Several problems with Nate Slaver’s methodology:
1) uses historical precedent for polling agency accuracy and skew
Wouldn’t polling agencies modified their methodology over time to become more accurate?
Why does he discount the USC Dornsife and PPD tracking polls? Similar polls have been shown to be quite accurate over time.
2) Goes “conservative early, aggressive late”
This is clearly stated as part of part of the methodology.
They are willing to put much more weight on new polls.
But the swing over time in the polls is indicative of something not captured in his methodology.
3) The adjustment for poll skew is to the mean, not to the shape of the prior distribution of the sample representativeness
This is a departure from the Bayesian methodology he purports to follow.
To adjust for sampling skew, it is not enough to assume a normal distribution, but rather one would need to adjust the entire distribution.
But this could generate spurious results, such as what happened with the “unskewed” polls in 2012 that showed Romney could win.
It’s better not to adjust for poll skew at all.
I’m not saying Slaver is wrong, just that these are potential sources of error in his results.
Yes. He TOTALLY missed it with the NCAA and NBA championships this year. That's a good thing.