Read on Twitter
Polling aggregators like @NateSilver538 at @FiveThirtyEight are making the same mistakes they made in 2016, giving poor-performing state level polls too much weight. Consequently, they exaggerate the odds of a Biden win at 87% (they gave Clinton 71% odds). 1/
I've been digging into their data to understand the election. First, kudos to them for making everything available and taking-on hard work. This is meant to be constructive criticism.
There is a very weak relationship between how they grade pollsters & 2016 performance
There is a very weak relationship between how they grade pollsters & 2016 performance
This matters because their polling aggregations and prediction models give more weight to pollsters with better grade, but their grading system is not well-aligned with 2016 performance at the state level.
Even worse, their grading system seems to be biased against pollsters that show large Trump margins in 2016 relative to actual. I took the difference between a ranking based on 2016 results & their poll ranking. Their rating bias was pro-Democratic Party.
Take a concrete example. 538 gives @trafalgar_group one of the worst ratings in their system (11/15), but they should be in the best group based on their 2016 performance. They had very low errors & correctly called MI, FL, and several other swing states no one else got right.
As it plays out, 538's models give Biden several swing states (AZ, NC, FL, and MI) that Trafalgar has Trump leading. 538 is using polls that had a pro-Clinton bias by a large margin in 2016 & giving them more weight than Trafalgar, without any obvious correction.
Given these results ( https://www.thetrafalgargroup.org/ ), it seems highly misleading to say that Biden has a 87% change of winning. These swing state polls are much tighter than a superficial reading would suggest based on 2016 performance. I don't have any confidence one way or another.
