Missed Signals in the Presidential Campaign
This post isn't about lament or triumph. It's not about taking a position on politics. That's not what we do here; we're a platform for venture capital. It is about the data behind politics and how flawed assumptions can lead a campaign, much like a startup itself, to take actions and make plans that sow its demise.
Political campaigns can be viewed very similarly to startups. They start with an idea, a candidate. Then it's on to raising some money and searching for product-market fit. Presidential campaigns find or don't find this fit during the primaries. Those that emerge from this test effectively operate like unicorns trying to gather as much market share as quickly as possible.
They spend hillocks of cash and hire people at impossible rates, ramping up from a one-room operation to a nationwide network that touches nearly every American in some form. Campaigns, like startups, rely on data to guide their moves and inform their decisions. Statisticians working in Python and R pore over new batches of numbers, coming in from disparate sources in disparate forms. Abstracting meaning from it all has become paramount in a country whose presidential elections now tilt on the tiniest of margins.
Misreading the data or operating with false assumptions about the data, or false assumptions about product-market fit, can subvert an entire startup, even a unicorn. The same, again, is true for campaigns. Hillary Clinton's campaign failed to see the market move. It made assumptions about the data and its extrapolations that may have been true in 2008 and 2012, but simply weren't correct in 2016.
That is why she lost.
The Clinton campaign followed the construct built by the 2008 and 2012 teams that piloted Barack Obama to victory. It assumed that the market was operating with the same tendencies and rules, and that its product would fit just so. But the rules weren't the same. It wasn't just that voters voted differently. The biggest difference was that different people voted. People who had turned out for Obama in some cases didn't turn out at all for Clinton. They simply didn't vote. And many who hadn't voted at all before now cast their ballots for Trump.
To be clear, it was an easy mistake to make. Spotting the error that had been missed, its effects then magnified by so many others and then magnified more through myriad projections, was nearly impossible.
Nate Silver, the founder of FiveThirtyEight who had projected 99 states out of 100 correctly in the previous two elections, said this in a video at his site:
"I worry that when people look back on this election, they'll find a way to normalize it and make it seem as though it was obvious and inevitable that this was going to happen. Don't minimize this in the rear-view mirror."
Even the one major poll that consistently projected Trump ahead in the race, the USC/L.A. Times poll, didn't quite get it right. It showed Trump ahead by three points in the popular vote, which he will likely lose by several hundred thousand votes.
But there was a shred of insight inside that poll, which did more than simply ask people for whom they expected to vote. It asked voters to rate, on a scale of 0 to 100, the chance that they would vote for Trump, Clinton, or a third party candidate. That measurement by itself was more granularity than most polls drew out, and it helped capture ambiguity within different voting blocks. But the USC/L.A. Times poll asked one more key question of those it surveyed: rate on a scale of 0 to 100 the likelihood that you will vote–thus giving it another critical edge.
Most polls, with their binary inputs, didn't reflect the nuance of a human electorate. And these polls were the fuel for projections at sites like FiveThirtyEight and the New York Times, which proved to be far off. On the eve of the election, FiveThirtyEight gave Trump a 29% chance to win, which was the biggest number among the major sites making these projections. The New York Times' Upshot put Trump's chances at 15%.
Early on election night, multiple reports out of the Trump camp indicated that its own staffers didn't expect to win. Trump's data team, based in San Antonio, had projected Trump with only a 7.8% chance of winning the election as late as Oct. 18.
But in the two weeks that followed, the campaign noticed that early voting seem to indicate that older white and rural voters were voting at a higher rate than was normally projected. Matt Oczkowski, the head of the Trump data team, told Bloomberg that they adjusted their projection model to account for bigger turnout among this group. The model still showed only a 30% chance of Trump winning, but it showed a narrow path to victory: one that led directly through Wisconsin, Michigan, and Pennsylvania, ground that the Clinton campaign assumed couldn't be lost.
Based on this insight, Trump spent a good chunk of time madly campaigning in these states during the last two weeks of the election. His campaign saw a flaw in its larger, favored competitor and it poured all of its resources into the breach.
Some of the biggest startup successes follow similar patterns. Stripe went from a startup to being worth multiple billions of dollars by seeing a weakness in card processing incumbents: their APIs were hard to use and took major effort from developers at startups to be implemented. As the larger companies in the space ignored startups, Stripe rode their growth, evolving its product with them.
Similarly, the Clinton campaign didn't see its weaknesses clearly, which were even deeper than an outsized rural vote. Many of the urban voters that turned out so heavily for Barack Obama in 2012 simply didn't show up for Clinton. The campaign's projections assumed a level of turnout similar to Obama's. It was an assumption that likely lost Clinton the election.
In Wisconsin, where Clinton lost to Trump by 27,000 votes, 60,000 fewer voters cast ballots in Milwaukee County compared with 2012. Clinton received 66% of the vote in Milwaukee County, meaning that a similar turnout just in that county would have won her the state. The situation was similar in Michigan's Wayne County, where Detroit is located. Turnout like that of 2012 would have given Clinton the state. It will never be known for sure, but Trump's focus on these two states in the closing weeks, his countenance and words dominating news cycles there, may well have been the difference.
The Clinton campaign wasn't alone in making this mistake. It's something we see in the startup market all of the time, of course. Venture capitalists assumed for years that the grocery delivery market couldn't be cracked in a big way because of previous failures in the space, such as Webvan. Instacart seized on the rise of the gig economy and found a seam that nobody else had considered—and Sequoia Capital was wise enough to believe that the market had changed enough to take a chance after missing in this space previously.
Startups need to constantly review their own assumptions about their space and their data—and how those might differ from the market at large. When discrepancies root up, it can either signal opportunity, or it can be a warning that a company has chosen the wrong tact.
Microsoft, once the biggest software incumbent of all, famously missed the rise of mobile. That miss was one of the things that's made Apple the extraordinary story that it is. The Clinton campaign was the effective incumbent here. It missed the signal.
It's not that the Trump campaign saw the signal with clairvoyance. But it saw the possibility of it and how any victory, however slim the possibility, would have to arise. This path being the campaign's only chance at success, as conventional routes had been closed, it put everything against it.