Over on the right hand side of this site is a projection of how the current polls would translate into seats at a general election tomorrow, if there was a uniform swing. On twitter and suchlike I sometimes see if referred to as UKPR’s current prediction, but I’m afraid it isn’t. Polls don’t predict the next election, they measure support now, so the polling average here isn’t my best guess for the shares of the vote at the next election, it’s a measure of support in an election tomorrow. Of course, there isn’t an election tomorrow, and if there was, the polls probably wouldn’t be as they are – if there really was an election tomorrow then the last three weeks would have been full of manifestos, policy announces, campaigns and debates which may or may not have had some impact.
It’s also worth noting that while uniform national swing is not a bad guide by any means, it can certainly be bettered. To start with it’s definitely worth dealing with Scotland seperately based on Scottish polling figures, it might also be useful to include some assumptions about incumbency effects in seats with new MPs, and some degree of random variation at the margins.
I deliberately don’t make predictions this far out, given the huge amounts of unknowns. I tend to find most people who do predict this far out with any degree of confidence are – probably unconsciously – merely predicting what they would like to be the case. It’s rare to find someone confidently predicting a Labour victory who wouldn’t like a Labour victory (or who has an ideological axe to grind against the Tory leadership), or vice-versa on the Conservative side. Given the prominence of Nate Silver and other election prediction sites at the last US election I would expect a plethora of more academic and sensible election prediction models come the actual election (hell, I know for certain of several groups of academics working on various models), but so far virtually the only prediction I have seen that moves beyond wish-fulfillment to actually come up with a poll-based model is the attempt by Steve Fisher at Oxford here, with an explanation of the model here.
Steve’s model is a simple one – it is purely based upon voting intention polls and how they have tended to relate to the election result that follows*. We cannot assume that the polls will remain unchanged in the run up to the next election, given that in past Parliaments they have tended to change. Past change has not been a random walk, with equal likelihood of government’s gaining or losing in the polls – this is the key to Steve’s model. In the past the polls have rended to regress towards the result of the previous election (usually in the form of the government recovering). What Steve has done therefore is to take the current polls, and then factor in the sort of size and scale of changes that have typically happened to the polls over the last years of previous Parliaments, then based a prediction on that. At past elections this would have proven to be a more accurate predictor than just taking the current polls. That is not to say that that it is a particularly accurate prediction, only that in the past it would have been more accurate than assuming no change.
On that basis, if the polls over the next year behave like the polls in the last year of previous Parliaments the most likely result come the general election is a Conservative lead of 5 points over Labour, which would produce a hung Parliament with the Conservatives the largest party. The most important word in that sentence is probably the “if”, and perhaps the most important thing to note in Steve’s projection are the large prediction intervals around it. Steve’s model predicts the Conservative vote will be 37%, plus or minus 8.5 (so between 29% and 46%), the Labour vote at 32%, plus or minus 6.4 (so between 26 and 39). These are huge gaps. Of course, results towards the centre of those ranges are still considered more likely, but it underlines the imprecision of the projection, and the limitations on using current polling data to predict a general election a year away. Polls a year out from the election are not a very good prediction of the election. It would be wrong to say that anything could happen (Steve’s model, for example, suggests it is unlikely that Labour would get over 40, or that the Conservatives would fall below 29), but certainly a lot of different outcomes could happen.
It also reflects the sheer variety of elections. One criticism I’ve seen of Steve’s model is that this election will be different because of the coalition, the UKIP factor and the realignment of the Lib Dem vote. That may very well be true, but we could say the same about other elections – 1964 had two late changes of leader, 1966 wasn’t a whole term, 1974 was different because the Liberals started contesting all seats, 1979 was different because of the Lib-Lab pact, or the winter of discontent, 1983 was different because of the Falklands and the SDP split, 1992 was different because of Thatcher’s removal, 1997 was different because of the sheer scale of the landslide. 2001 was different because Labour never really had any mid-term blues to come back from. The infrequency of elections means that almost by definition each one has things that make it unique and different – yet Steve’s out-of-sample predictions shows the model would been a better tool at predicting those past elections from 20, 12 or 6 months out than just looking at what the polls 20, 12 or 6 months out were saying (it also underlines the difficulty for political scientists in coming up with any decent models at all – you only get 16 data points and they are all weird).
That doesn’t mean it would have been a particularly good prediction at those past points, just that it was better than the alternative of just looking at the polls 20, 12 or 6 months out. The polls now are a snapshot of public support now, they are not a prediction of what will happen in May 2015. If polls move in the sort of way they have in the run up to past elections we can expect the Conservatives to significantly recover. If they don’t, then they won’t, simple as that. Polls do not move by magic, drawn towards past election results by some invisible force. If they narrow, it will be because of the economy, because of changing attitudes to the parties, because, perhaps, of different factors weighing upon people’s political choices as an election becomes more imminent… that, however, is a post for another day.
(*I should also add that this is NOT Steve’s personal prediction of the election – it’s an attempt to see to what degree you can predict election results months in advance using just national poll data. I expect if Steve was making a personal prediction he probably would ponder what the impact of the economy, the party situation etc would be, but that would be a very different and more subjective model.)
Filed under: o