Do August polls predict the next election?

There is a new YouGov poll in tomorrow's Sun. Topline figures, with changes from YouGov's previous poll are CON 42%(nc), LDEN 28%(+2), LDEM 17%(-1). YouGov's last poll was done just one day before this, with their fieldwork periods overlapping, so it's almost certain that the differences between the two polls are just random sample error.

The Sun, however, has got very excited about the poll, saying "it is the most cruicial opinion poll for five years" as "every August poll in the summer before a spring general election has predicted the real result accurately to within one percent since 1996."

This is, to put it kindly, not entirely accurate. To start with, YouGov's poll in August 2004 were not within 1% of the 2005 election result, it showed Labour and the Conservatives equal on 34%. Populus too showed Labour and the Conservatives equal in their August 2004 poll. Two August 2000 polls from MORI showed Labour on 51%, evidently not within 1 point of the 42% they got at the 2001 general election.

The claim stems from a couple of weeks back when Nick Sparrow wrote to Mike Smithson in response to him querying whether August polls are a bit dodgy, pointing out that in fact ICM's August polls prior to elections have a record of being very close to the actual result.

If you look at ICM's past polls their August polls before elections do indeed come very close to the subsequent election result. However, it is only ICM's polls where it works, not other companies. Neither is it

the case that ICM's polls don't change between August and the next year's election, in fact they then diverge away from it again.

It's just possible that this means something, perhaps August polls show underlying support at a time when the news agenda is very quiet. However, if a lack of political news did allow polls in August to see through to underlying levels of support and predict the next election, we would expect to see the same sort of thing with other pollsters, and we don't. With only three data points (1996, 2000 and 2004), it's just as likely that it is purely a happy co-incidence. Voting intention polls tell you how people would vote in a hypothetical general election tomorrow, they can't show you how people would vote in a year's time.