The poll aggregators were wildly off as they gave Hillary Clinton an over 90% chance of winning. Nate Silver was the most pessimistic because of his theory of correlated forecasting errors:

State outcomes are highly correlated with one another, so polling errors in one state are likely to be replicated in other, similar states….If Clinton loses Pennsylvania despite having a big lead in the polls there, for instance, she might also have problems in Michigan, North Carolina and other swing states.

The correlating factor appears to be the white working class vote which abandoned Clinton in the Rust Belt States. For reasons we do not yet fully understand, this vote for Trump did not appear in polls. The polls were then all off the mark in all Rust Belt states.

The others poll aggregators assumed independence and gave less thought to a simple theory of voting that might generate independence let alone whether such a theory might be plausible. If they hand, it would have led them to a more plausible way to look at the data, like Silver, and hence better predictions.

(HT Georgy Egorov though I may not be doing justice to his point.)