Belated thanks to John Rentoul for his excellent article on last week’s forecast. This is not to say that I necessarily agree or disagree with his views on the desirability of particular election outcomes, just that I think it is a good discussion of lots of issues in politics and forecasting.
This is the first of a series of posts picking up on points he made. The others are on parliamentary arithmetic, leadership effects in electoral cycles, and predicting elections instead of sport.
John Rentoul was right to say that there was a sense in which last week’s forecast was already out of date at the time it was published. The most recent polls before the forecast did not have such big Labour leads as the ones earlier in that week.
I have been using the UK Polling Report average as the measure of current polls for the forecast. This is a weighted average of polls over the last 20 days. More recent ones are weighted more highly but the average will still reflect older data.
For long-range forecasting there is a good argument for not just using the most recent polls (however adjusted) since they will be too volatile and the forecast will bounce around too much.
For instance, the UK Polling Report Average for polls up to 16 July had a Labour lead of 3 points, but this morning’s YouGov poll has a 7-point Labour lead while Populus shows a tie. These are big differences politically if they were actual election results, but not statistically given the properties of the sampling. There is a good discussion of the volatility of recent polls by Anthony Wells here.
Even using averages there are fluctuations which mean that an unusual forecast is unlikely to herald a stable departure from the previous pattern, unless there are very clear reasons why the change occurred. There were no such reasons to explain last week’s exception and so this week the forecast has reverted to trend.