2016 Postmortem

from enough people to extrapolate the results from their raw data. What they do, instead, is go through their sample and adjust their sample to make it look demographically like the sample that they believe will vote, and that is substantially based on the previous election cycle exit polls, plus voting record, for Likely Voter polls.

There are several sources of error contained in this method, but the most obvious (and most lethal) source of error is a change in turnout between elections. In essence, the method they use would be very accurate if they already had the exit polling for the election they are trying to predict, but of course they don't have that yet.

Turnout changes from election to election are hard to detect in advance. When both candidates have high unfavorables, as they did this year, they can be a bear. No matter how one tries to adjust the Likely Voter screen, it will be wrong.

Checking against exit polling shows that turnout changes, high third party candidate involvement and most specifically, high turnout in some areas shifted the results. But it also confirms that the voting tallies were accurate.