What we don't know yet: limitations and open questions in our UK forecast
Abstract. The method rests on four elections and fewer than twenty final polls, sets aside pollster adjustments that were usually right about Labour, and did not foresee Labour's 2024 overstatement; its seat model, simulation and election-day model each lean on simple assumptions. This paper sets out every known weakness and, for each, the evidence or work that would address it.
A forecast that claims to correct the polls owes its readers an unusually clear account of where the correction might fail. This paper collects every weakness we know of in the Polling Forecast method, from the evidence base to the software that gathers the data. For each we say what the problem is, how large we believe it to be, and what evidence or work would reduce it. Some of these are limits of the data and cannot be fixed by effort alone; others are unbuilt parts of the method. We try to be clear which is which. The method itself is summarised in Forecasting a British general election when the polls are wrong: an overview.
A small evidence base
The correction at the centre of the method is estimated from 4 general elections (2015, 2017, 2019 and 2024) and 18 final polls whose raw tables survive. That is not many. The table below gives the raw answers' miss at each election, with the same polls' published headline in brackets, and the number of polls behind each row.
| Election | Polls with tables | Con | Lab | LD | Reform | Green |
|---|---|---|---|---|---|---|
| 2015 | 2 | −6.6 (−5.7) | +1.6 (+1.5) | +2.0 (+1.8) | +0.9 (+0.8) | +1.1 (+0.9) |
| 2017 | 4 | −2.1 (+0.6) | −1.5 (−3.9) | +0.7 (+0.7) | +2.1 (+2.1) | +1.0 (+0.6) |
| 2019 | 6 | −3.3 (−1.6) | +1.8 (+1.0) | −0.5 (−0.5) | +1.1 (+0.8) | +0.4 (0.0) |
| 2024 | 6 | −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) |
Three features of this evidence limit what it can bear. The first is sheer size. With four elections, the standard error of a party's mean miss is roughly half the between-election standard deviation of that miss, and for the Conservatives the raw miss has ranged from about 2 to about 7 points. A second concern is that the polls within an election are not independent observations. They share the same electorate, the same late campaign and often similar weighting targets, so eighteen polls carry much less information than eighteen independent draws; the effective sample is closer to four elections than to eighteen polls, in the spirit of Kish's (1965) design effect. Shirani-Mehr et al. (2018) show that much of the error in election polls is shared by all the polls of a given election, so that total error is considerably larger than the reported sampling margin suggests.
The second feature is the uneven spread of polls. 2015 rests on two polls, because most pollsters' final tables for that year could not be recovered. Many other final polls from 2015–2024 are missing for the same reason: pages that have been taken down, archives that cannot be reached, or banners that merge Wales with an English region so that England cannot be isolated. The polls that survive are not a random sample of the polls that were published, and if the pollsters who keep their tables online differ systematically from those who do not, the correction inherits that difference.
The third is that the raw-answer set is not the same as the primary benchmark of published final polls. It includes Number Cruncher's final polls of 2019 and 2024, whose England figures are rebuilt on an estimated basis and are used only in sensitivity analysis for the headline benchmark, and it lacks YouGov's 2024 final poll, which is in the headline benchmark but whose raw stage is not stored. Each set is internally consistent, since both columns of the backtest describe the same polls, but the two sets should not be read as interchangeable.
There is also a deeper limit. All four elections were held with the Conservatives in government. If the lean in the raw answers reflects something about who answers polls when a particular party is in office, rather than something stable about the Conservative electorate, the correction may be pointing the wrong way at the next election.
What would address it. More elections, above all. Bringing 2010 into the benchmark is already planned, and 2005 is held as a legacy check, though older polls often lack the tables needed. Recovering the missing 2015–2024 final polls by hand from archive copies would enlarge each year's sample and reduce the selection problem. Neither can make four elections into forty, and the honest position is that the method's estimates will remain noisy for some time.
