The overall number is nothing but the result from each individual state. Your rebuttal doesn't make any sense. If you look at the results of each part contributing to the whole of course you can draw conclusions about what led to an inaccuracy in that result.
The. States. Are. Not. Independent. Data. Points.
There is variable bias! And it's not omitted!
When I say there's a 10% chance of a Hillary blowout, that means the
collective probabilities of these individual states equates to a 10% overall chance.
How do you distinguish between a 10% or a bad model? If you look at the individual states, which have much higher probability of being a blowout than 10% and when one state is a blowout it raises the probabilities of the other states of being a blowout, then how do you make the assessment you are asking to make?
The parts are all interrelated.
Each poll 538 looks at is weighted and adjusted. Each one is an opportunity to evaluate results. Which polls turned out to best represent the final result? How did the model treat them? I am baffled how so much data would be useless in your eyes. These are the decisions the model is making.
You're evaluating PAST RESULTS with no predictive power for future elections. Sigh.
I'm not saying the data is useless. And I'm not saying Nate shouldn't adjust his weights to polls after this election. He should.
What you don't seem to get is this aspect of adjusting the model doesn't actually have anything to do with whether the model's core algorithms are correct or not.
Here's my point. Even if Nate weighs every pollster correct...his model can still be wrong. Because his other inputs that are
unrelated to polls may be completely wrong and there's no fucking way of knowing.
How do we test if the UE rate has an effect? How about jobs data? What if THOSE data points are the problem with his model???