First combined forecast for the 2024 general election

By Stephen Fisher, John Kenny, Paul Furey, and Polina Ryzhuk. 30^th May 2024.

Here we combine seats and vote share forecasts from a variety of sources including betting markets, polls, statistical forecasting models, and citizen forecasts. As well as updating weekly, incorporating new forecasts with different methods as they become available, there may be developments in the methodology. Comments welcome.

All the different sources of seat forecasts in Table 1 point to a very large Labour majority, well over 100.

Table 1. Seat Forecasts

	Betting Markets	Complex models	Simple models	MRP	Average
Con	139	111	193	127	142
Lab	405	445	391	436	419
LD	43	36	30	36	36
Reform	0	0	0	0	0
Green		1	1	0	1
SNP	19	19	14	30	20
PC		3	3	3	3
Labour majority	159	239	132	221	188

In the “Simple Models” above we use uniform change in England, Scotland, and Wales separately. It is well known that with the big swing from the SNP to Labour in Scotland, there is a separate political dynamic north of the border that must be accounted for. Less well known is the more modest rise for Labour in Wales. Since 2019, Labour are up 19 points in Scotland, 11 points in GB as a whole, but by just 3 points in Wales, according to the average of the two polls from Wales this month.

Accounting for that makes a slight but important difference, especially for estimates of the number of Plaid Cymru seats. While it is not ideal to use GB change in the share of the vote as a proxy for change in England only, that is what we have done. We hope that pollsters publish England-only as well as GB vote intention figures in future.

Since we anticipate various Multilevel Regression and Poststratification (MRP) model-based forecasts between now and the election, we have created a separate category for these models. For our 2019 Combined Forecast, MRP models were included in the “Complex models” category. In Table 1 the MRP column shows the average of the YouGov and Survation MRP models from March 2024. While they are somewhat old, there has been relatively little change in public opinion since then.

Given discussions about whether MRP models have been estimating constituency change that is too proportional to prior strength, it will be interesting to see whether the MRP models continue to forecast such a large number of Labour seats. The estimates from the March MRP models which, somewhat by design but also for want of data to the contrary, have been estimating that the drop in the Conservative vote is broadly proportional to the share of the vote they are defending in each constituency. The average GB vote share for the two MRP models in Table 1 has the Conservatives on 25% and Labour on 43%. That 18 point lead is narrower than polls are currently showing. A uniform change projection (using Scotland and Wales polls separately) from that March MRP vote-share average would suggest 372 seats for Labour instead of the 436 in Table 1. That is a difference of 64 seats. The corresponding difference for the Conservatives is even greater, at 85 seats. With so much at stake, the degree to which the swing in each constituency will be proportional to the share of the vote Conservatives won last time is an important unknown factor that could make a big difference to the party fortunes.

The March YouGov MRP estimates involved less proportional change than those from the Survation model. Variation between MRP models during the campaign may prove substantial, as they were in 2017 between the YouGov and the Ashcroft MRP models.

Table 2. below gives some indication of how confident different forecasting methods are in different possible governing outcomes. Averaging the probabilities suggests a 95% chance of a Labour government and an 86% chance of a Labour majority. As we have previously noted, citizen forecasts are very sensitive to question wording. It is probably no accident that of the polls we have found, the one with the greatest proportion of citizens expecting a Labour majority is the only one where respondents were given the option of saying they think Labour are “on course for a landslide” (see Table 62 here).

Table 2. Probabilities (%)

	Betting markets	Models	Polls	Citizens	Volunteered	Average
Lab Majority	92	98	100	54		86
Lab 100+ majority			67			67
Con led Gov	8		0	12	0	5
Lab led Gov	95		100	84	100	95

For the calculation of implied/pseudo probabilities for both the citizen forecasts and those volunteered by participants in the Good Judgement Project, we applied the commonly accepted notion that the Conservatives need a majority (or something very close to a majority) to govern. Probabilities from polls are based on the proportions of pollsters with GB Labour leads of sufficient magnitudes: above 12 points for a majority and above 19 points for a majority over 100, using uniform change calculations (with Scotland and Wales separate as discussed above).

