Now that most of the results are in, it’s clear that the final projection of our MRP model properly told the story of the 2024 election.
Our model correctly projected the Labour landslide, captured well the extent and shape of the Conservative vote collapse, perfectly called the striking Liberal Democrat successes across the South and East, and also anticipated the growing presence of both Reform UK and the Green party as electoral forces in British politics. Additionally, it rightly anticipated that Labour would make significant gains across Scotland but struggle comparatively in Wales.
Every party was within our bounds on both share of seats and share of votes and overall, we called 92% of the seats correctly. This is the same level of accuracy that our 2017 MRP delivered. Our model properly picked up on the extent and breadth of tactical voting and understood how the coming proportional decline in Conservative vote share would spread across seats.
There are things in the results that we will need to build into our future work. While we foresaw Labour’s struggles in its safest seats and a potential fragmentation of its voter base coming, our model could have done better at picking up the extent of that. Labour’s decline in particular constituencies where there are a high number of ethnic minority voters is something we will particularly look to address. While our overall vote shares were largely on-point, Labour’s was too overstated, and the Conservatives was understated.
Our projections versus election results
YouGov’s final MRP seat projection of the campaign was Labour 431, Conservative 102, Liberal Democrat 72, SNP 18, Reform 3 Plaid Cymru 3 and Greens 2. With all but two constituencies declared, this compares to the results of Labour 412, Conservative 121, Liberal Democrat 71, SNP, 9, Reform 4, Plaid 4, and Green 4.
Our final vote share projection was Labour 39%, Conservative 22%, Reform UK 15%, Liberal Democrats 12%, Green 7%, SNP 3% and Plaid Cymru 1%. With all but two constituencies declared, this compares to the results of Labour 35%, Conservative 24%, Reform UK 15%, Liberal Democrats 13%, Green 7%, SNP 3% and Plaid Cymru 1%.
This was not a straightforward election. The electorate has shifted significantly over the last decade in a number of ways. First, the independence movement in Scotland in the wake of the referendum vote in 2014 shattered the dynamics of almost all first-past-the-post constituency battles there. Second, the Brexit debate unmoored various voter groups that were previously closely aligned to particular parties. The third – related – point, is that the 2019 Tory election-winning coalition was a unique blend of north and south cultural and economic conservatives and liberals. This coalition has fragmented beyond recognition, with each of Labour, Reform UK, the Liberal Democrats, and even the Greens benefiting.
Before last night, there was no historical precedent in this country for an election in which the government loses half its vote share between elections, and that vote does not move as part of a two-party swing. Past vote data has been instructive to pollsters but given the scale of the Conservative collapse it has been more of a signpost than a road map to how this extraordinary election would play out on the ground.
All of this made political modelling for the 2024 election especially difficult and makes our accuracy rate at the seat level of 92% of constituencies called correctly all the more pleasing.
Why our MRP performed so well
In the face of this dynamic and fast changing political landscape, we built a complex multi-level modelling and post-stratification (MRP) model similar to those we had used before to successfully call elections such as the UK in 2017 and Spain in 2023.
MRP constituency projection models work by first estimating the relationship between a wide variety of characteristics about prospective voters and their opinions – in this case, which party they will vote for at the general election – in a ‘multilevel model’. It then uses data at the constituency level to predict the outcomes of seats based on the concentration of various different types of voters who live there, according to what the multilevel model says about their probability of voting for various parties (‘post-stratification’).
A number of factors helped us. One of the most significant was a market-leading and completely new innovation; our “unwinding” algorithm. This algorithm was designed by YouGov to help the MRP model deal with the unique, high-change situation which the data was pointing toward and for which we had no precedent or prior data to help guide us. The unwinding algorithm, which has since been taken up by other pollsters and forecasters, learns from historical results what the typical distribution of party vote shares tends to look like (for each party), and re-fits constituency-level shares in the posterior distribution to better reflect this variance.
This has the effect of ‘unwinding’ the posterior distributions produced by MRP models to better reflect the spread of constituency-level results at British general elections, and in turn helps to control the proportionality of between-election party swings. Unwinding helped us to correctly call Conservative seats in particular such as Hereford and South Herefordshire, Huntingdon, and Broxbourne.
Another factor that made a notable difference in delivering an accurate seat projection was the quality of our data. High precision, big data, and machine learning models such as MRP work best when you feed good quality data into the model – as with any other branch of AI. Because YouGov owns its own panel, we have a unique relationship with the people taking our surveys. As a result, we have a deep understanding of who they are, which helps us build and execute models at scale which are truly market leading and help us become the most accurate when it comes to the MRP modelling. In the three projections we produced over the course of the campaign, we spoke to over 150,000 people meaning we could put a vast volume of quality individual level data into the models.
What this means for the future
YouGov has been using MRP modelling for almost a decade. Subsequently, we have built up both a track record of success and – crucially – an approach based on a combination of the latest technology and considerable experience. Of course, as a data and analytics business we constantly look for ways to improve our methods and we will be doing a thorough review of any areas where we can take learnings from for next time to make our MRP even more accurate and how we can apply our success into other areas of YouGov’s work.
As always, we will work with our panel, engaging them to ensure they continue to share the breadth and diversity of their thoughts, behaviours and opinions so that the people sharing their data with us keep on representing the complex nature of the public.
