From votes to Commons seats: UK swing models and their test
Abstract. We turn nation-level vote shares into 632 seat winners by moving each seat's 2024 result by its nation's change. Given the exact 2024 nation shares, uniform swing called 479 of 631 seats correctly from the notional 2019 result, ahead of a uniform–proportional blend (461) and proportional swing (445); all three badly under-called Labour, so the seat model, not the polls, is a major source of error in any projection.
From national shares to seats
A general election is 650 separate contests, and a forecast of national vote shares says nothing directly about who wins them. Between the polls and a seat total there has to be a model that says how a change in a nation's vote is spread across its constituencies. This step gets less attention than polling error, but it matters as much. As the backtest in section 6 shows, a seat model given the exact national result of the 2024 election still misjudged Labour's seat total by more than 80 seats.
This paper describes the seat model Polling Forecast uses. It sets out the baseline (the 2024 result in each current seat), the nation-level inputs, and the three swing models the code offers, with their formulae. It works through one real constituency under each model. It then reports the test we used to choose between the models, explains why all three failed in the same direction, and states the rule for replacing the default. It closes with the model-based alternatives that others use and that we do not.
The seat model is deliberately separate from the question of polling error. Whatever reading of the polls a reader chooses (a lens, in the site's terms; see Lenses: reading the same British polls several ways), the output is a set of vote shares for England, Scotland and Wales, and the seat model turns those shares into seats in exactly the same way.
The baseline: 2024 on today's boundaries
Seats and shares
The starting point is the 2024 general election, fought on the boundaries that will be used at the next one. Seats/SeatBaseline.cs loads all 650 constituencies from the view web.vw_ConstituencyBaseline, which is built from the UK Parliament's constituency results. For each seat it holds the nation, the English region where relevant, the 2024 winner, and the 2024 vote share of each modelled party: Conservative, Labour, Liberal Democrat, Reform UK, Green, the SNP, Plaid Cymru and Others. Restore Britain, which did not exist in 2024, is given a share of zero everywhere. The seat also carries the notional 2019 result on the same boundaries, which is used only for the backtest in section 6.
Parties that do not stand in a nation are folded into Others there. The SNP's column in an English seat, for example, is added to Others and then removed, so each nation's seats contain only the parties that nation's model tracks. Each seat's weight is its number of valid votes, computed as electorate × turnout / 100.
Nation shares from the seats themselves
The model needs, for each nation, the 2024 share of each party as a baseline against which to measure change. We do not take these from published national totals. We compute them from the seats themselves, weighting each seat by its valid vote and leaving out the Speaker's seat:
Here \(n\) is a nation (England, Scotland or Wales), \(p\) a party, \(i\) a seat, \(s_{i,p}\) party \(p\)'s 2024 share in seat \(i\) in percent, and \(w_i\) the seat's valid vote. \(B_{n,p}\) is the nation's baseline share.
There are two reasons for doing it this way. The first is consistency. If a projection feeds a nation its own 2024 shares, every seat should come back exactly as it was in 2024, and that holds only if the baseline is computed from the same seat data the model moves. Published totals can differ slightly because of rounding, the treatment of minor parties or the Speaker. The second is the Speaker. In 2024, as by convention, the main parties did not stand against the Speaker in Chorley, so including that seat would slightly understate them and overstate Others. The decisions log requires the Speaker to be left out of every party-comparison calculation, and equation (1) does so.1
Seats that are not modelled
Northern Ireland's 18 seats are outside the prediction. Its party system is separate and the GB parties' polling has nothing to say about it. In every projection those seats keep their 2024 winners, and so does the Speaker's seat. The remaining 632 GB seats, 631 excluding the Speaker, are projected. The site's seat totals always add up to 650, so a party's total includes any Northern Ireland seats it held in 2024, which in practice affects only parties that do not stand in Great Britain.
Nation-level inputs
The seat model moves each seat by the change in its own nation. It therefore needs a target vote share for every party in each of England, Scotland and Wales. These come from three separate polling averages, all built in Polling/PollAverager.cs and gathered in ForecastService.CurrentPolls.
- England
- The England figure inside each GB poll, taken from the pollster's own tables or rebuilt from its English regions. The average uses the latest primary-grade poll from each pollster whose fieldwork ended in the last 28 days, with equal weights, so a pollster that publishes weekly counts no more than one that publishes monthly.
- Scotland and Wales
- Scotland-only and Wales-only Westminster polls. Because these are rarer, the window is 120 days, again with the latest poll per pollster and equal weights. If a nation has no poll in that window, its target is its 2024 baseline moved by the change in the GB average since 2024, party by party (uniformly), and the page labels the source accordingly.
