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

Lenses: reading the same British polls several ways

Abstract. A lens is a vector of assumed polling errors subtracted from today's nation-level shares, followed by a rescale to 100. We show our projection beside the published figures, five past-miss scenarios, three pollster-based readings and the reader's own, and we do not average them, because only one of them has passed an out-of-sample test and the spread between the rest is the useful information.

Polling Forecast · updated 9 November 2026 · 18 min read · lenses, polling error, scenarios, seats at risk, pollster house effects, what-if

What has changed since this paper was written

Updated 9 November 2026. The rest of the paper is as first published, except where a sentence below says it was corrected.

The problem a lens solves

Every forecast of a British general election starts from the same raw material: a few dozen voting-intention polls. What separates one forecast from another is mostly what it assumes about how wrong those polls are. Most forecasts make that assumption silently. They treat the published average as unbiased and add a band for sampling error, or they apply a house correction that the reader never sees.

Silence is hard to defend given the record. Final polls have missed at every general election since 2015, and not only by the margin that sampling error would explain (Sturgis et al., 2016; Prosser and Mellon, 2018). The misses also differ from one election to the next in size and, for some parties, in direction. Taken across many countries and decades, polling error has not been getting steadily worse (Jennings and Wlezien, 2018), but no one has shown that it is predictable enough to correct away. The honest position is that several readings of today's polls are defensible, and that the choice between them changes the seat outcome a great deal.

Polling Forecast makes that choice explicit. We call each reading a lens. A lens takes the same polling averages and applies one stated assumption about polling error. Every lens then goes through the same seat model (From votes to Commons seats: UK swing models and their test) and, when needed, the same simulation (Simulating UK polling error: where the chances come from), so any difference between two lenses comes only from the assumption. This paper defines a lens formally, describes each one the site offers, explains how lenses are combined into a scenario set and a count of "seats at risk", and sets out why we show them side by side rather than averaging them.

What a lens is

Intuition

A lens answers one question: by how many points are the polls wrong about each party? If we believe the polls overstate Labour by one point, the lens takes one point off Labour. Because vote shares must add up to 100, the shares are then rescaled. Nothing else changes. The same seat model, the same baseline and the same simulation settings are used whatever the lens.

The sign convention follows the one used everywhere on the site. Polling error is poll minus result, so a positive error means the polls overstated a party. A lens stores the error it assumes, and subtracts it:

\[ \text{adjusted share} = \text{polled share} - \text{assumed error} \]

Formal definition

Let \(n \in \{\text{England}, \text{Scotland}, \text{Wales}\}\) index the three nations of Great Britain and \(p\) index the parties modelled in nation \(n\) (Conservative, Labour, Liberal Democrat, Reform UK, Green, Restore Britain and Others everywhere, plus the SNP in Scotland and Plaid Cymru in Wales). Let \(v_{n,p}\) be today's polling average for party \(p\) in nation \(n\), in percent, after parties not modelled in that nation have been folded into Others. A lens \(L\) is a set of adjustment vectors \(a^{L}_{n}\), one per nation, where \(a^{L}_{n,p}\) is the error in points that the lens assumes the polls are making about party \(p\) in nation \(n\). A party with no entry has \(a^{L}_{n,p} = 0\); in particular no lens adjusts Others directly.

Applying the lens takes two steps. First subtract the adjustment and clip at zero:

\[ \tilde{v}^{L}_{n,p} = \max\!\left(0,\; v_{n,p} - a^{L}_{n,p}\right) \tag{1} \]

Then rescale so that the nation's shares add up to 100:

\[ v^{L}_{n,p} = 100 \times \frac{\tilde{v}^{L}_{n,p}}{\sum_{q} \tilde{v}^{L}_{n,q}} \tag{2} \]

Equations (1) and (2) are exactly what Lens.Apply in Elections.Core/Lenses/Lens.cs does, through the Normalised method on the share vector (negative values set to zero, then everything scaled to a total of 100).

