June 8th 2017 is a day that the election predictor/modellers will want to forget
It is little comfort to the election predictors/modellers that Wikipedia has now decided to record for posterity how successful they were in predicting the party seat outcome of GE17. The chart is above.
As can be seen only the YouGov model based on 50k+ of its own interviews came out of this well.
Throughout the campaign the forecaster/modellers aimed to produce projections of the party seat totals which, are course, based on the outcome of 650 separate first past the post elections in the different constituencies. This is sharp contrast to the BREXIT referendum or, say, the final round of the French election where it is a binary choice of two based on national totals.
The reason the GE17 modellers got it so terribly wrong was that their main data sources, the opinion polls, had, with one very honourable exception, a huge polling fail. Never was the saying garbage in garbage out so appropriate.
Virtually all of the models were following standard swing theory in their approach to seat predictions which meant a big error on top of everything with their LD seat projections. The party increased its MP total by 50% with a reduced national vote share. The exception was YouGov with its own exclusive and large polling data source.
A problem was the overwhelming CON landslide narrative which dominated everything and was reinforced by the adjustments almost all the pollsters had made following the GE15 polling fail. Any pollster that produced numbers that didn’t fit the general perception were attacked and their findings ignored.
This meant that sharp move away from the Tories after the manifesto saga was much less noticeable at the time and things like the consequential drop in CON 65+ support are only now being observed.
Lots of lessons from GE17 then which no doubt they will try to avoid next time.
The problem, of course, is that there is a huge appetite during a campaign for information on how the battle is going in terms of seats. Gamblers in particular are a key audience. The forecaster/modellers satisfy that need.
No doubt the Wikipedia table will be rolled out next time as a reminder to treat projections with caution.