Each average is then restricted to the parties modelled in that nation, with anything else added to Others, and rescaled to 100. The chosen lens is applied to each nation separately (Lenses: reading the same British polls several ways). The result is a target share \(V_{n,p}\) for every party in every nation. A GB figure, where one is shown, is the three nations combined with weights equal to their 2024 valid votes; it is an output, not an input.
Two features of this design are worth drawing out. First, the model never uses a GB swing to move a Scottish or Welsh seat when nation polls exist. The SNP and Plaid Cymru are not in England, and swings in Scotland in particular have long diverged from those south of the border, so GB-wide swing would be a poor guide there. Second, England is not broken down by region. Every English seat moves by the same England-wide change, which means regional differences in swing, such as a party gaining faster in the North than in London, are invisible to the model. Regional polls are loaded (The UK data pipeline: collecting, reading and checking British pollsters' tables), but no regional layer has been built into the seat model.
Three swing models
Let \(s_{i,p}\) be party \(p\)'s 2024 share in seat \(i\), \(B_{n,p}\) its baseline share in the seat's nation \(n\) from equation (1), and \(V_{n,p}\) its target share in that nation. A swing model is a function that returns a projected share \(\hat{s}_{i,p}\) for every party in the seat. The winner is the party with the largest projected share, and the margin is its lead over second place in points.
All three models share a normalisation step, which we write \(\mathcal{N}\). It sets any negative share to zero and rescales so the shares add up to 100:
Uniform swing
Uniform swing adds each party's change in the nation to its share in every seat:
If the Conservatives fall five points in England, they fall five points in every English seat. The idea goes back to the Nuffield studies and was set out as a general account of British electoral change by Butler and Stokes (1969). Its appeal is that it needs nothing but the last result and the national change, and that it has, over many elections, called a large majority of seats correctly.
Uniform swing has a well-known flaw at the extremes. A party that falls by more than its share in a seat is pushed below zero, which is impossible. The code catches this in \(\mathcal{N}\): the negative share is set to zero and the seat is rescaled, which slightly raises every other party there. The floor is not a rare edge case. In the backtest of section 6, where the Conservatives fell by about 22 points in England, uniform swing pushed them below zero in 53 English seats before the floor applied. The most extreme was Liverpool Riverside, where a notional 2019 share of 7.8% became −14.1%; the party actually won 3.6% there in 2024. The floor gives the right winner in such seats, but it means uniform swing no longer reproduces the national change exactly.
A new party is handled by uniform swing without special treatment. Restore Britain's 2024 share is zero everywhere and its baseline is zero, so equation (3) gives it its national target share in every seat. The docstring states the same for any party that barely stood last time: its vote is spread evenly until there is evidence of where it sits.
Proportional swing
Proportional swing scales each party's share in every seat by its national ratio of target to baseline. A party that halves its national vote halves its share everywhere. This fixes the negative-share problem, since a share can only shrink towards zero, but it creates a different problem for parties rising from a small base, which it multiplies up enormously.
The code treats parties differently depending on whether they had at least 1 point nationally in the baseline. Let \(\theta = 1\) point, let \(\mathcal{E}_n = \{p : B_{n,p} \geq \theta\}\) be the established parties and \(\mathcal{M}_n\) the rest (new or negligible parties). First,
New parties are thus placed evenly at their national share, exactly as under uniform swing. The established parties are then scaled together so that they fill whatever room the new parties leave:
and finally \(\hat{s}^{\,P}_{i,p} = \mathcal{N}(x')_p\), which here only guards against rounding. The rescaling in (5) is not a detail. When a party's national ratio is large, the unscaled shares in (4) can sum to far more than 100 in seats where it was strong, and (5) then squeezes every established party, including the one that grew.
The blend
The third model averages the two, with equal weights by default:
The rationale in the code is that uniform swing overstates a falling party's losses where it was weak, proportional swing overstates them where it was strong, and the blend sits between. Both components already sum to 100, so the outer \(\mathcal{N}\) changes nothing in practice. The weight \(\lambda\) is a constructor parameter, but only 0.5 is offered on the site and only 0.5 was tested.
Summary of the models
| Model | Change in each seat | Negative shares | New parties (under 1 point last time) |
|---|---|---|---|
| Uniform (default) | National change in points, added | Floored at zero, then rescaled | Added evenly at national share |
| Proportional | Share × national ratio, then rescaled to 100 | Cannot occur | Added evenly; others scaled down to make room |
| Blend | Equal-weighted average of the two | Via the uniform component | As both components |
Table: The three swing models in Seats/SwingModels.cs.
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.