The rescale in (2) matters more than it looks. If the adjustments in a lens sum to zero across parties, the adjusted shares already total 100 and the rescale does nothing. If they do not, every party moves, including those the lens did not touch. An illustration with round, invented numbers: suppose England polls at Conservative 22, Labour 27, Liberal Democrat 11, Reform 23, Green 12, Restore Britain 4 and Others 1, and a lens assumes the polls understate the Conservatives by 3 and overstate Labour and Reform by 1 each. After step (1) the shares are 25, 26, 11, 22, 12, 4 and 1, which total 101. Step (2) scales each by 100/101, giving 24.75, 25.74, 10.89, 21.78, 11.88, 3.96 and 0.99. The Liberal Democrats, Greens, Restore Britain and Others were not adjusted, yet each loses about 1% of its share. In practice this effect is small, because past errors roughly cancel across parties, but it means a lens's stated adjustment and its net effect on a party are not always identical.

Per-nation adjustments

The site predicts all of Great Britain, but England is where polling error can be measured most cleanly: the SNP and Plaid Cymru do not stand there, and England holds 543 of the 632 GB seats (see Measuring British polling error in England, 2015–2024). Most of the evidence behind the lenses is therefore England's. Scotland and Wales have their own, shorter records, built from Scotland-only and Wales-only final polls at the same four elections (Scotland and Wales: sparse polls and no steady lean).

A lens handles this with two fields. Adjustment holds the England vector. NationAdjustment optionally holds a separate vector for Scotland and for Wales. The rule in AdjustmentFor is simple:

\[ a^{L}_{n} = \begin{cases} a^{L,\text{own}}_{n} & \text{if } n \neq \text{England and the lens has an entry for } n,\\[2pt] a^{L}_{\text{England}} & \text{otherwise.} \end{cases} \tag{3} \]

So a lens with no nation-specific entries applies England's adjustment to the GB-wide parties in all three nations. The SNP and Plaid are left alone in that case, because England's vector has no entry for them.

Keeping Scotland and Wales as polled

Some lenses rest on evidence measured only in England. Our projection uses a correction estimated from England raw samples, and the pollster-record lens uses each pollster's England final-poll record. Applying those England corrections to Scottish and Welsh polls would be an extrapolation with nothing behind it. Worse, the one test we have points against correcting those nations at all: correcting Scotland's and Wales's polls by their own average past miss made them worse in a leave-one-out test (the bottom row for each nation in the table below).

NationElectionPollsError by party
Scotland20153SNP +0.3, Lab 0.0, Con +0.6, LD −1.6, Green +0.5, Reform −0.3
Scotland20173SNP +4.1, Lab −0.9, Con −2.9, LD −1.2, Green +0.8, Reform −0.2
Scotland20193SNP −1.3, Lab −1.6, Con +2.2, LD +0.6, Green +0.2, Reform −0.4
Scotland20246SNP +2.4, Lab −0.5, Con +1.0, LD −1.6, Green −0.6, Reform −0.7
ScotlandLeave-one-outas polled 2.10, corrected 2.25 (mean abs. error)
Wales20151Lab +2.1, Con −2.2, Plaid +0.9, LD +1.5, Reform −1.6, Green −0.6
Wales20171Lab −3.0, Con +0.4, Plaid −1.4, LD +0.5, Reform +3.0, Green −0.3
Wales20191Lab −0.9, Con +0.9, Plaid +0.1, LD 0.0, Reform −0.4, Green 0.0
Wales20241Lab +8.0, Con −0.2, Plaid −2.8, LD −1.5, Reform −4.0, Green −0.7
WalesLeave-one-outas polled 1.77, corrected 2.31 (mean abs. error)
Live Scotland and Wales: final polls of each nation minus its result, points. Computed from the database on 19 November 2026.