The 2024 Labour miss and the Labour weak spot
The largest single failure in the record is 2024. In England the published final polls' Labour error was +4.3 points and the raw answers' +6.1 (positive meaning overstated). Our method, tested as if it had been run before the election, did not fix this. The forward test predicts each election only from those before it:
| Election | Pollsters | Ours: ½ × mean of all earlier | ½ × previous election | All of previous election |
|---|---|---|---|---|
| 2017 from 2015 | 1.58 (−3.9) | 1.17 (−2.3) | 1.17 (−2.3) | 2.03 (−3.1) |
| 2019 from 2015, 2017 | 0.79 (+1.0) | 0.90 (+1.8) | 1.15 (+2.6) | 1.44 (+3.3) |
| 2024 from 2015, 2017, 2019 | 2.19 (+4.8) | 2.18 (+5.7) | 2.00 (+5.2) | 1.43 (+4.2) |
| Mean | 1.52 | 1.42 | 1.44 | 1.64 |
The bracketed Labour errors show the problem plainly. Every approach leaves Labour several points too high in 2024. Taking the comparison wider, across both possible starting points and several amounts of correction, does not change that:
| Approach | Leave-one-out | Forward | Lead error (LOO) | Lead error (fwd) | Labour 2024 (fwd) |
|---|---|---|---|---|---|
| Published figures, no correction (the pollsters) | 1.67 | 1.52 | 5.5 | 5.0 | +4.8 |
| Published figures minus ½ their past miss | 1.57 | 1.57 | 5.3 | 5.6 | +5.1 |
| Published figures minus all their past miss | 1.68 | 1.98 | 5.6 | 6.3 | +5.3 |
| Raw answers, no correction | 1.99 | 1.84 | 6.0 | 5.2 | +6.1 |
| Raw answers minus ½ their past miss (ours) | 1.47 | 1.42 | 4.6 | 4.7 | +5.7 |
| Raw answers minus ¾ their past miss | 1.34 | 1.43 | 4.1 | 4.7 | +5.6 |
| Raw answers minus all their past miss | 1.46 | 1.69 | 4.2 | 4.6 | +5.4 |
| Half and half, each minus ½ its past miss | 1.51 | 1.50 | 5.0 | 5.2 | +5.4 |
The reason is not mysterious. Before 2024, the published polls' Labour error had changed sign from one election to the next (overstated in 2015 and 2019, understated in 2017) and averaged close to zero, and the raw answers' Labour lean was modest. An estimator that learns the size of past misses from 2015–2019 could not have predicted a Labour overstatement of 2024's size. Catching it would have required a reason, not a record: a model of differential turnout, or of where don't knows would go. In the 2024 tables we hold, don't knows were concentrated among people who had voted Conservative in 2019 (at Ipsos, 15.2% of 2019 Conservatives against 9.0% of 2019 Labour voters), so setting them aside removed Conservatives from the sample. That is a candidate mechanism, not a demonstrated one.
The Labour weak spot is more specific than the 2024 miss and, in one respect, more worrying. The pollsters' own turnout and don't-know adjustments lowered Labour at every one of the four elections, and in three of them (all but 2017) that was the right direction. Our method starts from the raw answers, so it sets those adjustments aside and replaces them with a Labour correction of only −1.0 points. On Labour alone, in the forward test, our method was further from the result than the pollsters in 2019 and 2024. The overall gain comes mainly from the Conservatives and the Reform family, where the raw answers' lean has been steadier. If the next election resembles 2024 more than it resembles 2017, our reading is likely to overstate Labour by more than the pollsters do. The lens table on the site shows what a repeat of the 2024 miss would do to the seat picture:
| Lens | Group | Lab | Con | Reform | LD | Seats changed |
|---|---|---|---|---|---|---|
| Our projection | Our method | 317 | 154 | 36 | 72 | – |
| Polls as they stand | Polls | 329 | 143 | 37 | 72 | 14 |
| Average miss | Past misses | 319 | 167 | 32 | 70 | 23 |
| 2015 miss | Past misses | 279 | 210 | 30 | 60 | 60 |
| 2017 miss | Past misses | 395 | 117 | 13 | 72 | 85 |
| 2019 miss | Past misses | 318 | 154 | 36 | 72 | 4 |
| 2024 miss | Past misses | 273 | 200 | 42 | 70 | 70 |
| Pollster records | Pollsters | 308 | 169 | 32 | 70 | 16 |
| Best for Labour | Pollsters | 347 | 134 | 34 | 72 | 34 |
| Best for the Conservatives | Pollsters | 325 | 150 | 36 | 72 | 11 |
| Best for Reform UK | Pollsters | 318 | 136 | 51 | 73 | 25 |
What would address it. Two tests are open. The first is a Labour-specific rule that keeps the pollsters' own adjustment for Labour, or blends it with the raw answers, while correcting the other parties from the raw answers. This has been proposed but not run, and it carries an obvious risk of fitting the rule to the one party where it happens to help. The second, better, route is a mechanism rather than a patch: our own turnout step and don't-know step, each estimated from data outside the polls and each tested leave-one-out. A don't-know step that returned don't knows to their previous party was tested at several strengths and did not beat the pollsters out of sample, so it is not used. A turnout step has not been built (see below).
