Prediction Machines: Looking Backward to Understand America’s Obsession with Behavioral Forecasting

Another heart-stopping U.S. national election has come and gone, even if, at this writing, President Donald Trump refuses to accept the result. One of the highest-stakes political contests of Americans’ lifetimes, it would have been nerve-wracking no matter what. But the experience of watching the returns roll in was all the more wrenching because our experts in prediction were off their game. After offering months of daily intelligence—on the candidates’ respective standing in battleground states; on likely shifts in the composition of the House and Senate; on the ways differential turnout rates, the expansion of mail-in voting, the battle over a new Supreme Court nominee, and the public health crisis might affect the course of the election—and updating our national vocabulary to incorporate phrases like “shy Trump voters” and “red mirages,” those whom we trust to forecast elections in the United States projected a decisive victory for Joe Biden and significant Democratic gains in Congressional races. That is: in 2020, they seem to have failed, and badly, once again.

By the time this essay is published, a rash of detailed post-mortems on the 2020 polls will already have been written, and there will be more to come. But it seems clear that professional prognosticators did not come close to measuring the strength of Trump’s support among white voters without college degrees—or for that matter, Latinos—and were completely off base about senior citizens trending in a big way for Biden. The old confidence that high participation benefits Democrats disproportionately and the new one that voters in COVID-19 hotspots would turn against Trump likewise evaporated in the days after November 3.¹

You might wonder: did 2016 teach us nothing? That was the year that we were assured with percentages too lopsided to leave room for doubt—a 98.2% chance according to the Huffington Post; a 99% likelihood according to Sam Wang’s Princeton Election Consortium model—that Hillary Clinton would be the next United States president. Disbelief, shock, and recriminations followed. Even so, four years later, it proved difficult to resist these quantitative glimpses of the future, slivers of certainty in terribly uncertain times. In 2020, pollsters and data scientists trotted out new-and-improved forecasting models, tweaked to better their 2016 performance, which, analyses suggested, had misfired primarily because state surveys had failed to properly weight voters’ educational levels. Nearly all of these shiny new productions, up until the very end, painted a bleak picture for incumbent Donald Trump, who appeared to lag significantly in the projected popular vote, the Electoral College calculus, and an array of battleground states that had been key to his victory in 2016. Buoyed by such reports, many on the left cautiously traded their anxieties for the comforting projection of a post-Trump politics.

Biden of course did win the election. But the polls’ inability to capture the closeness of the race in any of the battleground states and their underestimation of Trump’s support in all of them but Arizona fed conservative narratives about media bias, bolstering Trump’s spurious claims of voting fraud, his refusal to concede, and popular distrust in the basic mechanisms of American democracy.

Will there be any repercussions for the polling industry that claimed so much collective national attention during what felt like an interminable election season? This seems unlikely. Polls are a fixture of our politics and an obsession of our reportage. That said, they have never been an uncontested feature of U.S. political culture. Their recent, public failures expose a simmering unease with the power of quantitative modeling to define the social world.

Indeed, a similar disquiet ought to pervade the whole of our big-data-infused, algorithmically shaped present...