Welcome to an Unprecedented Election Day Experiment

Want to see who’s winning the presidential race? Follow real-time turnout projections from Slate and VoteCastr.

Several people cast their ballots on electronic voting machines on the first day of early voting at the Provo Recreation Center, on October 25, 2016 in Provo, Utah.
The sun rises near the White House on November 8, 2016.

Photo by Zach Gibson/Getty Images

Be sure to visit slate.com/votecastr for all VoteCastr estimates and analysis.

This Election Day will be different—regardless of how it ends. This time, for the first time, you won’t have to wait until the polls close to find out what happened while they were open. In partnership with the data startup VoteCastr, Slate will be publishing real-time projections of which candidate is winning at any given moment of the day in seven battleground states, any of which could decide who is the next president of the United States.

This, as you may have heard, is controversial. It will break a decadeslong journalistic tradition whereby media outlets obey a self-imposed embargo on voting information under the unproven theory that it might depress turnout on Election Day. But as our Editor-in-Chief Julia Turner put it this summer when she announced the VoteCastr partnership: “The role of journalists is to bring information to people, not to protect them from it.” For the first time, you’ll have access to the same kind of data that campaigns use to monitor voting activity and frame their thinking throughout Election Day. We teamed up with VoteCastr because we don’t think there’s any good reason the candidates and their teams should have a monopoly on that kind of information.

Why do we trust the VoteCastr team? Its staff is stocked with leading data experts from both sides of the aisle. Those staffers include Blaise Hazelwood, who served as the political director for the Republican National Committee and managed Election Day reporting for George W. Bush in 2004, and Ken Strasma, the chief of microtargeting for Barack Obama in 2008.

Here’s how the VoteCastr system operates. By combining proprietary, large-sample polls taken prior to Election Day with targeted, real-time tracking of voter turnout on Tuesday, VoteCastr will make rolling projections of how many ballots have been cast for each candidate in each of the states we’re tracking: Florida, Iowa, New Hampshire, Nevada, Ohio, Pennsylvania, and Wisconsin. If you visit Slate at 11 a.m. EST on Tuesday, you’ll see projections for how many votes have been cast for Hillary Clinton and Donald Trump in each of those states as of 11 a.m. (VoteCastr will also analyze the vote in Colorado, albeit using a different technique. More on that in a bit.)

It’s crucial to remember these projections are being made in real time. Even if we were to assume the VoteCastr models are perfect—and we won’t—they can’t tell us who will win a particular state, only who is winning that state at a specific moment in time and who might win if current trends continue. When it comes to who might win, the emphasis should be on might. There are too many unknowns for us to be able say with confidence that what we think is happening in the present will continue to happen in the future. It’s entirely possible, for instance, that Trump voters will be more likely to cast their ballots in the morning and that Clinton voters will be more likely to cast theirs in the evening—or vice versa. There just isn’t enough historical data to give us meaningful insight on that type of voter behavior. Over the course of the day, we expect Clinton’s and Trump’s respective shares of the total vote in each state to shift as turnout waxes in some areas and wanes in others.

Slate readers will be able to watch live as those vote totals update throughout Election Day. They’ll also be able to sort the data in a number of different ways. We’ll make it possible, for instance, to compare real-time turnout in Trump-leaning counties and Clinton-leaning counties, as well as to gauge turnout in counties grouped by age, income, and race. If our real-time trackers are seeing that turnout in Pennsylvania’s middle class or predominantly black counties has surpassed 2012 levels, you’ll know that. If turnout in Ohio counties that are predominantly white or lower class does the same, you’ll know that, too.

We also hope to use the VoteCastr model to bring some empiricism to all the anecdotes that pop up in the news on Election Day. If there’s rain in Cincinnati, a viral photo of long lines in Las Vegas, or an unplanned appearance by Tim Kaine in Philadelphia, we won’t have to speculate about whether those events will cause turnout to rise or fall. We’ll be able to look at the numbers and draw conclusions—albeit tentative ones—ourselves.

* * *

Here’s a longer description of how this whole thing works. The project can be broken down into two phases: what happens before Tuesday, and what happens on the day itself. In the lead-up to Election Day, VoteCastr conducted large-sample surveys in eight battleground states. Unlike a typical media poll that might ask hundreds of respondents dozens of questions, these surveys presented thousands of people with just a handful of queries each. The results were then run through predictive models to determine the probabilities of each voter in each of the eight states casting a ballot for Clinton, Trump, Gary Johnson, or Jill Stein. (VoteCastr did not include Evan McMullin in its models. The independent candidate is only on the ballot in two states we are tracking, Colorado and Iowa.)