What "raw" means differs between pollsters
The method treats each poll's weighted answers from everyone asked, before any turnout or don't-know step, as its raw reading. This is a convenient definition, but it is not a uniform object across pollsters.
Almost every pollster's first stage is already weighted, usually to recalled vote at the previous election as well as to demographics; only Ipsos prints genuinely unweighted figures, and among the historical polls only Qriously's 2019 data were unweighted at source. So "raw" means "before the turnout and don't-know steps", not "as collected". Some pollsters have already filtered their widest printed base: Survation's excludes those who say they are certain not to vote, and Opinium's base is people who "might vote". For Opinium there is no turnout weight between that table and the headline, so its raw answers equal its published headline on the parties and our correction applies to effectively the same numbers as its headline. YouGov's current headline is the output of its MRP model, and its raw stage is the constituency-prompted question from which that model is built. Some pollsters, such as Freshwater Strategy, print every table on a likely-voter base, and cannot be read at the raw stage at all.
The consequence is that the correction, estimated from a mix of pollsters at past elections, is applied to a different mix of pollsters today, whose raw stages are not quite the same quantity. The table below shows the polls currently behind our projection, with each pollster's raw, published and corrected figures for England.
| Pollster | Fieldwork end | England from | Raw | Published | Ours |
|---|---|---|---|---|---|
| Find Out Now | 19 Nov 2026 | RegionSum | 27.7 / 20.3 | 29.7 / 20.9 | 26.7 / 22.2 |
| YouGov | 16 Nov 2026 | Column | 28.0 / 24.0 | 27.0 / 24.0 | 27.0 / 26.0 |
| More in Common | 15 Nov 2026 | PopulationShares | 30.2 / 23.7 | 30.2 / 26.3 | 29.2 / 25.7 |
| Survation | 12 Nov 2026 | RegionSum | 31.5 / 21.6 | 31.1 / 22.5 | 30.5 / 23.5 |
| Opinium | 6 Nov 2026 | RegionSum | 28.5 / 20.8 | 28.5 / 20.8 | 27.5 / 22.8 |
What would address it. A correction estimated per pollster, or per type of raw stage, would remove the mismatch, but at four elections the per-pollster records are too short (between none and three checked final polls for the current pollsters) to support it. A better medium-term answer is to harmonise the raw stage ourselves where the tables allow: rebuilding a common base from likelihood-to-vote tables and past-vote cross-tabulations. Every pollster's past-vote table has gaps, however, most often the previous election's non-voters, so this would be approximate.
Coverage: England only, and only some pollsters
The raw-answer correction is measured and applied in England only. Scotland and Wales are left as polled in our projection, both because the raw-stage evidence for those nations is too thin and because correcting their polls by their own past miss made them worse (see Scotland and Wales: sparse polls and no steady lean). England holds 543 of the 632 seats in Great Britain, so this matters less for the seat total than it might, but it does mean that our projection and the published polls cannot differ in Scotland or Wales.
