Remember Left Foot Forward’s bold extrapolation of the political polls which I had a bit of fun with on this site on 5 March? It projected polls from January to March as a straight line which would have Labour overtaking the Conservatives round about polling day.

The habit seems to be catching. Labour List, another blogging site with a leftward lean, has taken the process a step further. But instead of fitting straight lines to the scatter of points, it’s decided that a polynomial trend is a better option.
 
Its justification for this is impenetrable, and as one comment on the site asks, why should one believe that a second-order polymomial  is the best fit to the data? It generates a nice curve, true, but one might have hoped for a better reason than mere aesthetics.
 
The site credits Peter Kellner of YouGov for his kind advice, so maybe there is more to it than meets the eye. But I think if you’re going to extrapolate using polynomials, you should provide some sort of justification for doing so. As another comment puts it: “Can it really be claimed that if the rate at which the Conservative lead falls slows down, then that means the lead is likely to start growing again soon?”
 
    
 
Discouragingly for its readers, the curves Labour List projects (above) show the Conservative lead widening just enough to ensure a majority in the House of Commons on May 6. But it does helpfully point us to the source of the polynomial idea, a US site that in September 2008 unveiled its “curvilinear polynomial regression model” to predict the proportion of the popular vote that Senator McCain could expect to win in an election then still 60 days away. That site doesn’t provide any rationale, either.
 
It concluded, 60 days before the US election, that Mc Cain would lose the popular vote by a margin of 3.1 per cent. The actual margin was 7.2 per cent. So it got the result right, but I’m not sure the accuracy of its prediction really constitutes convincing evidence for the model. Here's its Septmber 3 projection.