I'd guess that there are a number of factors (aside from the conspiracy theories, GOP or Dems):
- People were embarrassed to admit voting for Trump (i.e., he was viewed
as the less respectable candidate), but that shame didn't translate to
the actual ballot. That doesn't mean that people voted for Trump on a
whim; it just means that they weren't keen on admitting that to someone
else, even a pollster they'd never see again.
- Exit polling was not done at all locations, for obvious reasons. So
projections were based on a regression analysis that fits estimates to
the sampled locations. That regression assumes, among other things, a
certain degree of polarization between demographics. It looks like that
polarization was even more extreme than expected (which was already
- Trump was simply a higher-variance candidate than the traditional Republican. This strategy makes sense in any contest where you're the underdog (as Trump was for most of the time)—if he were to play a low-risk strategy, he was almost guaranteed to lose. Employing a high-risk strategy increases the probability of a blowout loss, but it also increases the probability of a close win, which is what happened. We're seeing this all the time in sports, where endgame strategies by the trailing team are becoming more aggressive. That increase in variance translated to the polls. Five thirty-eight was very open about this—they pointed out that their model, though predicting a Clinton win, had about three times more variation (by some metric) in it than in past years.
I may have more to say about the election results themselves, but I'll save that for another post.
[Most of this post was drawn from a Facebook comment.]