The Polls and Predictors Were Off, But Not By as Much as It Seems
Going into Election Day, political forecasters predicted the most likely outcome was a victorious Hillary Clinton, a Democratic Senate, and the House still firmly in Republican hands. Instead, Donald J. Trump emerged as the 45th president of the United States, while the GOP managed to hold on to both houses of Congress.
Why were these top-line forecasts ultimately so wrong? A deeper dive into the state-by-state and down-ballot predictions we’ve been tracking since the primaries shows that the failure to predict the Republican clean sweep comes down to just a few limited (though consequential) mistakes. In the presidential race, all seven forecasters we examined (including poll aggregator RealClearPolitics) incorrectly projected Clinton would win Florida, North Carolina, Pennsylvania, and Wisconsin—collectively worth 74 electoral votes out of the 270 needed to win. On the Senate side, the two universal misses were Pat Toomey's unexpected reelection in Pennsylvania, and Ron Johnson's defeat of former Senator Russ Feingold in Wisconsin.
That’s not to say that the forecasters didn't wildly underestimate many of Trump’s ultimate margins of victory. Take Ohio, for instance, where Trump was leading Clinton by more than eight points on Wednesday morning: an average of four pre-election estimates predicted he'd win by a mere 1.6-point margin.
Of the four predictors we tracked that assigned probabilities to Clinton's chances of winning, the odds ranged from a respectable 71 percent in FiveThirtyEight's "polls-plus forecast"—which includes polls, as well as economic and historical data—to a sky-high 98 percent in the Huffington Post's mainly poll-driven model. Senate Democrats were also universally, if less heavily, favored to retake control of the chamber.
As a result, no individual forecaster can claim an “A” grade for Nov. 8. That said, FiveThirtyEight could arguably take home an honorable mention for its relatively more Republican-bullish outlook. The site's founder Nate Silver even received a Twitter apology of sorts from the Huffington Post's Ryan Grim, who over the weekend had savaged him for "putting his thumb on the scales."
For forecasters looking for a silver lining amid the wreckage of their months of hard work, the overall accuracy rate of the models we analyzed remained quite high.
As this table shows, prediction market aggregator PredictWise correctly projected the winner in 42 of the 46 states called as of Wednesday morning, for a 91.3 percent accuracy rate. Five others matched this record, including FiveThirtyEight, the New York Times' UpShot model, the Huffington Post, Bing Predicts, and the Cook Political Report. When looking at Senate races, PredictWise correctly called 29 out of 31 races so far, a track record also mirrored elsewhere. (This count excludes the California Senate race, where two Democrats faced off.) As for the House, PredictWise managed to correctly predict the partisan makeup of nine out of every ten House delegations, based on those seats called so far.
What's left now is for the pollsters, pundits, and prediction markets to figure out what, if anything, they can do to improve their forecasts for when the next elections roll around in just two years.