A forecast of the 2017 UK general election
Today I posted the first of a series of forecasts of the 2017 UK general election. You can find it at www.electionforecast.co.uk.
Before you read the forecast, you may wish to ask yourself whether you really need an election forecast, and whether your time wouldn’t be better spent by reading the parties’ manifestos (when they are published). Analysis of the horse-race and analysis of policy aren't zero-sum (better outlets do more of both), but there is a balance to be struck.
You should also note that a similar forecasting methodology produced a
flawed forecast in 2015. Most of that was due to the failure of the
polls, but there was also a failure on our(*) part to model properly the
ways in which polls can go wrong.
Finally, you should know that there is much less data available about individual constituency contests in 2017. As a result, some of the constituency level predictions can be extreme. This doesn't matter if you are interested in aggregate outcomes, but it does matter if you are interested in specific seats.
With those disclaimers out of the way, here’s what I believe will happen:
- the Conservatives will finish as the largest party, with between 37 and 49 percent of the vote, and between 359 and 468 seats.
- Labour will end up with between 23 and 33 percent of the vote, and between 110 and 206 seats.
- the Liberal Democrats will end up with between 7 and 19 percent of the vote, and between 3 and 20 seats.
These are 95% forecast intervals, so if the model is well calibrated there should be a 95% chance of the outcomes falling within these ranges. Whether the model is in fact well calibrated we'll only know after the election.
Subjectively, the ranges make sense to me. I think it is difficult to imagine the Conservatives winning less than 37% of the vote: this would represent a deterioration of their position since the 2015 election.
These intervals are large. There is a considerable difference for the Labour party between finishing with 110 seats — which would represent the worst performance for a second party since 1931 — and finishing with almost twice as many seats.
The width of these intervals is a feature, not a bug. Had our forecast intervals been broader in 2015 we would have been less embarrassed by the outcome.
Having said that, these forecasts do rule out two possibilities. The probability of Labour finishing as the largest party is, for all intents and purposes, zero.
Slightly more adventurously, the probability of the Conservatives winning a majority of 326 seats or more rounds to 100%.
These forecasts are broadly comparable to the forecasts implied by betting markets (the forecasting record of which is also poor). At the time of writing, a £1 bet on a Conservative majority will win you £1.05 at Ladbrokes, and £1.04 on Betfair. This implies a probability of less than five percent. Given the favourite-longshot bias, the de-biased probability is lower still.
It is true that the best guess for the Conservative seat tally (412) is slightly higher than the mid-point of the spread betting markets (399), or the seat tally implied by a uniform national swing. The divergence with respect to the spread betting markets is, I think, due to these markets' optimism regarding the Liberal Democrats. The divergence with respect to uniform national swing is due, I think, to a failure to model properly changes in the Conservative vote in Labour-held seats as a result of the UKIP collapse.
In the next days and weeks I’ll try to set out the basis of the model, and in particular the changes introduced since 2015 to deal with the correlations of errors across the parties. Although I can’t guarantee that all parts of the model will remain the same (I did not expect this election, and have been racing to catch up), the main elements should stay constant.