- Tactical voting tested and left out (16 October). Switching against the Conservatives at a fixed rate raised the seats called correctly in 2024 (from 479 of 631 to 544), but when the rate was chosen on two of 2015 to 2024 and tested on the third it never beat uniform swing alone; anti-Reform switching made no difference before 2024. Tactical voting stays a reader's tool on the what-if page.
- The map (November). The UK map can show the general elections from 1997 on the boundaries of their day, council areas and census figures. The seat model is unchanged.
A worked example: Makerfield, 2019 to 2024
To see how the models differ, take one real seat and project it across the last election, where the answer is known. This is the same exercise as the backtest in section 6: start from the notional 2019 result on the 2024 boundaries, give the model England's true 2024 shares, and compare with what happened. Makerfield, in Greater Manchester, is a useful case because it involves the two things that dominated 2024 in the English seats Labour held: a collapse in the Conservative vote and a surge for Reform UK from the Brexit Party's 2019 base.
The England baseline \(B\) (notional 2019, vote-weighted as in equation (1)) and target \(V\) (2024), with the national change and ratio, were:
| Party | England 2019 notional | England 2024 | Change (points) | Ratio |
|---|---|---|---|---|
| Conservative | 47.80 | 25.89 | −21.91 | 0.542 |
| Labour | 33.34 | 34.50 | +1.15 | 1.035 |
| Liberal Democrat | 12.62 | 13.19 | +0.57 | 1.045 |
| Reform UK (Brexit Party in 2019) | 1.96 | 15.36 | +13.40 | 7.855 |
| Green | 3.05 | 7.32 | +4.27 | 2.401 |
| Others | 1.24 | 3.75 | +2.51 | 3.031 |
Table: England's notional 2019 and actual 2024 shares, computed from the seats (Speaker excluded).
Makerfield's notional 2019 result was Labour 45.27, Conservative 34.29, Brexit Party 13.08, Liberal Democrat 4.79, Green 2.58.
Uniform swing. Add each change: Conservative \(34.29 - 21.91 = 12.38\); Labour \(45.27 + 1.15 = 46.42\); Liberal Democrat \(4.79 + 0.57 = 5.36\); Reform \(13.08 + 13.40 = 26.48\); Green \(2.58 + 4.27 = 6.86\); Others \(0 + 2.51 = 2.51\). Nothing is negative and the total is already 100, so \(\mathcal{N}\) changes nothing. Labour holds the seat with a lead of 19.9 points over Reform.
Proportional swing. Every party had more than 1 point in England in 2019, so all are scaled by their ratios: Conservative \(34.29 \times 0.542 = 18.57\); Labour \(45.27 \times 1.035 = 46.83\); Liberal Democrat \(5.00\); Reform \(13.08 \times 7.855 = 102.70\); Green \(6.20\); Others \(0\). These add up to 179.31, so equation (5) scales each by \(100/179.31\): Reform 57.28, Labour 26.12, Conservative 10.36, Green 3.46, Liberal Democrat 2.79. Reform gains the seat by 31.2 points. Multiplying a party that was strong in this seat by nearly eight produces an impossible figure, and rescaling it down takes Labour with it.
Blend. Average the two projections: Reform \((26.48 + 57.28)/2 = 41.88\), Labour \((46.42 + 26.12)/2 = 36.27\), Conservative 11.37, Green 5.16, Liberal Democrat 4.07, Others 1.26. Reform gains the seat by 5.6 points.
What happened. Labour held Makerfield with 45.21%, ahead of Reform on 31.80%, the Conservatives on 10.88%, the Liberal Democrats on 6.79% and the Greens on 4.41%. Uniform swing called the winner correctly and was close on Labour's share. It underestimated Reform by about 5 points and overestimated the Conservatives by about 1.5: Reform did gain more in a seat where the Brexit Party had been strong, just not in proportion. Proportional swing was far out, and the blend inherited enough of its error to call the wrong winner.
Today's projection for Makerfield under each model can be seen on its seat page, which shows the 2024 result and the projection under our projection.
The backtest: notional 2019 to 2024
Design
To choose a default, we asked each model to do the job it is used for, on an election where the answer is known. Elections.Pipeline/Commands/SeatBacktest.cs starts from the notional 2019 result in each of the 631 GB seats (the 632 on the 2024 boundaries less the Speaker's), gives each model the exact 2024 vote shares for England, Scotland and Wales computed by equation (1), and counts the seats where the projected winner matches the actual 2024 winner. Independents, who won five seats, count as Others.
Supplying the true nation shares isolates the seat model from polling error. Any seat it gets wrong, it gets wrong even with a perfect poll. The test is therefore an upper bound on how well the model can do in real use.