The other piece of the pre–Election Day puzzle is early voting, which now accounts for an estimated 30 to 40 percent of the general election vote. Local officials collect and report information about who voted early in each state in advance of the election, and VoteCastr then compares that public info with its own private voter files. To understand how this works in practice, consider my early ballot, which I cast in Iowa City last week. Though VoteCastr doesn’t know who I voted for, it can make an educated guess based on the things it does know about me: my age, race, and party registration. Our friends at VoteCastr tell me the model believes there’s a 97 percent chance I voted for Clinton. When my name shows up on the list of people who voted early in the Hawkeye State, VoteCastr will use that number to fill in the blank. These voter preference estimates allow VoteCastr to make more specific forecasts about the early voting split than most other modelers, which simply sort returned ballots by party registration. (For what it’s worth, the model got it right in my case: I voted for Clinton.)

That’s the easy part. If everyone voted before Election Day, the final outcome would be pretty easy to predict even without a fancy model. The challenge for VoteCastr and other prognosticators is to figure out which voters will make the trip to their local polling stations and which will stay home. That’s where the day-of tracking comes in. VoteCastr will have hundreds of field workers stationed at preselected precincts around the country. Those field workers will be reporting official turnout numbers as they’re provided to them by poll workers throughout the day. By selecting a representative mix of precincts, VoteCastr will extrapolate the turnout in similar precincts that aren’t being tracked, in the same way it used large-sample polling to draw probabilistic conclusions about how I was going to cast my vote without surveying me directly.

Let’s assume there’s a particular precinct in Wisconsin in which pre–Election Day polling suggests voter preference for Clinton and Trump is split 50-50. If a field worker stationed there reports that 100 votes were cast in the first hour of voting, VoteCastr won’t simply assume that 50 of those votes were for Trump and 50 were for Clinton. The model will also factor in how likely it believes Clinton supporters in that precinct are to vote compared to Trump supporters. Let’s consider a simple hypothetical in which each of Trump’s likely voters in our Wisconsin precinct is more likely to vote than each of Clinton’s likely voters. If projected turnout is low, then we can assume the more-energized Trump supporters will vote in greater numbers than Clinton supporters. If turnout is high, then we can assume there will be more parity—that the high turnout is an indication that less-energized likely Clinton voters did show up to vote on Election Day.

VoteCastr’s projections will work a little differently in Colorado, where the vast majority of ballots are cast by mail ahead of Tuesday. On account of that, VoteCastr won’t be tracking real-time turnout in Colorado. Instead, it will rely on the same technique it used to sort the early vote in other states, making projections by comparing its private voter files—which include the poll-derived preference scores for each voter—with the public roll of who returned their ballots. That means the VoteCastr projections for Colorado won’t change throughout Election Day like they will for other states; the numbers we have in the morning will be the same ones we have all day.

* * *

That’s how the VoteCastr system will work in theory. It’s possible, however, that reality will introduce a few surprises. For starters, projections are only as good as the models that make them, which are only as good as the polling data they use. If there’s some sort of shy-voter effect, or respondents are being in any other way dishonest with the pollsters (or themselves) about who they are going to vote for, the projections could easily miss the mark. Likewise, there’s the possibility of a late swing in the race that VoteCastr polling misses. As with all polling, some uncertainty is unavoidable. (VoteCastr tells me its own polls have generally been in line with publicly available polls.)

There’s also the not-so-small issue of Clinton’s and Trump’s respective ground games, or lack thereof. The VoteCastr model relies heavily on 2012 voting to predict the likelihood of whether someone will vote in 2016. The consensus four years ago was that Obama’s get-out-the-vote operation was better than Mitt Romney’s, but only by a relatively slim margin. There’s good reason to believe that Clinton’s advantage over Trump will be considerably larger than that. If that turns out to be true, Clinton supporters may be more likely to vote than comparable Trump supporters simply because they’re more likely to get knocks on their doors on Election Day from campaign volunteers. The VoteCastr model can’t account for that.

Setting aside the model, there’s one more problem to consider: me. On Election Day, I will be doing my best to provide the context and analysis readers need to understand the VoteCastr data. But I’m only human, and I’m coming in with my own preconceptions. Based on what I’ve observed during the final months of this campaign, I believe Clinton is the favorite to win the general election. And so there is always a chance that I will (perhaps unconsciously) seek out patterns that confirm that prediction, or I will fail to spot evidence suggesting the opposite outcome. Conversely, now that I’ve acknowledged this bias, it’s possible I’ll overcompensate and swing too far the other way.

Ultimately, it’s best to think of this Slate-VoteCastr collaboration as a real-time Election Day experiment. The VoteCastr projections are based on vast amounts of data, but this specific model has never been tested on this scale. If it works as planned, it’ll provide us with fascinating insights into what’s happening on the ground on the final day of one of the most unpredictable campaigns in recent memory. It can also help us answer long-standing electoral mysteries, like whether certain kinds of voters go to the polls at different times of the day. And if all else fails, we’ll still be giving you a taste of how campaign insiders see Election Day and how they see you as voters. We think that’s a valuable exercise, and we’re excited to have you along for the ride.

Read VoteCastr’s official methodology documentation here.