Read on Twitter
What if I told you the polls were actually pretty decent this year?
Pretty simple stuff but I went through and laid out each state's final popular vote prediction from @NateSilver538 and compared the difference...
Pretty simple stuff but I went through and laid out each state's final popular vote prediction from @NateSilver538 and compared the difference...
that column on the right looks pretty bad right? a red color means the prediction was underestimating Trump, blue means it was underestimating Biden. And hey that's a lot of red and the average is 4.5% that's a bad miss and we all underestimated Trump! But...
We forget that many states are still counting mail-in votes, not just the ones we're all obsessing over. So lets look at the % unreported and build in a small guess to consider that we're seeing a strong Biden lean in mail-in votes everywhere (except AZ weirdos).
Now it's getting interesting, still plenty of red but the average is only a 1.56% miss. And some states start to look insanely accurate like GA, FL, AZ, MA etc... and 29 states within a 3% margin of error.
Then we carve further because states like Montana don't get much polling data since they are so lopsided.
So over there on the right hand side is filtering out the states that had predictions thinner than 6% points. And we get a 1.43% predicted miss average.
So over there on the right hand side is filtering out the states that had predictions thinner than 6% points. And we get a 1.43% predicted miss average.
I think when all is said and done the demise of the accuracy of polling will be greatly overstated.
There does seem to be a fairly consistent underestimating of Trump voters which needs to be understood, but this "polling disaster" I keep hearing about is part of the red mirage.
There does seem to be a fairly consistent underestimating of Trump voters which needs to be understood, but this "polling disaster" I keep hearing about is part of the red mirage.
Only caveats in the data here is Alaska and New York have significant outstanding votes that I can't predict all that easily. And the mail-in bump is a modest guess laid out evenly among states, not a county level analysis which would be better.
So a major takeaway of 538 is that they always build in margins of error and uncertainty into the models and predictions but those are general errors in either direction. What they should have built in was the possibility that there was a specific error favoring one side.
The context of Trump being an anomaly and the pandemic being a wild variable the possibility that an error was built in everywhere that slanted a specific direction was real and undercalculated. For logical reasons that error was more likely to bias against Trump.
If that factor was built in, it would have bracketed its certainty of predicting the winner as it had consider that it could be compounding a bias. State by state stuff actually looks good. But that damn fox with his 90/100 odds for Biden was a dumb overconfident little dweeb.
Maybe a better way to calculate the average error I'm projecting. Separate the Biden bias misses and Trump bias misses and the average lands at 2.5% average error from actual. And close call states lands at 1.47%. Still not too bad.
An error with Ohio was pointed out to me. True. I had 43.4 for trump supposed to be 53.4
This makes the current error 4.7% and predicted errors 2.54% for and predicted swing states error 1.47%.
This makes the current error 4.7% and predicted errors 2.54% for and predicted swing states error 1.47%.
Had an error in NC too.. listen im not a data scientist I just play one on Twitter.