For these lenses the code calls KeepNationsAsPolled. It sets an empty adjustment vector for Scotland and for Wales, so by equation (3) those nations take no adjustment and their polls pass through unchanged apart from the trivial rescale. The method only does this when a nation actually has an error record loaded. Without one, the lens falls back to England's adjustment, which is the behaviour before the Scotland and Wales evidence existed.

The catalogue

The site groups lenses under four headings, held in LensGroups: Our method, Polls, Past misses and Pollsters. The grouping is not cosmetic. As section 5 explains, the past-miss group alone defines "seats at risk".

Our projection (the default)

The headline lens is our projection, code our-reading, the only lens in the Our method group. It starts from what respondents told the pollsters before the pollsters' own turnout and don't-know adjustments (the raw stage), takes the latest such poll from each pollster whose full tables we hold in the last 28 days, and removes half of the average amount raw samples have missed by at past elections. The method, its evidence and its shrinkage factor of 0.5 are the subject of Correcting the raw answers: the method behind our UK projection, and the reasons for starting from raw answers rather than published headlines are in Why we start from British pollsters' raw answers rather than the published headline.

For present purposes what matters is how it is expressed as a lens. Let \(v_{E,p}\) be the published England average and \(r_{E,p}\) our England figure. The adjustment is simply their difference:

\[ a^{\text{best}}_{E,p} = v_{E,p} - r_{E,p}, \qquad p \in \{\text{CON}, \text{LAB}, \text{LD}, \text{REF}, \text{GRN}, \text{RB}\} \tag{4} \]

Substituting (4) into (1) returns \(r_{E,p}\) for each named party. Others are not adjusted, so after the rescale in (2) the named parties can differ from our figure by a few hundredths of a point when the two averages treat Others differently. Scotland and Wales are kept as polled.

The lens is built only when at least two pollsters' raw tables from the last 28 days are loaded and the historical backtest exists. Otherwise ForecastService marks "polls as they stand" as the default and the site says so. Today the projection draws on 5 pollsters: Find Out Now, YouGov, More in Common, Survation and Opinium.

Polls as they stand

The polls lens has an empty adjustment. It is the published England figures, averaged with equal weight per pollster, together with the Scottish and Welsh averages. This is the number other forecasts use, and we show it for comparison. Until 25 September 2026 it was the site's headline; the decisions log records the switch to our projection on the grounds that the raw-answer correction was the only approach to beat the published figures out of sample.

Average past error

The average-error lens assumes today's polls will miss the way they have on average. In England its adjustment is the mean, across the four core elections (2015, 2017, 2019 and 2024), of each election's average final-poll error:

\[ a^{\text{avg}}_{E,p} = \bar{e}_{p} = \frac{1}{T}\sum_{t=1}^{T} e_{p,t} \tag{5} \]

where \(e_{p,t}\) is the average error of the published final England polls for party \(p\) at election \(t\) and \(T\) is the number of elections, currently four. The Conservatives' mean error is −2.6 points, so this lens raises them by about that much before the rescale; Reform's is +1.4, so this lens lowers them. The full table of past errors is below.

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

In Scotland and Wales this lens uses each nation's own mean error, computed the same way from its nation-only final polls, and it includes the SNP and Plaid Cymru. The evidence text shown with the lens says plainly that those nations show no steady lean and that this correction made their polls worse in testing. That is why it is labelled a scenario, not an estimate.

Replaying a past election's miss

There are four replay-YYYY lenses, one for each core election. Each subtracts exactly the error of that year's final polls:

\[ a^{\text{replay-}t}_{E,p} = e_{p,t} \tag{6} \]

In Scotland and Wales the lens uses the same year's nation-only error, including the SNP and Plaid, where that year was measured. The four lenses are the concrete form of a question readers often ask: what if the polls are as wrong as last time? They are also the most vivid way to see how different past misses were. The 2015 and 2024 misses both understated the Conservatives and overstated Labour in England; 2017 did the opposite. Replaying 2017 therefore moves seats in the opposite direction from replaying 2015 or 2024.

