Models are great but doesn't this all hinge on quality of the data inputs? In 2016 polls fielded the day before the election had Clinton up by 10 points in Wisconsin. Why did the polls and the election differ so greatly?
Note that these polls are of Likely Voters. A big source of systematic error in 2016 state polls was Likely Voter models, used to weight raw poll results by voting propensity. In the Midwest in 2016, lots of people without college degrees said they preferred Trump - but based on previous elections, pollsters expected few would vote. This kind of shift is hard to predict - pollsters do not intend to make the same mistake again, but they might make another. (National polls are much less noisy because these kinds of regional errors often cancel out when averaged over the country.)
So do 538 and The Economist try to take these polls, back out their biases, and reapply different ones? Or, to borrow a phrase from Trump, "unskew" them?
No, these models weight the polls by quality (variously judged), they don't reweight by voter. This weighted average of polls is then used, along with factors like time until the election and economic conditions, to give a probability distribution over election outcomes (which no poll can give you, no matter how you weight it).
"Unskewing" polls dates back to 2012 (Dean Chambers and UnskewedPolls.com). He was doing what you describe - applying his own likely voter model to all polls, based on party affiliation. It did not work. Better to use each pollster's likely voter model as it is, since then you have an ensemble and possible cancelation of error.
Even though there were more people supporting Trump in Wisconsin (or rather any non-left candidate), they couldn’t tell that in polls openly because is goes against vocal minority.
Online surveys are even worse than in-person, because they are hosted on particular sites, having particular audience, leaning or biased in particular way.
E.g. Poll on cnn.com and fox.com will give you opposite results.
Also different age/wealth/profession groups have particular leanings on political spectrum and particular average level of computer proficiency.
E.g. a farmer in Idaho would likely to vote for any Republican candidate and simultaneously less likely to participate in any online polling.
Here's the last poll Google collected in Wisconsin in 2016: https://datastudio.google.com/s/kcPRGFaMaO0