Results
| Model | Seats correct (of 631) | Share correct | Labour seats projected (411 won) |
|---|---|---|---|
| Uniform swing | 479 | 75.9% | 329 |
| Blend | 461 | 73.1% | 315 |
| Proportional swing | 445 | 70.5% | 290 |
Table: Seat-model backtest, notional 2019 to 2024, exact nation shares. These are fixed historical results.
Uniform swing won clearly and became the default. We reproduced these figures independently from the baseline data while writing this paper, which also gives the breakdown below.2
| Model | England (of 542) | Scotland (of 57) | Wales (of 32) | Conservative seats (121 won) | Lib Dem seats (72 won) | Reform seats (5 won) |
|---|---|---|---|---|---|---|
| Uniform swing | 403 | 47 | 29 | 240 | 36 | 2 |
| Blend | 387 | 44 | 30 | 226 | 38 | 25 |
| Proportional swing | 374 | 42 | 29 | 195 | 49 | 61 |
Table: The backtest by nation, and projected seats for three other parties.
What went wrong
All three models under-called Labour by between 82 and 121 seats, and uniform swing also gave the Liberal Democrats only half the seats they won. The errors are one-sided. Under uniform swing, 82 seats were projected Conservative that Labour won and 37 were projected Conservative that the Liberal Democrats won. The two largest parties' misses and the Liberal Democrats' miss all point to the same cause: the anti-Conservative vote in 2024 was not spread evenly. It concentrated where it could defeat a Conservative.
Two mechanisms were at work, and no swing model can see either. The first is tactical voting. Where the Liberal Democrats were the challenger, Labour voters lent them support, and the reverse where Labour was the challenger. A uniform model moves Labour and Liberal Democrat shares by the same national amount everywhere, so it cannot put a Liberal Democrat surge in Surrey and a Labour one in Kent. The second is uneven swing more generally. Labour's England share rose by about one point overall, but not evenly. In the 99 English seats where Labour led the Conservatives by more than 20 points in 2019 (notionally), its share fell by 12.7 points on average; in the 83 seats where the two were within 10 points it rose by 1.4; and in the 274 seats where the Conservatives led by more than 20 it rose by 5.3.3 Labour won 411 seats on about a third of the GB vote because its vote was efficiently distributed, and efficiency of that kind is exactly what a model that applies one national change to every seat assumes away.
Proportional swing did worst for a specific reason, visible in the Makerfield example. In 2019 the Brexit Party did not stand in seats the Conservatives had won at the previous election; it has a notional share of zero in 263 of England's 543 seats. In the remaining seats, mainly Labour-held ones in the North and Midlands, it was relatively strong. Proportional swing multiplied those shares by nearly eight and gave Reform 61 seats, against the 5 it won, while leaving it at zero wherever it had not stood. The blend halved that error but did not remove it. Uniform swing, by adding the same 13 points everywhere, was closer to the truth for Reform even though it undercounted its best seats.
A related effect is that none of the models reproduces the national shares exactly once the floor and rescaling are applied. Weighting the projected English seats by their 2024 valid votes, uniform swing implies a Conservative share of 26.5% against the 25.9% it was given, the blend 27.8% and proportional swing 29.1%; proportional swing's implied Reform share is 11.0% against 15.4%. The gap is smallest for uniform swing, which is another reason to prefer it.
Uniform swing also beat the others on vote shares: its average absolute error on the projected shares, across every party entry in every seat, was 4.0 points, against 4.5 for the blend and 5.5 for proportional swing.
How good is 75.9%?
The 2024 election was unusually hard for a seat model. The swing from Conservative to Labour was large, a new party rose sharply, and tactical voting was extensive, so a 2019-to-2024 test is a demanding one. We have not run the same test on earlier elections (2015 to 2017, or 2017 to 2019), which the constituency data would allow. Until that is done, we cannot say whether 76% is typical of uniform swing or a bad year for it. We also cannot say whether the ranking of the three models is stable. It is quite possible that proportional swing does better at an election without a party rising from a highly uneven base.
The rule for replacing uniform swing
The decisions log sets a simple rule: a demographic or tactical-voting model must beat 479 correct seats on this backtest before it can replace uniform swing as the default. It is written into the documentation of SwingModels.All, and the methodology page states that such a model "will replace uniform swing only if it beats it on this test".
That model is planned, not built. The ingredients exist in the database: 40 demographic variables for every seat from the 2021 and 2022 censuses, the estimated 2016 EU referendum Leave share for GB seats, the 2024 and notional 2019 results, and by-election results since 2011 (The UK data pipeline: collecting, reading and checking British pollsters' tables). None of them is used in the seat model today.
