A small, somewhat quirky polling firm in Atlanta went against the grain. It predicted Trump victories in Michigan and Pennsylvania on the eve of voting and turned out to be right when all of the national firms were wrong.
Trafalgar Group and its seven employees discovered during comparison polling that some Trump voters would not disclose how they planned to fill out their ballots. After all, the liberal news media and Mrs. Clinton had labeled Trump followers as deplorable people.
So the company used robotic calls for which Trump voters seemed more comfortable. They also added a “neighbor” question, figuring that a respondent would be more willing to answer truthfully if a neighbor was voting for Trump.
The also created a demographic of people who had not voted in a half-dozen years or so but planned to vote for Mr. Trump. To capture these voters, Trafalgar created a large sample, about 1,000 respondents, compared with a usual poll of 400 or so.
The result: A day after the election, when searchers looked at the Real Clear Politics webpage for Michigan polls, there was a long list of predictions, all of which were in blue because they had Mrs. Clinton ahead — except one posted on Nov. 6. It had Mr. Trump winning by 2 percentage points. He was leading by 0.3 percent with all precincts reporting. That outlier poll was by Trafalgar Group.