A replay lens assumes the whole error vector recurs at full size. That is a strong assumption, and the forward test in Correcting the raw answers: the method behind our UK projection shows why we do not trust it as an estimate: taking all of the previous election's miss off the published figures did worse on average than leaving them alone. The replay lenses are scenarios. They say how the seats would fall if a known historical miss repeated, not how likely that is.

Pollster lenses

Three lenses, all built in Lenses/PollsterRecords.cs, look at pollsters rather than elections. They exist because the equal-weight average hides a choice (which pollsters count) and an assumption (that house effects cancel).

Each pollster corrected by its own record (pollster-record). For every pollster in today's England average, take its own mean error per party across its primary-benchmark final polls since 2015. A pollster with no checked final poll is given the industry mean error of equation (5) instead. The adjustment is the average of these across today's pollsters:

\[ a^{\text{rec}}_{E,p} = \frac{1}{K}\sum_{k=1}^{K} \hat{e}_{k,p}, \qquad \hat{e}_{k,p} = \begin{cases} \dfrac{1}{|T_k|}\sum_{t \in T_k} e_{k,p,t} & \text{if } |T_k| > 0,\\[6pt] \bar{e}_{p} & \text{otherwise,} \end{cases} \tag{7} \]

where \(K\) is the number of pollsters in today's average, \(T_k\) is the set of elections at which pollster \(k\) has a primary-benchmark final poll, and \(e_{k,p,t}\) is that poll's error. Because every pollster carries equal weight in the average, subtracting each pollster's own error from its own figure and then averaging is the same as subtracting the average of their errors from the average of their figures. So (7) is exactly an adjustment to the England average and runs through the same machinery as every other lens. The lens is built only when at least three pollsters are in today's average, and it keeps Scotland and Wales as polled.

The evidence text is candid: records are one to four elections long and methods change between elections, so "a house effect from 2019 may not hold now". The pollsters' records themselves, and how we rate them, are covered in Rating the British pollsters: accuracy, the Labour–Conservative lead and openness.

Only the pollsters best for Labour, the Conservatives or Reform UK (friendly-lab, friendly-con, friendly-ref). For a party \(f\), rank today's pollsters by their current share for \(f\) and keep the top half, but never fewer than three: \(m = \max\!\left(3, \lceil K_f / 2 \rceil\right)\), where \(K_f\) is the number of pollsters with a figure for \(f\). The lens is skipped if \(K_f < 4\). The adjustment is then the gap between the full average and the average of the chosen subset \(S_f\):

\[ a^{f}_{E,p} = v_{E,p} - \frac{1}{|S_{f,p}|}\sum_{k \in S_{f,p}} v_{k,p} \tag{8} \]

where \(S_{f,p} \subseteq S_f\) is the subset members that report party \(p\) (for Restore Britain, only those that prompt it). Applying (8) in (1) turns the England figures into the subset's average. These lenses do not keep Scotland and Wales as polled: the same point shifts are applied to the GB-wide parties there too.

These are deliberately partisan views. Their evidence text reads: "A pollster being good for [the party] today is not evidence that it is right." Their use is to show how much of the headline depends on which pollsters one chooses to believe, which the equal-weight average otherwise hides.

The reader's own what-if error

The what-if tool lets a reader set their own error for Labour, the Conservatives, Reform, the Liberal Democrats, the Greens and Restore Britain. The server builds a lens called custom from those values, clamping each to ±10 points (the page's sliders run to ±8), and applies it with equations (1) and (2). It has no nation-specific entries, so by equation (3) the same error applies in England, Scotland and Wales and the SNP and Plaid stay as polled. Unlike every other lens it starts from the polls as they stand, not from our projection, and it shows the central projection only, without the simulation, so that it can recalculate on every slider movement. The same lens code serves the reader-prediction page. Readers' lenses are never cached or stored.1

Every lens through the same seat model

The table below runs every lens through uniform swing, the default seat model, and counts seats for the four largest parties. The last column counts the seats whose projected winner differs from our projection.

