USA Today's Jim Norman asks why the polls all predicted the Democrats would do better than they eventually did. I believe he unwittingly answered his own question.
As he notes, in 2012 the polls systematically overestimated the Republican turnout. In 2014 they overestimated the Democratic turnout. I'd guess the answer is buried in that pair of facts.
Following their miss in 2012 pollsters rechecked their assumptions about turnout and adjusted their modeling of the electorate to reduce pro-GOP bias. I'd guess they adjusted it too much, believing that the Democrats vaunted "ground game" would succeed in getting the young and minorities to the polls in a midterm election. That belief proved to be false.
At a minimum a pollster needs separate models for presidential elections and midterm elections. Needed are models that reflect the older, whiter, more conservative midterm electorate and the younger, more diverse, more liberal presidential electorate.