Table 3 summarises forecasts for vote shares. The models (that systematically take into account stable differences between pollsters) do not currently produce substantially different estimates from simple averaging of the most recent poll from each pollster. The betting markets appear to expect a small drop in the Reform UK share relative to current polls, but no corresponding rise in the Conservative vote. Averaging across the different methods, Labour are expected to win with a 20 point lead over the Conservatives: enough for a majority of over 100.

Table 3. Vote share forecasts (%GB except where stated)

	Poll aggregates	Betting markets	Models	Average
Con	23	24	24	24
Lab	45	42	45	44
Lib Dem	9	11	9	10
Reform	11	8	11	10
Green	6		6	6
SNP (in Scotland)	31			31
PC (in Wales)	14			14
Lab-Con lead	21	18	21	20

Methodology

The idea of combining forecasts from different sources has a good track record, though it has to be admitted that our attempts for the 2017 general election and the 2016 Brexit referendum did not work out so well. For reviews of previous combined forecasts here for 2017 and here for 2019. Also worth noting is the experience of the pollyvote.com combined forecast of the US presidential elections.

The basic approach is to combine forecasts by averaging them within each category and then average across categories. Since the different sources do not all present similar figures that can be averaged on a like-for-like basis there are various judgement calls we have had to make on how to treat the data.

Polls

For vote shares, we use the various available polling averages. Some admittedly are quite sophisticated, allowing for house effects, but they are nonetheless estimates of current public opinion and not future votes.

To estimate a pseudo-probability of a Labour majority from the polls, we take the proportion of pollsters with a GB Labour lead greater than 12 points, on the basis that this is the level required for a Labour majority under traditional uniform change assumptions. Similarly, the proportion of the most recent polls from each pollster with a lead of 19 points or more serves as a pseudo probability for the chances of a Labour majority of over 100 (making allowances for current polling in Scotland and Wales).

Statistical models

We have divided statistical models into simple (poll average plus uniform swing seats projections), Multilevel Regression and Poststratification (MRP), and other complex models (anything more elaborate than uniform change but not necessarily particularly complex). Within these categories we simply average the available estimates of seats and shares. We have not excluded any models based on quality, but they do have to be statistical models as opposed to personal guesses.

Betting markets

For the seats forecasts we are using the mid points of the spread betting for the markets. The markets might collectively imply fewer or more seats forecast for the parties in total than there are seats to be won in Britain. This is because the markets are separate for each party and do not need to be consistent collectively.

Betting markets for shares of the vote are typically done in bands. We take the mid points of the bands with the highest implied probabilities weighted by their implied probabilities from the betting odds. We then average across bookies.

Citizen forecasts

Some polls ask people what they think that the outcome of the election will be. We call these ‘citizen forecasts.’ Different pollsters use different survey questions but they can be combined to generate pseudo-probabilities. We use the proportion of poll respondents who think there will be a Labour majority, excluding “don’t knows” and re-percentaging, as the pseudo-probability of a Labour majority. We treat the question of whether Rishi Sunak will continue as prime minister to be equivalent to the Conservatives getting a majority, because they do not have any clear governing partners in the event of a hung parliament. Where appropriate, the probability of a Labour led government is taken to be 100 minus the probability of a Conservative led one. But that is not the case for questions where respondents are given the chance to say they think there will be some “other” kind of government. We use the most recent poll with a citizen forecast survey question from each pollster and only those in recent weeks. Only a few pollsters ask citizen forecast questions.

Volunteered forecasts

These come from the Good Judgement Project, which encourages people to forecast the outcomes of various events in order to develop and improve their forecasting skills. The people providing their predictions to the Good Judgement Project are not necessarily experts, but nor can they be seen as a representative sample, and so we report their predictions separately as ‘volunteered’ forecasts.

Estimates come from mid-afternoon on 29^th May 2024.

Sources:

Betting Markets:

Models (Elaborate):