Methodology
National polls do not convert easily on a one-to-one basis to seats in Parliament. For example, in 1997 the Labour Party won 63% of the seats on 43% of the vote. In order to derive an estimate from national polls we use an "ensemble" of two different predictions: our own swing model and an average of pollsters' MRP forecasts.
First, our swing model takes the simple principle of “uniform national swing”—the idea that support for parties rises and falls across all constituencies in the nation by the same magnitude—and augments it with specific regional polling, where it is available, from Scotland, Wales, London and so on. For example, the vote share of the Conservative Party in Macclesfield in any given general election—a seat that they have held since 1918—is modelled as a function of its performance in the prior general election, as well as how the Tories are performing in both nationwide polls and specific polls from the north of England.
To estimate our swing model, first we trained a model using 9,398 individual constituency-level election results between 1959 and 2019 along with polling data (polling trend lines for each election cycle were estimated using the MGCV package in R). We fit a multinomial logistic regression model using the LASSO method, a statistical technique that eliminates or reduces the impact of certain variables in order to maximise accuracy on unseen data.
However, our swing model does not take into account the local effects of parties' campaigns in individual constituencies; the effect of tactical voting; or idiosyncratic seat-level changes in parties' support. Multi-level regression and post-stratification models (MRP) are a novel technique for producing seat-level estimates by segmenting voters into fine-grained socio-demographic groups. Ahead of 2017 there was one MRP model by YouGov, a polling firm; followed by MRP models from three firms ahead of the 2019 election. After adjusting these MRPs for the shift in national polls between the time they were conducted and the election, we then took an average of them, and calculated the blend of the MRP average and our swing model that minimised prediction errors for every party in every constituency.
There have been many more MRPs this year: Electoral Calculus; Focal Data; Ipsos; More In Common; Savanata; Survation; YouGov and The Economist itself (using data from WeThink) have all released them. Given the novelty of the method, most of these pollsters do not have track records and so there is no objective way to know which one might be best. Therefore we have produced our 2024 forecast using an average of all the MRPs' seat-level predictions (weighted 78%) and our own swing model (weighted 22%). As pollsters' estimates of the nationwide vote share differ, we have adjusted our final blended forecast to take into account our own assessment of parties' nationwide support from our poll tracker.
Finally, the point estimates from our ensemble model were fed into another model that uses large-sample survey data to analyse the correlation between voters in different constituencies across the country. For example, if Labour gains one vote in Macclesfield, there is a high chance that they will also gain a vote in other small-town northern seats like Ossett and Denby Dale, but the party is less likely to see a concurrent gain in a well-heeled London suburb like Richmond Park. After 10,000 simulations of these correlated vote shares, we estimated the 95% confidence intervals for every party's vote share in each of Great Britain's 632 constituencies (ie, excluding Northern Ireland).
Our model will continue to update until July 4th absorbing the latest polls nationwide and regional polls along with any other MRP models that are produced.
