Chief Research Officer
I 'stress test' predictions of a Labour victory, using historically proven methods to see what they tell us about Conservative support.
Coming back from a break, I’ve found that Labour are continuing to poll strongly. The PM hasn’t yet managed to lift the Conservatives’ poll numbers persistently into the 30’s.
While vote intention polling makes for grim reading for Conservatives, we think a different set of numbers are worth bringing into the mix. In this week’s Bi_Focal, we’ll take a look at how leader ratings, local election results and voter expectations relate to election results - and how this complicates ‘case closed’ forecasts for the next election.
It is true that the Conservatives’ current poll numbers are bleak. Our latest poll here at Focaldata has Labour on 48% against the Conservatives’ 28%, a twenty point lead. That’s one of the more promising results for the Conservatives in a while. Ever since Liz Truss and Kwasi Kwarteng’s autumn budget statement, poll after poll has given Labour leads in excess of 20 points. A poll we did with Sam Freedman in mid January had Labour 25 points ahead, on 49% against 24% for the Conservatives. YouGov’s poll a week ago had Lab 47%, Con 24%.
When you look at the conversion of votes to seats, the picture only gets trickier for the Conservatives. The latest MRP model from Electoral Calculus gives Labour a majority of over 400, and puts the Conservatives on fewer seats than the SNP. Our projections - which take account of the new seat boundaries - are one of the more pessimistic models for Labour around. They suggest that Labour would need a swing of around 12-13% for a majority. History suggests that electoral swings of this size are rare. In 1997, the great landslide of modern times, Blair had a swing of 12.5%. But current polling implies a swing to Labour of over 16%, which would be a triumph eclipsing even Blair’s.
The case for why the Conservatives might do extremely badly has been made by a range of voices. Chris Cook in the FT, Arieh Kovler on his sub-stack, among others, have made their cases convincingly. The New Statesman has a long form piece on the death of conventional Conservatism. Most analyses are now concentrating on what form, shape and size a Labour victory will take, and which election it will most closely resemble. These takes will only ramp up in the aftermath of the resignation of Scottish First Minister Nicola Sturgeon and the now volatile political landscape of Scottish politics (a topic I’ll take up in the next issue of Bi-Focal).
The point of Bi-Focal is to offer unexpected and deep insights into public opinion, with a particular focus on UK politics. To that end, we really wanted to test whether things were quite as bad for the Conservatives as the polls suggest. Rob Ford’s excellent sub-stack suggests that governments can sink to extreme lows of 20-25 point deficits and still win, but the length of time that Labour has led in the polls suggests that the coming years will be difficult terrain for the Conservatives.
This week we look at three fundamentals-based methods to forecast general election outcomes. They share two key features, 1) being successful in projecting historical results, and 2) not being driven by voting intention (i.e asking “If a General Election was held tomorrow, how would you vote?”). These methods are:
The idea here is not to make an election prediction. Rather, it’s to see how much these different perspectives might contradict the dominant view of what will happen in 2024. Before engaging with the analysis, we didn’t know whether these methodologies would confirm the 20-25% point leads that Labour have, or suggest something else.
It’s been a habit of us pollsters to keep a close eye on ‘leader ratings’, alongside headline polling numbers. Ipsos have been helpfully tracking these since 1977.
Why do they matter? Why are they so predictive? There are several possible explanations. Ultimately leadership is the main question facing an electorate at an electorate - who should govern. We could hypothesise that electorates self-sort on party choices by conceding to a latent leader preference they hold. Leader ratings might also be predictive because they provide a much more intuitive response to a political question, and avoid the response ‘Don’t know’. Current voting intention polls contain large numbers of these ‘Don’t knows’, but leader ratings give an indication over how these undecideds will break when it comes to the actual election.
In any case, the evidence talks.The leader ratings lead of the Prime Minister over the Leader of the Opposition has been very highly correlated to the election result for decades. Matt Singh (of Number Cruncher Politics) famously spotted this and used it to call the 2015 election, which almost everyone else got wrong.
What does this mean for the next election? Given how poorly the Conservatives are polling right now, Sunak’s leader ratings are much better, and Starmer’s much worse, than we would expect. In our latest poll last weekend, Sunak is much more popular than his party. His net satisfaction / dissatisfaction is ‘only’ -15. Meanwhile Starmer’s personal ratings aren’t great on +6. Inserting this in the historical trendline gives a predicted election lead for Labour of only 3% (or 6% if you use the Ipsos numbers from January). This is far less than the 20% plus that current polls suggest.
The question though, is about the direction of travel. Will Sunak’s ratings decline to match his party’s, or will the Conservatives’ support increase to match their leader? So far, it’s likely to be the former.
When we looked at this back in December (“Ipsos Dec” on the graph), Sunak’s ratings implied a very close election. When his ratings were projected back onto the historical trend line (in blue) Labour and Conservatives came out almost tied. Now though, polling from the first two months of this year has seen Sunak becoming more unpopular, and the trend, though rather unclear, seems to point towards a 1997 style swing, delivering a 2010 seat count. That would be a Labour victory, but far from current predictions of an electoral wipeout for the right.
A second methodology comes from Local Election Results. Although local elections only take place in certain local areas each year, John Curtice and Stephen Fisher use the results to calculate the ‘Projected National Share’ (PNS) - the implied national share of the vote if the whole country had voted.
