Which of course raises the question of how accurate the model is. I don’t know the answer, but apparently the reason it differs from pure polling is that this far from election day it considers fundamentals like the economy and domestic and foreign policy to be disproportionately important. As election day draws nearer the polls become more heavily weighted. Below is a brief blurb from a Twitter post by someone involved with it, the editorial director of data analytics at 538 and ABC News.
Addressing some confusion about the 538 model:
Broadly speaking there are two types of uncertainty in the model: Uncertainty about how states vote (which we show you) and uncertainty about how they relate to each other (which we don’t show you).
Specifically, we have uncertainty about the correlations between states, informed by (a) a model of the polling data and (b) a model of historical fundamentals data.
Given the disproportionate amount of data and lack of measurement error in election results, we have a much better prediction of historical fundamentals-based correlation than polling correlation.
This influences the “weight” put on each variable. That works like so:
The model has the ability to guess how much polls should revert towards the fundamentals by Election Day. It does this by adding a random variable drawn from a multivariate distribution.
This amount of movement (by design) is not mean-zero across states or simulations. Polls can move up or towards fundamentals — by less when we’re close to the election and more when we’re far away.
Usually this splits the difference between the polls and fundamentals, but the model can push vote shares outside the range of the observed polls and fundamentals if that’s what the model needs to do to make the correlations between states match expectations.
This gives the appearance of us adjusting the forecast to Biden in some close states, eg WI, but it’s because the polls and fundamentals elsewhere indicate he should be doing better than he is in WI — and the model is balancing those correlations between states,
Important! This adjusting will decrease as we get closer to Election Day (and is gone on E-Day) since there is less (no) uncertainty to simulate about how things will move by Nov 5
Finally, this also has effects for the national popular vote. Less uncertainty about the correlation for the fundamentals than the polls increases weight on the fundamentals since the national forecast comes from aggregating the state estimate.