correct; the published headlines' errors have moved around too much for the same trick to work on them. Tested on 4 elections, each predicted without its own result, the method was no worse than the pollsters at 4 and better at 3. Tested forward in time, it was better on average than the pollsters and than any rule that trusts only the last election.

The gains are modest and the evidence is thin. The method would not have caught the large overstatement of Labour in 2024, and on Labour alone it has sometimes been further out than the pollsters. We lead with it because it has the better record, and we show the pollsters' own figures beside it, with the ranges from every past miss, because four elections cannot rule out that the next one will be different.

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.

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.

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. Where a pollster prints two versions of the question, we use the one behind its headline. For YouGov this is the constituency-prompted question; the standard question is stored for comparison only.↩

  2. Equal weights ignore sample size. That is deliberate: differences between pollsters' methods are much larger than the sampling error of any one poll, so weighting by sample size would mainly weight by house style.↩

  3. A party-specific \(\lambda_p\) would add five parameters estimated from four elections. Any such scheme would need to pass the same leave-one-out and forward tests before replacing the single value; it has not been tested.↩

  4. In the backtest the corrected shares are not rescaled to 100, whereas today's reading is. Because the main parties' corrections nearly cancel, the difference is a few hundredths of a point.↩

  5. This is why two slightly different figures for the 2024 Labour miss appear on the site. The primary benchmark (all six checked final polls) is used for the published-error tables and lenses; the raw-answer set (the six 2024 polls with both raw and headline tables) is used for our correction and its backtests.↩

How to cite

Polling Forecast (2026). Correcting the raw answers: the method behind our UK projection. Retrieved 20 November 2026 from https://www.pollingforecast.com/research/uk-our-reading. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like