90-year-old Chinese American Cy Yen Liu, center, checks her ballot as she waits behind ballot-filling booths at a polling station in Flushing section of the Queens borough of New York, Tuesday, Nov. 8, 2016. Liu, originally from Shanghai, China, immigrated to the United States 50 years ago. (AP Photo/Alexander F. Yuan)
© AP

Polls are not predictions, though they inform them, and predictions are not guarantees. To hear the criticism being hurled at America’s pollsters, though, you would think a terrible travesty has been committed, that the widespread failure to predict Donald Trump’s victory represented a dereliction of some duty to democracy.

It is nothing of the kind. Polling is an imperfect science and Tuesday night’s result was within the national polling margin of error. Nate Silver, a popular diviner of polling data, wrote on Wednesday: “If Clinton had done just 2 points better, pollsters would have called the popular-vote margin almost on the nose and correctly identified the winner in all states but North Carolina.” But she didn’t.

Where the polls offered least was in the rust-belt states which handed Mr Trump his victory. In Wisconsin, for example, all the polls leading up to the vote had Mrs Clinton leading by as much as eight points. Trump beat her by 1 per cent. Mrs Clinton’s team believed so strongly that Wisconsin was a lock, based on their internal models, that she did not visit that state between securing the nomination and election day.

One might imagine in this age of “big data” they should have known better. Enter a search term on Google and before you know it every website you visit is blitzing you with targeted advertising. Amazon has a good idea of what you will buy next. Facebook learns a disarming amount about us from our posts and our circles of friends. Our iPhones track us as we move. The crumbs we leave along our digital trail add up to great insight and great fortunes.

But polling and political predictions are not yet a truly big data business. Pollsters deal in past election results, responses from 1,000 or 2,000 people at a time, census data and voting records. In smaller communities, the polling can get pretty sketchy.

A proper big data approach to predicting voters’ intentions would plunder all kinds of public and private data sets — the petabytes of information that make up our digital identity — to eliminate some of the unevenness of polling. Personal credit scores, for example, might tell you if people with histories of financial delinquency preferred Mr Trump. Internal Revenue Service data could have led his campaign to people who take a lot of tax deductions. Mrs Clinton could have examined health insurance pools to find voters who were grateful for the plans introduced under Barack Obama’s Affordable Care Act.

Such data sets, however, are not readily available, and pollsters suffer from more basic limitations. When everyone had a landline and their name and number appeared in the phone book, polling was a simpler business. Today, many younger people own only unlisted mobile phones. And those of us who still have landlines know better than to answer an unknown number calling during dinner. Finding a good mix of people to poll is hard.

Shortly before the election, Jim Messina, the manager of President Obama’s re-election campaign in 2012, wrote in The New York Times praising the Clinton campaign’s mastery of voter targeting. “‘Big data’,” he said, “is a buzzword, but that concept is outdated. Campaigns have entered the era of ‘little data’.” Little data campaigns are driven by “direct, highly personalised conversations with voters both on and offline”.

These conversations create a more granular picture of voters, which guide how candidates directed their time and efforts. They, presumably, are what guided Mrs Clinton’s disastrous decision not to visit Wisconsin during the key campaign period.

A couple of changes since 2012 may have derailed the approach that proved successful for President Obama. The first is that young people, who are most likely to engage in direct, online, personal conversations, were not as important in this election as they were in 2012.

The second is that, for many voters, supporting Mr Trump was not something they bragged about. It was telling that some pollsters found that Mr Trump did better when voters shared their intentions with a recorded rather than live voice on the phone. Many didn’t want to share the truth with a real human being.

Third, I suspect Mr Trump was far more savvy about voter data than anyone imagined. On election night, the CNN host Anderson Cooper joked about visiting Mr Trump at his home in Palm Beach a few months ago and being stunned by how small his campaign team was. The young man snoozing by the pool, Mr Trump’s campaign manager told him, was “our digital guy”.

I don’t believe it. Mr Trump and his companies are seasoned consumer marketers. The feedback from his much derided Twitter rants must have given him a strong sense of where his support lay — and the kind of messages they wanted to hear. Hiring Stephen Bannon of Breitbart News, a svengali of the digital right, and using as a consultant Roger Ailes, the recently dethroned chairman of Fox News, will also have given him powerful and novel insights into how to direct and corral voters.

Polling remains a precious tool for politicians and the media. But it is fallible. This year, statisticians set a low probability on a Trump victory. They made a “Black Swan” out of what was closer to a coin toss — to Mrs Clinton’s great cost.

The writer is author of ‘What They Teach You at Harvard Business School’

Get alerts on Donald Trump when a new story is published

Copyright The Financial Times Limited 2019. All rights reserved.
Reuse this content (opens in new window)

Follow the topics in this article