The big day is a week away. Election Day has almost come, a fact we imagine any American using any form of media already knows given the near constant stream of reminders that election day is Nov. 3 and they ought to register and vote.
So, what’s going to happen? That is the constant question that will be asked, re-asked and examined from every possible statistical angle for the next seven days or so. Millions of Americans will be watching with bated breath on Election Night to see how things turn out — and whose predictive model did the best job predicting the outcome.
The stakes are high for election forecasters, as things didn’t quite turn out as expected last time. As Karen Webster noted six days after the 2016 election in a commentary: “November 8, 2016, was not a very good day for Big Data.”
She noted that the best work of some of the world’s smartest data scientists “using some of the most sophisticated predictive analytics tools in the world” to predict the election’s outcome missed spectacularly.
Two weeks before the election, most models showed Hilary Clinton was all but certain to win the 2016 election. But when the voters actually turned up at the polls, they narrowly elected Donald Trump the 45th president of the United States.
This time around, everyone is a whole lot more gun-shy about making definitive predictions about the election — chastened by the miss last time. As for what went wrong in 2016, Webster noted in her post-election column that a series of errors — seven deadly data sins, even — that created the malfunction.
She believes experts polled the wrong people, took answers at face value when they shouldn’t have and neglected to remember that polling is a time slice and people sometimes change their minds. Pollsters got stymied by statistics and got misled by groupthink.
Perhaps most critically, many of those problems were invisible to the experts which meant bad data got fed into sophisticated models that were all pre-programmed with a blind spot.
“Big Data doesn’t always deliver better data — or results,” Webster noted in her 2016 column. “All the algorithms in the world can’t overcome flaws in sampling or targeting criteria that can make the output of Big Data exercises unreliable — and undiscoverable — until it is too late.”
What Big Data in this case missed was what small numbers of voters in small counties in Michigan and Pennsylvania that no one was targeting with polls or predictive models were about to do. They were going to change decades of previous voting patterns and flip to supporting a Republican presidential candidate instead of a Democrat. Erie County, Pa., was one, and Monroe County, Mich., was another.
And while the shift wasn’t massive — between the two states, less than 100,000 votes flipped — it was just enough within the algorithms’ predictive blind spots to change the election’s outcome.
Unfortunately, if you’re looking for the proper understanding of how to read all the data right now to create a completely unassailable prediction model for 2020, you’re on the wrong site. We’re more interested in PYMNTS in the underappreciated lesson to be culled from Big Data’s big miss four years ago — the importance of tapping targeting correctly.
How It’s Playing Out This Time
If 2016 was a story about failing to consider voters in oft-overlooked geographies, the story this time is that no one wants to run the risk of making the same mistake twice.
That’s particularly true in a state like Pennsylvania, where misunderstanding the likely behavior of roughly 42,000 voters was critical in generating an upset outcome.
Part of the strategy now, according to local media, is to look at campaign advertising, where saturation on all possible airwaves becomes the goal.
But this year, advertising dollars are flooding into Facebook — because, as Nick Fitz, CEO of online donations company Momentum, explained to CNBC — most digital campaigns now rely almost entirely on Facebook regardless of how the people running them feel about the platform.
“For better or worse — mostly worse — Facebook is the de facto place you go,” Fitz said. “It’s the cheapest and most effective way to get in front of the right people.”
The right people, in this case, being the quarter-billion or so users Facebook has in North America spread across Facebook.com, Messenger, Instagram and WhatsApp. But it’s also because many other ad-supported sites have retreated from politics, leaving Facebook as the only game in town.
And among tech companies, Facebook simply makes it easier. While Google allows political advertising, it has always limited the targeting tools for political ads. Meanwhile, Twitter and TikTok have both banned political advertising entirely and Snap fact-checks them.
By contrast, Facebook does not fact-check ads due to what it says is a matter of free speech principles. And the company allows political advertisers to use the same kind of targeting tools as corporate advertisers.
The Home Stretch Changes
Facebook’s comparatively more permissive policy allows campaigns to do the all-important micro-targeting that ended up being so crucial to the eventual outcome of the last election. But that’s generating some controversy, as well as some unexpected friction points.
The social media giant is currently involved in disagreement with some New York University researchers requesting that they please stop attempting to collect data on how Facebook targets political ads.
The project is observing how roughly 6,500 volunteers who use a special browser extension, are targeted by keeping a record of what political ads they are shown. The goal, according to NYU, is to offer journalists, policymakers and researchers an impartial view into which ads Facebook displays to whom.
Facebook claims the project is tantamount to data scraping, which is firmly against the site’s rules.
“Scraping tools, no matter how well-intentioned, are not a permissible means of collecting information from us,” Facebook privacy official Allison Hendrix wrote in a 16 October letter to NYU, The Wall Street Journal reported.
But as of Tuesday (Oct. 27), there will be no new political ads to target on Facebook, as the site announced it will stop accepting political ads for the seven days prior the election. However, NYU said it has already gathered data on more than 200,000 ads, and that it has no intention of stopping its program.
How big an effect will microtargeting have this time around, when COVID-19 has radically altered the landscape and sent the economy as a whole reeling? Only time — and correctly calibrated data models — will tell.