By Stephen Fisher, Rosalind Shorrocks and John Kenny.
Lots of the forecasts for the 2017 British general election were wrong in pointing to a Conservative majority. Most even suggested an increased majority when the party actually lost one. The well-known exceptions are the exit poll and YouGov multilevel regression and post-stratification (MRP) model. Less commonly remembered are the polls that were within a point of the Conservative-Labour lead and so, by standard uniform change calculations, provided an indication that the Tory majority was in jeopardy.
During the election campaign we ran an exercise combining the different forecasts for shares of the vote, seat tallies, and probabilities of a Conservative majority. Inevitably the combined forecast is only as good as the average of what goes into it, and that average was poor. But since the exercise involved classifying and averaging forecasts by the methodology they used, it is possible to reflect on which types of forecasting method performed better than others.
The figure below shows predictions of the Conservative share of the vote over the course of the campaign from the Political Studies Association (PSA) expert survey, betting markets, forecasting models and opinion polls. Strictly speaking polls are snapshots not forecasts, but the final polls performed as well as the experts and better than all the other methods. Consistently the worst source of predictions for the Tory share were the betting markets.
Note: Due to an error in our 2ndJune 2017 betting market calculation (and no way of correcting it at this stage) the betting market line runs direct (interpolates) from 26thMay to the final 8thJune 2017 estimate.