Predicting the election with modern statistics

Posted 30 November 2019

Electoral Calculus is a long-term supporter of using modern mathematical and statistical techniques to get the best election prediction. While no prediction method can be completely accurate, our testing has shown that the new regression-based techniques are better than or no worse than the classical polling analysis performed by most pollsters. We have tested this with pre-election polls from 2017, 2015 and 2010 (details).

Our method works as a three-stage process:

• Take a medium-sized poll of respondents and train an algorithm to learn (or regress) how people's voting intention typically depends on aspects of their demographic and political profile, such as age, sex, class, occupation status, past vote, etc.
• Apply those relationships to the known characteristics of groups within each constituency using (anonymised) census data and past election results, to get a forecast for how people in the seat are likely to vote on average. We call this the baseline prediction.
• Make a small adjustment to every seat to reflect changes in public opinion since the baseline was last calculated. This adjustment is based on uniform national swing (UNS) modelling. That modelling is not recommended for large changes in public opinion, but should be reasonable for small movements.

This allows us to have a regression-based prediction which can change day-by-day as public opinion evolves. Our testing shows that these methods can be very effective with sample sizes of about 5,000 respondents.

Comparison with recent YouGov MRP poll

Respected pollster YouGov conducted a large-scale MRP poll in late November. The abbreviation MRP stands for "multi-level regression and poststratification", which means it is a regression-based poll which uses similar techniques as Electoral Calculus. The poll surveyed around 100,000 respondents, which is a very large compared with typical polls of between one thousand and two thousand respondents. The sample dates were from 20-26 November 2019.

To compare the YouGov and Electoral Calculus predictions we need to adjust for the fact that they were conducted at different times when the parties had different national levels of support. We can do that by applying the third-stage UNS adjustment mentioned above. This can be done either to switch the Electoral Calculus prediction to be on the same basis as YouGov, or vice versa. The results of the two experiments are shown in the two tables.

The first table (A) shows the predictions on the basis of the YouGov vote share where the Conservatives are exactly 11pc ahead of Labour. YouGov calculated the Conservatives would win 359 seats to have a majority of 68. The Electoral Calculus estimate for the Conservative seats is slightly smaller at 345, implying a majority of 40.

The second table (B) shows both baselines converted into the Electoral Calculus poll-of-polls as at 30 November. This has a reduced Conservative lead of 8.5pc. (That poll of polls was based on a sample of 8,341 respondents from 21-28 November.) The YouGov baseline translates into 346 seats for the Conservatives with a majority of 42. The Electoral Calculus baseline gives 336 seats to the Conservatives, with a majority of 22.

Neither table shows a massive difference between the two baselines. Electoral Calculus is slightly lower for the Conservatives and the SNP, and slightly higher for Labour and the Liberal Democrats. Historically, a prediction is quite good if it is within 20 seats of actual result for the major parties, and these two baselines are well within that tolerance.

One baseline might be a bit better than the other, or the truth could be somewhere in between, or they might both be miles off. We will have to wait and see.

The Electoral Calculus prediction will be regularly updated from its own baseline until election day to give the public the most up-to-date information about likely political opinion.