Our projection also uses only pollsters whose full tables we can read. At present that is 5 pollsters (Find Out Now, YouGov, More in Common, Survation and Opinium); pollsters without readable tables, or without a reader written for their format, contribute nothing, although they are in the published average we compare against. Part of any gap between our figure and the published average therefore comes from the choice of pollsters rather than the correction, and the site shows that part separately. If the pollsters we can read happen to share a house effect, our figure inherits it.
A further limit is that the correction is additive: it removes a fixed number of points from each party whatever its current level. It was learned while the Reform family polled between about 2 and 15 per cent in England, and the Greens and Liberal Democrats at lower shares than some current polls show. If polling error scales with a party's share, a fixed correction will be too small for a party polling well above its historical range. The site's 80% vote bands for the Greens and Liberal Democrats are probably too narrow for the same reason.
What would address it. Readers for more pollsters' formats, where the pollsters publish tables at all; a raw-stage error record for Scotland and Wales, which needs nation-only polls with full tables; and a test of proportional against additive correction once there is more than one election at which the challengers polled at today's levels.
From votes to seats
The seat model is uniform swing within each nation (Butler and Stokes, 1969). In a backtest from the notional 2019 result on the 2024 boundaries, given the true 2024 nation shares, it called 479 of 631 seats correctly, against 461 for a blend of uniform and proportional swing and 445 for proportional swing. That is the best of the three, but it is not good. All three models under-called Labour badly: 329 seats for uniform swing, 315 for the blend and 290 for proportional swing, against the 411 Labour actually won. The error arises because Labour's vote rose least where it was already strong and most where it was needed, and because anti-Conservative tactical voting concentrated the Liberal Democrat and Labour votes where each could win. Uniform swing cannot see either effect. It also adds new parties such as Restore Britain evenly across every seat in a nation, which is certainly wrong in detail.
It follows that even a perfect reading of the national vote would produce a seat projection with substantial error, and that this error has recently had a direction. Our seat ranges come from the simulation's noise, not from a model of how the swing varies from seat to seat, so they do not represent this structural under-call.
We have loaded forty census and deprivation variables for every seat, and constituency estimates of the 2016 referendum vote, but none of them is used in the seat model yet.
What would address it. A seat model that uses each seat's demographics and the tactical situation there. The test it must pass is set in advance: it must beat 479 of 631 on the same 2019-to-2024 backtest before it replaces uniform swing, and ideally it should also be tested on 2015-to-2017 and 2017-to-2019, where boundaries did not change. Multilevel regression and post-stratification (MRP), as developed for British elections by Hanretty, Lauderdale and Vivyan (2016) and Lauderdale et al. (2020), is the natural candidate. We have built nothing of this kind yet, we do not use other organisations' MRP estimates, and stand-alone MRP releases are excluded from our polling averages. We make no claim about how an MRP model would do on our backtest. The published MRPs' record at past elections, and each one since 2024 against our projection, is in MRPs and mega-polls in Britain: what they measure, how they have done, and how we treat them.
Uncertainty: the simulation and the election-day model
The simulation replays the joint error pattern of one of four past elections in each run, scaled by a standard normal factor, adds one point of independent national noise per party and seat-level noise of \(3\sqrt{s/25}\) points for a party on share \(s\). Its main limitation is the small number of patterns. Four patterns span only four directions in the space of possible errors; a miss of a new shape (Reform understated rather than overstated, say) can arise only from the independent noise or from a negatively scaled pattern, and a miss much larger than any of the four is improbable by construction, since the scale factor rarely exceeds two. The simulation also treats the error patterns of the published final polls as the relevant uncertainty, while the centre of the forecast is our corrected reading, whose past errors were somewhat smaller and differently shaped. We think this makes the ranges cautious rather than overconfident, but it is an assumption, not a result.
The seat-level noise is independent from seat to seat. In reality local deviations are correlated within regions and among seats of a similar type, and a model without that correlation understates the chance of many seats moving together. The noise settings themselves are fixed values, not quantities estimated in any test we report.
