The late shift toward Trump, like other periods of polling instability throughout the campaign, was consistent with a long-term pattern. Historically, the more undecided and third-party voters there are, the more volatile and less accurate the polling has tended to be. The relationship ought to be fairly intuitive: Theres not much a pollster can do when a voter hasnt yet made up her mind. In 1980, for instance, final polls showed Ronald Reagan leading Jimmy Carter roughly 43-40, with 17 percent of voters undecided or saying they planned to vote for independent John Anderson. Reagan wound up winning in a landslide, 51-41, making for a seemingly massive polling error. But Carter technically didnt underperform his polls: its just that Reagan hugely outperformed his as undecideds and Anderson voters broke his way. A milder example came in 2000, when a relatively high number of voters said they were undecided or planned to vote for Ralph Nader, presaging an upset win in the popular vote by Al Gore (George W. Bush had led by 3 to 4 percentage points in the final national polls). By contrast, the 2004 and 2008 cycles had very few undecided voters and highly accurate polling.
FiveThirtyEights model accounted for the high number of undecideds, which is part of the reason it gave Trump better odds than other forecasts. Im somewhat perplexed as to why these voters didnt draw more attention from other modelers or reporters, however, since they were often key to understanding the progression of the campaign. With a large fraction of voters not firmly committed to either candidate no doubt in part because of the historic unpopularity of both Clinton and Trump it didnt take much to move them from one candidate to the other, and so news events had more impact on the polls in 2016 than they did in 2012.