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

Three patterns in this table are stable enough to describe in words, though the figures will move as polls arrive. First, the gap between our projection and the polls as they stand is small compared with the gaps among the past-miss lenses. Our correction is deliberately modest (it is half the historical raw-answer miss), so it moves relatively few seats, while a full replay of a past election can move dozens. Second, the replay lenses fall on both sides of the headline, because past misses did not all lean the same way. Third, the pollster-choice lenses can move the seat totals by more than our correction does. Which pollsters one chooses to read is itself a large, and usually unstated, assumption.

Each row is a central projection only. For the headline lens the site also shows simulated ranges, discussed in Simulating UK polling error: where the chances come from; the scenario table does not simulate each lens, because the point of the comparison is to isolate the effect of the assumption.

Scenario sets and seats at risk

Lenses/ScenarioSet.cs gathers every lens into one object. For each lens it runs the projection and records the seat totals, the implied GB vote shares, the largest party and its margin over the majority line of 326, and the number of modelled seats whose winner differs from the headline lens. Northern Ireland and the Speaker's seat are held at their 2024 winners in every lens and are excluded from the count of changes.

For each seat it then records the winner under every lens. From these it derives the measure the site calls seats at risk. Let \(H\) be the headline lens and \(\mathcal{P}\) the set of lenses in the Past misses group: the average-error lens and the four replays, so \(|\mathcal{P}| = 5\). For seat \(i\), let \(w^{L}_{i}\) be the winner under lens \(L\). The flip count is

\[ F_i = \sum_{L \in \mathcal{P}} \mathbf{1}\!\left[w^{L}_{i} \neq w^{H}_{i}\right] \tag{9} \]

A seat is at risk if \(F_i \geq 1\): at least one past polling miss, replayed on today's polls, would give it to a different party from our projection. The home page reports the number of such seats, and the map can colour seats by \(F_i\) out of five. The scenario comparison also picks out seats with \(F_i \geq 3\), a majority of the five past-miss lenses, as the closest to a coin toss. For each seat the set records the most common alternative winner among the flips.

Some choices here are deliberate and worth stating. Only the past-miss lenses count. The pollster-choice lenses and the reader's lens do not, because they are not claims about how polls have actually erred. The polls-as-they-stand lens does not count either, because the difference between it and our projection is our correction, not a historical error. The measure is therefore a count of historical scenarios, not a probability. A seat that flips under one lens out of five is not "20% likely" to change hands; the probabilities come from the simulation. Nor is the average-error lens independent of the four replays: in England its adjustment is the mean of theirs, so a seat that flips under several replays will often flip under the average too.

The scenario set also records, for each party, the lowest and highest seat total across the polls-as-they-stand lens and the five past-miss lenses. The home page uses this as its plain-language range: "if the polls miss exactly as they did at a past election, this is where the largest party would land".

The evidence shown with each lens

Every lens carries two pieces of text beside its numbers: a Summary of what it assumes and an Evidence statement of how much support that assumption has. Both are generated from the data, not typed by hand, so they stay true as the evidence changes.

For the replay lenses, the summary lists that year's England errors in order of size ("CON understated by …, LAB overstated by …") and adds the two largest errors in Scotland and in Wales where measured. The evidence gives the number of final polls behind each figure. For the polls lens, the evidence counts how many of the four elections understated the Conservatives and overstated the UKIP/Brexit/Reform family in England, and states that correcting the published figures by their average past miss did not beat them consistently. For the average-error lens it says that the size of the miss varies a lot between elections and that correcting Scotland and Wales this way made them worse. For our projection it gives the leave-one-out record: better than the pollsters at 3 and no worse at 4 of 4 elections, an average miss of 1.47 points per party against 1.67, from 18 final polls whose raw tables survive.

