I'd say the 'how' lies in voter prediction models that don't match the actual voters who go to the polls. Every pollster looks at past data to try and predict the 'balance' of how many folks of each party, or indeed each ideology, will actually get out and vote. And for your generic elections, those tend to work well enough. Where they fall down is in 'weird' elections, when the general populace isn't behaving as they normally do. Then pollsters have to decide if they'll stick with the modeling that's worked for them forever, or try and guess how the zeitgeist will change the balance of who actually votes.
I think we've gotten to a point where even calling people on phones isn't necessarily accurately representing the voting public. A lot of people simply blow off phone calls from numbers they don't know. So your polling data comes from the sort of people who are more willing to answer odd numbers, or who WANT to participate in polls if the caller ID actually shows it's a pollster.