Biden AdministrationUkraine InvasionInflationEnergyCybersecurity

Georgia Senate races: Data scientist’s final prediction has Democrats taking both seats

January 5, 2021, 9:02 PM UTC

Given that most polls got the presidential election so wrong, folks are looking for a more reliable, scientific forecast for the crucial senatorial runoffs in Georgia. Thomas Miller, a data scientist at Northwestern University, is handicapping the races, marshaling what he considers the most sophisticated, data-rich, bias-free model in the business.

According to Miller’s analytics, online betting sites underestimate the true strength of the Democrats in the two runoffs, while the polls wrongly predict blowout victories for both. Miller’s methodology finds a middle ground between the two extremes that dominate the headlines and keep heads spinning.

Miller’s bottom line: With polls hours away from closing, both races remain competitive, but Democrats Jon Ossoff and Raphael Warnock are most likely to win—Ossoff by a narrow single point, and Warnock by a more substantial 2% margin. Miller also reckons that the odds on the political betting sites—whose forecasts are a lot more reliable than the polls—are wrong in favoring the Republicans to win at least one race and keep control of the Senate. Instead, Miller reckons that the Democrats enjoy a more than one-in-two chance of sweeping both races, and hence controlling both houses of Congress.

Miller uses three types of forecasts, each of which are proprietary. The first he calls the Preference Survey (conducted in partnership with Isometric Solutions, Panel Consulting, and Prodege LLC). It’s a conventional, daily poll of around 1,200 adult Georgia residents, asking whom they intend to vote for. The second is his Prediction Market data. For this yardstick, Miller uses the prices posted on the sole major political betting site for U.S. gamblers, PredictIt. Miller distills the 14 different point spreads on each race that bettors choose from on PredictIt to get odds that incorporate far more data—and, he claims, are much more reliable—than the win-or-lose, head-to-head prices that PredictIt itself displays on the two contests. (Miller’s data is posted on https://www.data-science-quarterly.com/.)

While Miller says the betting markets are much more accurate than the polls, as they proved in the 2020 presidential race, both have biases. The polls typically lean Democratic, while the gaming odds tilt Republican because the gamblers are typically male, well-educated, and often wager on football or basketball (all demographics that lean toward the GOP).

Hence, Miller also uses a third data set, the Prediction Survey, that combines the best elements of the polls and betting sites. It canvasses the same 1,200 respondents as the Preference Survey, but instead asks people whom they expect to win, not whom they will vote for and hope will prevail. He also adds an incentive that turns his audience into bettors: In addition to the $1.50 they receive each day for participating, they’ll get an extra $1.50 for each of the two races they call correctly. Both surveys poll only people who have been living in Georgia for at least a year. “Those people really understand what’s happening in their home state, so when they predict who’ll win and get paid a few dollars to get it right, their overall forecast is more reliable than either the standard polls or the betting odds separately,” says Miller. For Miller, the Prediction Survey is the new gold standard for handicapping political races.

Ossoff vs. Perdue

Let’s examine where the two races have been trending, and where they stand now, on all three measures. We’ll start with the election pitting Democrat Ossoff versus the Republican, one-term incumbent David Perdue. (All numbers are expressed as the margin of victory if the election were held the day the data is collected.) Just a few days ago, the numbers in the Preference Survey, or regular poll, were careening back and forth, with Ossoff leading by 3.8 points on Dec. 30, and Perdue jumping over four points ahead the next two days. But for the three days from Jan. 2 to the last reading on midnight, Jan. 5, Ossoff enjoyed a huge advantage of well over five points. “In other words, the polls are saying that he’ll win in a virtual landslide,” says Miller.

The Prediction Market odds, the number Miller derives from PredictIt, tell a far different story: Ossoff is leading by a slender 0.1%. “So according to the bettors, the race is still a toss-up,” says Miller. For Miller, it’s the Prediction Survey, where respondents get rewarded for calling the races, that chart the best course between the plain vanilla polls and the betting prices. It gives Ossoff a lead of 1.1 points.

But Miller’s analysis doesn’t stop there; he also adjusts for the most likely turnout, the factor that could swing both elections either way.

“The overwhelmingly top predictor for the candidate people vote for is race and gender,” he says. “That’s based on four categories, nonwhite male and female, and white male and female. The numbers who actually vote in each of those categories is key to calling the results.” Miller plugs the demographic data for the four groups across every one of Georgia’s 159 counties into his formula predicting turnout, based on which groups voted in which numbers in the 2016 and 2018 elections. “So the forecast has four times 159 or 636 components,” he says. Miller’s model foresees an unusually high number of voters from the groups typically favoring Democrats. That factor lifted Ossoff’s edge a bit in Miller’s model to just under 1.5%. Given the continued closeness in the betting odds, Miller rounds down his advantage to 1.0%, meaning the most likely result is an Ossoff victory of 50.5% to Perdue’s 49.5%.

Loeffler vs. Warnock

In the special election where Republican incumbent Kelly Loeffler, who was appointed to her seat last year, faces challenger Warnock, Loeffler actually led in the conventional Preference Survey on Dec. 31 and Jan. 1. Then Warnock suddenly opened a gigantic, enduring lead of over five points that mirrored Ossoff’s sprint. Once again, the bettors were less bullish on the Democrat: The Prediction Market reading that Miller derives from the 28 point spread prices on PredictIt has Warnock ahead by only 0.5%, better than Ossoff’s 0.1% but pointing to a tight race. The Prediction Survey, Miller’s golden mean, has Warnock as a 1.46% favorite. As with the regular race, Miller runs that number through his turnout model incorporating demographic data from the Peach State’s 159 counties. With turnout tilting Democratic, Warnock’s lead advances to around 2.0%.

At Fortune’s request, Miller translated Ossoff’s 1% lead and Warnock’s 2% advantage into the odds each will win. From the 1,200 respondents in the Prediction Survey, the chance that Ossoff will win in the regular election is 54.8%, and from the same survey, Warnock’s probability of victory is 59.6%. The paramount challenge is forecasting whether the Democrats will take both seats, the feat they must achieve to control the Senate. On a single coin toss, the chances of getting heads or tails is one in two, but the chance of getting heads twice in a row is one in four. Since both races are fairly close, you’d think that the Democrats’ chances of a sweep are less than 50-50. Indeed, the bettors on PredictIt are wagering that the Republicans will win at least one seat and maintain the majority.

Their thinking is misguided, says Miller. “The two races aren’t independent, like two coin tosses,” he says. “They’re highly correlated. They move together in lockstep. When Warnock moves up, Ossoff moves up. If Warnock wins, Ossoff is also likely to win. So the odds are 55% that the Democrats take both seats.” That result would move the upper chamber to a 50/50 split that puts the Democrats in control.

As Georgia goes, so goes the nation, and Georgia’s been going Democratic only in the past 10 days. No one’s chronicled the historic shift with greater precision than Tom Miller.