The election-day model rests on history in a different way. It was fitted to GB polls at ten elections from 1987 to 2024, and it assumes that the next parliament will behave like those: the Conservatives, Labour and the Liberal Democrats drawn part of the way back to their previous result, the challengers not. The evidence for the challengers is only eight party-elections. The current parliament, with several parties polling in the twenties and new parties entering, has no close precedent in that record. The model is also fitted to Great Britain as a whole and applied within each nation, pulling each nation's parties towards that nation's 2024 result, which assumes that the same pull operates everywhere.
| Days out | Elections | Weight fitted here | Weight on curve | Polls MAE | Last result MAE | Ours MAE | Ours better | Ours RMSE |
|---|---|---|---|---|---|---|---|---|
| 30 | 7 | 0.81 | 0.76 | 2.72 | 4.61 | 2.58 | 4 of 7 | 3.30 |
| 61 | 7 | 0.63 | 0.67 | 4.15 | 4.61 | 3.71 | 4 of 7 | 4.69 |
| 91 | 7 | 0.58 | 0.64 | 4.48 | 4.61 | 4.00 | 4 of 7 | 5.04 |
| 182 | 7 | 0.50 | 0.62 | 4.71 | 4.61 | 4.03 | 4 of 7 | 5.58 |
| 273 | 7 | 0.71 | 0.62 | 3.59 | 4.61 | 3.26 | 3 of 7 | 4.53 |
| 365 | 7 | 0.71 | 0.62 | 4.09 | 4.61 | 3.73 | 5 of 7 | 4.82 |
| 456 | 7 | 0.72 | 0.62 | 4.32 | 4.61 | 4.17 | 3 of 7 | 5.01 |
| 548 | 7 | 0.64 | 0.62 | 4.65 | 4.61 | 4.41 | 4 of 7 | 5.24 |
| 639 | 7 | 0.61 | 0.62 | 4.88 | 4.61 | 4.30 | 4 of 7 | 5.22 |
| 730 | 7 | 0.71 | 0.62 | 4.84 | 4.61 | 4.48 | 4 of 7 | 5.60 |
| 913 | 6 | 0.56 | 0.62 | 5.77 | 4.48 | 5.10 | 5 of 6 | 6.02 |
| 1095 | 5 | 0.54 | 0.62 | 5.40 | 4.47 | 5.00 | 3 of 5 | 6.56 |
What would address it. For the simulation: more elections, which enlarge the set of patterns; a test of the simulated ranges against the outcomes at past elections (how often the result fell inside the 80% range); and seat noise with a regional component. For the election-day model: 2029's result will be the first real out-of-sample test, and until then the leave-one-election-out scores in the table above are the only evidence.
Scotland and Wales
The Scottish and Welsh records are much thinner than England's. Scotland's benchmark has fifteen nation-only final polls across the four elections; Wales has four, all from YouGov, one per election, and the 2024 Welsh poll was taken about a month before polling day because YouGov's later Welsh figure came from its MRP model. Several final polls in each nation could not be retrieved. The Scottish and Welsh columns of GB polls are much less accurate than the nation polls (their misses in Scotland ran at 2.6 to 4.7 points against 1.9 to 2.5 for the nation polls), so they are kept for comparison only.