We do this because a scenario without its evidence invites a reader to treat it as a forecast. The replay of 2024 is the clearest example. It is the most recent miss and the one many people remember, and it is a legitimate warning that Labour's polling may be overstated, a warning we discuss in Correcting the raw answers: the method behind our UK projection. But the evidence statement makes clear that it rests on six final polls at one election, and the forward test shows that assuming the last election's miss repeats has not paid off in this period.

Why the lenses are shown side by side, not averaged

An obvious alternative would be to combine the lenses into one number, perhaps averaging their seat totals or weighting them by some measure of plausibility. We do not do this, for four reasons.

The first is that averaging has already been tested, in the form that matters. In England the average of the four replay adjustments is the average-error adjustment, equation (5). Subtracting it from the published figures is the "published minus average past miss" approach, which the leave-one-out test scored at 1.68 points per party against 1.67 for the published figures alone: no gain, and worse in 2017. Averaging the scenarios reproduces a correction that failed its test.

The second is that the lenses are not draws from one distribution. Our projection is an estimate that passed an out-of-sample test. The past-miss lenses are historical scenarios of very different likelihood. The pollster-choice lenses are deliberately one-sided. Averaging seat totals across such different objects gives a number with no clear meaning, and weighting them would need a probability for each scenario that four elections cannot supply.

The third is that seat totals are a non-linear function of vote shares under any swing model. The average of the seat totals under several lenses is not the seat total under the average lens. Where seats are close, averaging outcomes can produce a figure that no coherent set of vote shares would give.

The fourth is the plainest. The spread is the information. A reader who sees that one past miss gives the largest party a comfortable majority and another leaves it well short has learned something true about how fragile the forecast is. Averaging hides that. The site's design follows the decisions log: lead with a single best estimate backed by evidence ("rather than a cop-out 'it could be between these numbers'"), and put the alternatives next to it, each labelled with what it assumes.

Where we do want a probability, we use the simulation. It treats past error patterns as a source of uncertainty around the headline lens, draws them at random with random size and sign, and adds noise (see Simulating UK polling error: where the chances come from). That is a principled way to turn the same historical evidence into chances. The lenses do a different job: they show named, reproducible scenarios.

Limitations

The lenses inherit every weakness of the evidence behind them. There are only four elections in the core record, and the England error for each is the average of between two and six final polls. Scotland's and Wales's records are thinner still; Wales has one nation-only poll per election. The replay lenses can only reproduce misses that have happened. A future miss unlike any of the four, for example one driven by a party with no polling history such as Restore Britain, is outside the catalogue. Restore Britain has no error history at all, so no past-miss lens adjusts it except through the rescale.

The UKIP, Brexit Party and Reform errors are treated as one family. That is a modelling choice (the PartyFamilyHistory variant in the decisions log), and the parties' circumstances differed: the Brexit Party stood down in Conservative-held seats in 2019, for instance. The pollster lenses rest on records of one to four elections per pollster, and the equal-weight identity behind equation (7) holds exactly only before the rescale.

Finally, a lens changes only national shares. Every lens assumes the error is the same in every seat of a nation, so no lens can express a miss that falls unevenly, such as polls that understate a party only where it is already strong. That limitation belongs to the seat model and is discussed in From votes to Commons seats: UK swing models and their test and What we don't know yet: limitations and open questions in our UK forecast.

References

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

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 what-if page also offers the other lenses as starting presets. A preset copies only the lens's England adjustment into the sliders, so choosing, say, the 2017 preset there applies England's 2017 error in all three nations, whereas the replay-2017 lens itself uses Scotland's and Wales's own 2017 errors in those nations.↩

How to cite

Polling Forecast (2026). Lenses: reading the same British polls several ways. Retrieved 20 November 2026 from https://www.pollingforecast.com/research/uk-lenses. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like