The rule has a known weakness. A single test election gives one number, and a model with many parameters can be tuned to beat 479 on 2024 without being better in general. A fair version of the test would fit the demographic model on earlier elections only and then predict 2024, in the same leave-one-out or forward spirit as the tests of our polling correction (Correcting the raw answers: the method behind our UK projection). Any replacement will need to pass a test of that kind, not only exceed the headline figure.
Model-based alternatives
The main alternative to swing models is to model each seat's vote directly. Multilevel regression and post-stratification (MRP) fits a model of individual vote choice to a large poll, with respondents' demographics and their constituency's characteristics as predictors, and then uses census counts to predict the vote in each seat. Lauderdale et al. (2020) describe this approach for national and sub-national outcomes in the US and the UK, including the constituency estimates YouGov published at the 2017 election. Hanretty, Lauderdale and Vivyan (2016) show how national and constituency polling can be combined for forecasting. The exit poll takes a different route to the same end: Curtice and Firth (2008) set out how the BBC–ITV exit poll models the change in each party's vote across a panel of sampled seats as a function of seat characteristics, rather than assuming it is uniform, and converts the result into seat probabilities.
These methods can capture exactly what our backtest shows uniform swing missing: uneven swing tied to the kind of seat, and, with the right predictors, some tactical voting. We do not use them, for three reasons. The first is data. MRP needs respondent-level data from very large samples; we work from the published tables, which give weighted cross-breaks, not individual responses. The second is that we want the seat model to be separable from the polling correction, so a reader can see which part of a forecast comes from which assumption. The third is the rule above. A more complex model has to show it is better on the same test. MRP is not automatically superior: its projections also depend on how the poll is weighted, and its errors in any one seat can be large.
Some pollsters now publish MRP-based figures. The decisions log records that YouGov's headline voting intention has been produced through its MRP model since 2025. We use those figures only as a pollster's headline in the England average, and we exclude separately published MRP seat projections from the averages altogether. How the published MRPs have done, and how each one compares with our projection, is set out in MRPs and mega-polls in Britain: what they measure, how they have done, and how we treat them.
Limitations
The seat model assumes that every seat in a nation moves by the same national change, in points or in proportion. That is the root of its main weakness, and the backtest measures it: even given a perfect poll, uniform swing got about one seat in four wrong in 2024 and missed Labour's total by 82 seats. In a real forecast the seat model's error comes on top of the polling error.
There is no regional layer for England. There is no treatment of incumbency, candidate effects or by-election history. New parties, currently Restore Britain, are spread evenly, which is almost certainly wrong for any party whose vote is concentrated. Seats won in 2024 by independents are projected from their 2024 share as Others and move with the national Others share, which says little about whether a particular independent will stand again. Northern Ireland is not modelled at all.
The simulation (Simulating UK polling error: where the chances come from) adds random seat-level noise to represent local effects the swing model cannot see. That noise is independent from seat to seat and centred on zero, whereas the backtest shows the seat model's errors in 2024 were systematic, all in the same direction for the same parties. The simulated ranges are therefore likely to understate the uncertainty that comes from the seat model. We return to this in What we don't know yet: limitations and open questions in our UK forecast.
References
Butler, D. and Stokes, D. (1969). Political Change in Britain. London: Macmillan.
Curtice, J. and Firth, D. (2008). Exit polling in a cold climate: the BBC–ITV experience in Britain in 2005. Journal of the Royal Statistical Society: Series A, 171(3), 509–539.
Hanretty, C., Lauderdale, B. and Vivyan, N. (2016). Combining national and constituency polling for forecasting. Electoral Studies, 41, 239–243.
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.
-
The notional 2019 baseline used in the backtest is computed with the same equation, and the code weights each seat by its 2024 valid vote in both cases. The England "2019" shares in section 5 are therefore not identical to published notional 2019 totals, though the difference is small.↩
-
The reproduction used the same seat data and the same formulae as the C# code. It matched the recorded totals exactly (479, 461 and 445 correct; Labour 329, 315 and 290). The breakdown by nation, the seat totals for other parties, the implied national shares and the share errors in this section come from that reproduction.↩
-
Computed from the notional 2019 and actual 2024 results in the baseline data. The groups are defined by the notional 2019 Labour–Conservative margin and do not cover every English seat.↩
How to cite
Polling Forecast (2026). From votes to Commons seats: UK swing models and their test. Retrieved 20 November 2026 from https://www.pollingforecast.com/research/uk-seat-model. Live figures change as new polls arrive; the date above is when the text was last revised.