| Nation | Election | Polls | Error by party |
|---|---|---|---|
| Scotland | 2015 | 3 | SNP +0.3, Lab 0.0, Con +0.6, LD −1.6, Green +0.5, Reform −0.3 |
| Scotland | 2017 | 3 | SNP +4.1, Lab −0.9, Con −2.9, LD −1.2, Green +0.8, Reform −0.2 |
| Scotland | 2019 | 3 | SNP −1.3, Lab −1.6, Con +2.2, LD +0.6, Green +0.2, Reform −0.4 |
| Scotland | 2024 | 6 | SNP +2.4, Lab −0.5, Con +1.0, LD −1.6, Green −0.6, Reform −0.7 |
| Scotland | Leave-one-out | as polled 2.10, corrected 2.25 (mean abs. error) | |
| Wales | 2015 | 1 | Lab +2.1, Con −2.2, Plaid +0.9, LD +1.5, Reform −1.6, Green −0.6 |
| Wales | 2017 | 1 | Lab −3.0, Con +0.4, Plaid −1.4, LD +0.5, Reform +3.0, Green −0.3 |
| Wales | 2019 | 1 | Lab −0.9, Con +0.9, Plaid +0.1, LD 0.0, Reform −0.4, Green 0.0 |
| Wales | 2024 | 1 | Lab +8.0, Con −0.2, Plaid −2.8, LD −1.5, Reform −4.0, Green −0.7 |
| Wales | Leave-one-out | as polled 1.77, corrected 2.31 (mean abs. error) |
Neither nation shows a steady lean that a correction could exploit: the SNP was overstated in 2017 and 2024 but understated in 2019. Correcting by the average past miss made both nations' polls worse in leave-one-out tests, so our projection leaves them as polled. Between elections, Scottish and Welsh polls are sparse enough that the averages use a 120-day window, and when a nation has none in that window its 2024 result is moved by the GB change. The nation averages also currently draw on every nation poll stored, including some transcribed from secondary sources that have not yet been checked against the pollster's own tables; our evidence rules reserve such polls for sensitivity analysis, and the averaging code does not yet enforce that.
What would address it. Recovering the missing nation-only final polls; checking every current Scottish and Welsh poll against its source table before it enters the average; and, in time, a raw-stage record for the nations comparable with England's.
Steps not yet built, and the data behind those that are
A turnout step. Nobody publishes voting intention by likelihood to vote in a form that would let us rebuild a pollster's turnout step exactly. Our own turnout step would therefore need outside calibration, and the natural source is the British Election Study, whose validated-vote data record whether respondents actually voted. With it, turnout could be modelled by past vote and age and tested leave-one-out on the historical finals like every other step. This is planned but not started. It is the most promising route to a mechanism for 2024's Labour miss, if differential turnout was part of it.
Data quality in automatic ingestion. New polls are found and loaded automatically once a day from the pollsters' own sites, read by a reader written for each pollster's format and stored only if every arithmetic check passes. The checks are strict, but they test internal consistency, not truth. When a file belongs to a poll already in the database, the England headline rebuilt from its tables must match the hand-checked figures. When the poll is new, those figures are themselves created from the same tables, so there is no independent check. A reader that picked the wrong table, one that happened to add up, could pass. Layouts do vary in ways that matter: two Opinium files printed a headline table without the SNP and Plaid rows (it summed to about 97 per cent), and the reader had to be taught to recognise and skip it. England figures rebuilt from population shares, where a pollster prints no weighted bases, are accepted within 0.6 points; YouGov's whole-number tables within 0.75. Pollsters change their file layouts without notice, and a file that no reader recognises is logged but not loaded, so coverage can fall without anyone noticing.
What would address it. An independent check for automatically added polls, such as a comparison with the headline the pollster states in its release; a periodic audit of stored tables against their source files, which are all kept; and an alert when a pollster that normally publishes weekly has had nothing loaded for a fortnight.
Choices made by judgement
Several choices in the method were made by judgement rather than by an optimisation, and a reader should know which.
The most important is the shrinkage factor \(\lambda\) = 0.5. The rule we adopted before testing was that a correction must improve consistently out of sample, not win on average because of one election. The table below shows every value tried.
| Shrinkage | Mean abs. error | Mean abs. lead error | Elections no worse than pollsters |
|---|---|---|---|
| 0.00 | 1.99 | 6.0 | 1 of 4 |
| 0.25 | 1.72 | 5.2 | 2 of 4 |
| 0.50 (chosen) | 1.47 | 4.6 | 4 of 4 |
| 0.75 | 1.34 | 4.1 | 4 of 4 |
| 1.00 | 1.46 | 4.2 | 3 of 4 |
Taking off three-quarters of the past raw miss scored better on the leave-one-out average, matched a half in the forward test and takes more off Labour. It was not chosen because at 2017 it was worse than a half (a mean miss per party of 1.61 against 1.29) and marginally worse than the pollsters' own 1.58, a difference within the 0.05-point margin we treat as "no worse". Applying the consistency rule to a choice between two corrections that both pass it against the pollsters is itself a judgement. We think it is the prudent one with four elections, and we note that the choice is close.