Local elections, of course, are not the same as General Elections, and people vote quite differently in them. But the idea this time - another brainchild of Matt Singh - is that the Local Election results indirectly track General Election support. The governing party tends to outperform the last 4-5 years worth of lagging Local election results by about 7 to 8 percentage points.
The model takes a weighted average over each local election leading up to the general election, with the most recent years weighted higher. The results are then fed into a linear regression. Again there are very few data points here, but the pattern looks fairly clear.
The model has historically performed very well, with an R2 of 0.71. (That said, the relation is less strong when predicting one year out, with an R2 of 0.62, and we are talking about very few data points here). The relationship is strong, in fact, except for a big miss in 2017 when Theresa May did very well in the local elections before imploding in the summer election campaign.
We polled local election voting intention (showing Lab 49 Con 28), and took our numbers as equivalent to the PNS from past local elections but for 2023 (again, a big assumption!). The model gave a predicted government result at the next election of minus 4% - that is, a 4% Labour lead.
This is much smaller than current polling, and very much in line with the 3%-6% Labour lead prediction based on leader ratings above. The caption on these fundamental perspectives is that Labour’s lead may be much less large than is often assumed. The local elections in May will be very interesting.
Our third approach this week - and the one I personally find the most fascinating and opaque - uses the work of Andreas Murr, Mary Stegmaier and Michael S. Lewis-Beck on voter expectations. Their model has been extremely successful, predicting the results in 2015, 2017 and 2019 very accurately. We’ve replicated it using their open code and dataset.
The idea is that asking about people’s expectations of the result (“Regardless of who you intend to vote for, who do you think will win the next election?”) implicitly increases the sample size of the poll. Everyone thinks of their networks and how they think others will vote, so your 1,000 person poll becomes a 50,000 person poll, and your 10,000 person poll a half million sample size.
When we looked into this model, the first thing that struck us was how esoteric and un-intuitive it was. It’s not as simple as asking what people expect to happen, and taking the numbers as a proxy for voting intention (like a “Wisdom of the Crowds” type exercise would do). Instead there’s a simple regression model, predicting each party’s share of seats in terms of:
Put simply, it suggests that the difference between parties in people’s expectations of whether a party will win is linked to the change in that party’s share of seats between elections. The model is fitted to all past elections, and then used it to predict the election in question in terms of seats.
I first came across this approach just before the 2019 General Election when the LSE posted this great table and the 360 Conservative seat prediction really stuck in my mind. I recall many of the polls and MRP seat forecasts actually narrowing (and therefore becoming less accurate) as the election went on - and it's a fascinating sidebar of the paper that actually the most accurate election prediction of all is a voter expectations model using data a full quarter out before an election! Another interesting sidebar - and not to go all Michael Gove - but expert predictions were materially worse than polls and models at the last election.
Voter expectations can vary wildly year to year. But as you can see from the plots above, which replicates the analysis from their paper, what the model does is compress the wild swings in expectation to make a more refined estimate of changes in seat share. Data on voter expectations has also been only intermittent, meaning that in 2001 in particular accurate data was not available.
Turning to historical predictions, there is again a very strong trend, with one very distant outlier. The model has historically been extremely predictive (even accounting for how few data points we have), except for a huge miss in 1997 where the prediction was over 12% off. R2 is 0.6. The miss in 1997 is pretty striking - the problem seems to be that the 1997 model ‘learns’ from 1983 - 1992. Since Labour didn’t win in any of those elections, the model doesn’t think Labour can win in 1997. There’s also a miss for the SNP and Lib Dem results in 2015: the model has trouble handling parties other than the Conservatives and Labour.
What do voter expectations models have to tell us about next time? Last weekend, we also polled the question “Regardless of who you plan to vote for, who do you think will win the next General Election?”. 50% thought Labour, against 22% to the Conservatives. It’s worth noting this is a lot more optimistic about Conservative chances than I think elite opinion is. I’d struggle to say that 1 in 5 people I talk to think the Conservatives will win the next election! Feeding these numbers into the model (it’s worth remembering here that we fit the model we used their replication data that only goes up to 2017) gave Labour 338 seats and the Conservatives 238. This is the kind of result you would expect with Labour holding a 2010-ish style 4-8 point lead, pretty far from where we are today. Again worth noting quite how wide the confidence bands are though!
I’ll emphasise again that I’m not trying to predict the next election this far out - that’s pointless. The result could easily be a huge Labour majority or a very tight election. But I did want to stress test the “case closed” view that Labour are going to win a massive majority at the next election, and the Conservatives face decades in the electoral wilderness. This could of course happen, and it's where the polls are at (including our own).
Fundamentals analysis suggests that we should treat some of this midterm polling and in some quarters wish-fulfilment of electoral carnage for the Conservatives with a bit more caution. These methodologies should give us pause for thought around predictions that the Conservatives are about to face an electoral cataclysm.
The fundamental methodologies are all suggesting small but comfortable Labour leads and an outcome much closer to 2010/1979, but nothing like the 400/500+ Labour seat count we’re seeing predicted today. Time will tell whether the polls will move towards this result as we get closer to the election. Then again, it’s also worth reflecting also that the new Governments of 1979 and 2010 ended up spending at least a decade in power having been put into office with slim-ish to comfortable majorities.
You can find tables for the full poll results here.
Subscribe to get the Bi_Focal newsletter delivered directly to your inbox.