How surprising was Trump's victory? Evaluations of the 2016 U.S. presidential election and a new poll aggregation model

The U.S. presidential election results of 2016 surprised many poll-watchers, suggesting possible biases in estimated support for the major party candidates and posing a challenge for poll aggregation as a prediction tool. Using data from earlier elections and the 2016 campaign, we evaluated poll aggregation performance for the major prediction web sites. We found that a proportional bias, partly due to non-major party preference during polling, had a large impact on state-level estimates. A novel smoothing mixed effects model that is sensitive to both national and state-specific trends showed similar or superior performance to other aggregation methods over multiple elections. The improvement of the proposed model over competing methods was especially large in 2016, which we largely attribute to the late swing in voter support for the candidates, and suggests that the average bias in polling may be smaller than assumed. Simulations of electoral college outcomes indicate that, on the eve of the election, the probability of a Trump victory was about 47%. We suggest that an increased emphasis on fundamental statistical tradeoffs of bias and variance may be the key to further improved prediction.