Other settings were chosen rather than estimated: the 28-day window and equal pollster weights in the averages; the 0.05-point margin for "no worse"; the 0.3-point threshold below which we say our figure agrees with the polls; the minimum of two pollsters with full tables before our reading becomes the headline; treating UKIP, the Brexit Party and Reform as one family for error; the national and seat noise in the simulation; and the weights and thresholds of the pollster ratings. Each is defensible, none is the product of a formal search, and a different reasonable choice would change some published figures.
What would address it. Sensitivity analysis published alongside each choice (the shrinkage table is the model for this), and, for the settings that matter most, pre-registered rules for when they will be revisited.
Summary
The table gathers the main limitations and what would address each.
| Limitation | What would address it |
|---|---|
| Four elections, 18 final polls with raw tables | Bring 2010 into the benchmark; recover missing final polls from archives |
| 2024 Labour overstatement not predicted by any variant | A mechanism: our own turnout step (British Election Study validated vote) and a better don't-know step, each tested leave-one-out |
| Labour weak spot: pollsters' adjustments set aside | Test keeping or blending the pollsters' own adjustment for Labour, guarding against fitting to one party |
| "Raw" differs between pollsters | Harmonise raw stages from likelihood-to-vote and past-vote tables; per-pollster corrections once records allow |
| Correction measured in England only | A raw-stage error record for Scotland and Wales |
| Pollsters without full tables excluded | More readers; pollsters publishing their tables |
| Additive correction learned at lower shares | Test proportional correction as evidence at current shares accumulates |
| Uniform swing under-calls Labour, ignores tactical voting | A demographic and tactical seat model that beats 479 of 631 on the 2019–2024 backtest |
| Simulation built on four error patterns, independent seat noise | More elections; coverage check of past ranges; regional seat noise |
| Election-day model depends on 1987–2024 history | The 2029 result; continued out-of-sample checks |
| Sparse Scottish and Welsh data | Recover missing nation polls; check every nation poll against its source |
| No MRP | Build and backtest one against the same seat benchmark |
| Automatic ingestion checks consistency, not truth | Independent check for new polls; audits against stored source files |
| Shrinkage of ½ and other settings chosen by judgement | Published sensitivity analyses; pre-set rules for revision |
Table: Known limitations of the method and the evidence or work that would address each.
None of these limitations is hidden on the site, and several are displayed next to the forecast they qualify. The method's claim is narrow: over the elections for which the evidence exists, reading the raw answers and correcting them modestly was closer to the result than the published polls. Whether that holds at the next election is the real test, and we will publish the result either way.
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 rescaled (13 October). The ratings no longer assume an unproven pollster is a poor one (Rating the British pollsters: accuracy, the Labour–Conservative lead and openness).
- Tactical voting (16 October) was tested and kept out of the projection; it is a reader's tool (From votes to Commons seats: UK swing models and their test).
- 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, Right winner, wrong size: how every poll overstated Labour's lead in 2024.
References
British Election Study (Fieldhouse, E. et al.), British Election Study Internet Panel. www.britishelectionstudy.com
Butler, D. and Stokes, D. (1969). Political Change in Britain. London: Macmillan.
Hanretty, C., Lauderdale, B. and Vivyan, N. (2016). Combining national and constituency polling for forecasting. Electoral Studies, 41, 239–243.
Kish, L. (1965). Survey Sampling. New York: Wiley.
Lauderdale, B. E., Bailey, D., Blumenau, J. and Rivers, D. (2020). Model-based pre-election polling for national and sub-national outcomes in the US and UK. International Journal of Forecasting, 36(2), 399–413.
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.
How to cite
Polling Forecast (2026). What we don't know yet: limitations and open questions in our UK forecast. Retrieved 20 November 2026 from https://www.pollingforecast.com/research/uk-limitations. Live figures change as new polls arrive; the date above is when the text was last revised.


