A mention in a comment on Wang’s site led me to PollyVote and I thought some here might find it of interest - not for it’s reassuring take on the probable outcome (although that be nice) but out of interest in its methods and performance to date.
The method:
Think of it as 538’s PollsPlus (which includes econometric fundamentals) plus adding in other aggregators, and adding in four other approaches.
Problem I see with it? The forecast is for a two-party race. Also there is bleed between methods, for example prediction markets are informed by read of the other methods.
Anyone heard of this one before? Thoughts about it?
Might not be all that unrealistic: today Carl Bernstein broke the news that Libertarian VP nominee Bill Weld is worried enough about the potential to help Trump become President that he’s considering stepping out of the race. Of course Johnson would keep going even if he received a heavenly messenger telling him his involvement will definitely give the White House to Trump, but Weld’s dropping out would weaken Johnson’s spoiler ambitions.
That would only make sense if one assumes that the Libertarians will pull more votes from the Democrats than from the Republicans, which at the very least isn’t something one ought to take for granted.
And meta-studies like this have a tendency to just amplify noise. You’d be much better off figuring out which predictor is most reliable, and then just following that predictor.
Another aggregator (polls plus other stuff). If their track record is as good as they say it is, then I’ll add them to my list (which currently has 538 on top, followed by PEC, along with the prediction markets. If after the election they turn out to be as good or better than the others, then they’ll move to the top of the list.
Pleased to see the good prediction, but I won’t trust it until I can see and trust their track record (which means I can’t trust it until after this election, assuming they do well).
Do you have any basis for that opinion? Not snarky, honestly curious…
The data they supply seems to go against that proposition, and suggests that combining across methods decreases the potential errors in each other sort of method. Often that is a means of dealing with noisy signals, sample by different means (or at different angles, so to speak) and average it out: the idea being that the noise of different methods tends to cancel out leaving more true signal. Here’s one of their cites. It’s an old review but without data that contradicts it it seems likely still valid.
No magic bullet but maybe a wee bit value added?
iiandyiiii except that polling aggregation (in aggregate) is just 1/6th of the weighting … officially. Which I throw in because of course polls inform at least one of the other methods used.
So if other parties get 5% then the prediction calls for Clinton to get 53% of the remainder, i.e. 50.35%, round to 50.4%, and Trump to get 47% of the remainder, i.e. 44.65%, round to 44.6%.
Reading reviews here seems to be little debate over their past accuracy. Still as theWiki article about them suggests as a possible reason for their lack of popularity with the press: “PollyVote predictions are very stable and rarely change, whereas election observers and journalists are interested in excitement and newsworthiness.”