Bing Predicts kicked off the 2016 elections with the goal of providing user insight, starting with the primary season across all states and territories. Through all primaries, the model’s accuracy was 82 out of 98, or 84 percent. The following discussion analyzes more deeply how Bing Predicts did over the primary season.
In addition to predicting a single winner for each state, Bing Predicts also estimated the delegate assignments and kept a scorecard for the predicted final delegate count so that those following could extrapolate with us how many delegates each candidate could finish with.
Delegates predicted before June 7th primaries
With the focus now moving toward November, we have also added information in the Snapshot showing the electoral vote count predicted, updated on a daily basis.
Impact of non-traditional candidates
Up to Super Tuesday on March 1, Bing correctly predicted all of the primaries except for the Republican Iowa caucus. For those interested in understanding why: The models are built to evaluate how public anonymized web and social activity, polls, and other salient signals correlate to voting patterns. For existing politicians, this information is well-calibrated over time, so for the Democrats, the vote shares of the two main candidates were understood even before the very first primary in Iowa.
On the Republican side, due to the fact that there were multiple candidates in consideration that did not have prior election experience, the Iowa caucus was the first vote share available for candidates such as Mr. Trump and Dr. Carson. After this information became available, future primaries on the Republican side became easier for the models to calibrate.
One of the goals of the Bing Predicts election experience is to provide insight to users regarding which candidates should win a given state. In some cases, the outcome is effectively a toss-up. In the state of Missouri on March 15, both the Democrats and Republicans forged virtual dead heats at the top, with Clinton being declared the winner over Sanders by a margin of just .2 percent. For the Republicans, the outcome was withheld for a long time, with Trump leading Cruz 40.9 percent to 40.7 percent. The point about extremely close contests is that the result could go either way, and predicting the winner for such a case is equivalent to guessing whether a coin toss. For the Democrats, the Bing models thought it would be Sanders. For the Republicans, the models thought it would be Trump. In either case, though, one could argue it could have gone either way.
The takeaway here is mainly that our models seek to accurately inform in advance the winners for contests where there will be a clear winner. In cases where two candidates have almost the same count, difficult-to-predict factors about voter turnouts or voting variances can determine a winner.
Getting closer to the convention
As April came around, we observed that people wanted to know whether there would be contested conventions, so we built out a delegate tracker which was perhaps as important as finding out who would win each state. Because many states do not employ winner-take-all delegates, taking our predicted vote share and extrapolating out to the final delegate counts was important to let our users know whether the Republicans would have a contested convention. As the numbers showed, we did not think there would be one, and when Mr. Cruz and Mr. Kasich dropped out in early May, that projection became a certainty.
As we move to the conventions and then turn our focus to November, check regularly at Bing to see our latest projections for this election season.
- Walter Sun, Bing Predicts Team Lead