“My data show that both races are leaning Democrat, with Warnock holding the wider advantage, but that both are still close,” Miller says.

Miller is a data scientist at Northwestern University who has long believed that the odds posted on political betting sites are far more reliable than projections calculated after canvassing voters. Miller mined the data on PredictIt.org, America’s only major venue for political gaming, to get close to forecasting Biden’s relatively modest 74-electoral-vote margin—in contrast to the pollsters’ state-by-state averages displayed on RealClearPolitics.com at the time, which forecast a Biden blowout.

Miller deploys a highly original survey that asks not for whom respondents plan to cast their ballots, but the candidate they expect to win––and rewards them with a few extra dollars if they’re right. That enables him to escape the biases that limit the reliability of both conventional polls and, to a lesser extent, betting odds. “The problem with relying on one or the other is that typical polls have a slight Democratic bias, because they tend to over-poll Democrats and don’t fully adjust for the tilt, and gaming odds have a Republican bias, because the bettors are typically males who often wager on sports, are more highly educated, and have higher-than-average incomes that give them the discretionary money to gamble with,” Miller told Fortune.

Miller deploys three different methodologies, but considers his modified poll the most accurate. The first data set is what he calls the Prediction Market, reflecting numbers Miller crunches from PredictIt.org; the second is the Preference Survey, in which Miller conducts his own poll that has much in common with traditional polls. And the third is the Prediction Survey, which offers a financial reward for making the right prediction to the same same people polled in in the conventional Preference Survey. In other words, the Preference Survey asks people whom they will vote for and want to win, and the Prediction Survey asks whom they are willing to bet will win.

He looks at all three measures, but considers the Prediction Survey the one most likely to nail the final result. He translates all three measures separately into the margin of victory, expressed as the percentage of probable votes for each candidate. (See Miller’s data for all three categories for the Georgia runoffs at https://www.data-science-quarterly.com/.)

Sharp declines for the incumbents

The basis for the Prediction Market numbers are the adjusted betting odds that Miller derives from PredictIt.org. Miller doesn’t rely on the headline number for the head-to-head between Warnock and Loeffler and Ossoff and Perdue. Instead, he uses the quotes for 14 options on PredictIt.org that invite gamblers to bet on the margin of victory in each race, amounting to 28 sets of odds across the two elections. Until Dec. 30, Miller expressed his Prediction Market figures as each candidate’s share of the 100% total. Starting on Dec. 30, he’s been translating those betting odds into the leading candidate’s margin for victory, the same measure he’s using for his Preference and Prediction surveys.

In both races, Miller’s odds show substantial, stable leads for both Perdue and Loeffler until days after Christmas, followed by sharp declines that have made them both underdogs. The two falloffs likely rank among the fastest in any brief period in the annals of major senatorial races.

Let’s start with Perdue, now running for his second term. In early November, the incumbent was approximately a 75% to 25% or 3-to-1 favorite to win. As the absentee voting that began on Nov. 18 boosted the Democrats’ prospects, Perdue’s edge slipped to around 65% to 35%, until the period from Dec. 17–27, when his lead waxed to 70%. The big drop started on Dec. 27, when he stood at 67%. By Jan. 4, Perdue’s odds had dropped in a single week by almost 20 points, to just under 50%.

Loeffler followed a similar trajectory, suffering an equally big fall from a less commanding lead. Around Nov. 11, Miller had Loeffler as a 67% to 33% or 2-to-1 favorite. Warnock bounced back to around 40% from the start to the middle of December, but was still in about the same 60–40 hole on Dec. 27. Since then, Loeffler’s chance of winning evaporated from 60.4% to the mid-40s. Her drop in points closely parallels Perdue’s decline, but puts her in a tougher position since her post-Christmas lead was smaller.

As of Jan. 4, Miller reckons that his adjusted betting odds, translated into probable votes, make the Perdue-Ossoff race rank as a 50% to 50% toss-up, while Loeffler trails by only around 0.4 percentage points, or something like 50.2% to 49.8%. But keep in mind, the betting numbers tend to favor Republicans. That’s the reason Miller seeks to erase that bias by taking into account the other two measures. “You can’t just look at the prediction markets,” he says.

Miller’s Prediction and Preference Surveys are conducted in partnership with polling outfits Isometric Solutions, Panel Consulting, and Prodege LLC. The two surveys operate differently. The Preference Survey mirrors a typical political poll, with Miller as the pollster. It asks whether the respondents plan to vote, and whom they plan to vote for. It has the same slight Democratic bias as do the great majority of polls.

By contrast, the Prediction Survey asks the same people not whom they will vote for but to act as bettors and predict the outcome. “We’re not asking what candidate they really like, but the one they think will win,” says Miller. If they predict correctly, they’ll get money. Folks are paid $1.50 to participate, and an additional $1.50 for each correct call. “For both surveys, we poll only people who’ve lived in Georgia a year or more,” he says. “They have some understanding of what’s going to happen in Georgia.”

Most of the respondents plan to vote. But only the people who say they will vote are included in the Preference Survey. In the Prediction Survey, everyone is included, because nonvoting residents also have a strong sense of the trends taking shape in their communities and cities. The poll also gauges the margin of victory: The respondents say whether they would bet or pass on all of the 28 point-spread prices offered on PredictIt.org. “They’re laid out just as on the PredictIt.org site,” says Miller.

Raising the stakes

For both races, Miller’s Preference Survey––the conventional poll without the financial rewards––shows the Democrats winning big, although he notes the numbers have been jumping around. “In the Jan. 2 and Jan. 3 polls, our respondents have both Ossoff and Warnock winning by over five points, based on the way those respondents plan to vote,” says Miller. By conventional poll standards, that would amount to a decisive victory for both Democrats.

But the Prediction Survey has the advantage of landing between the two extremes. For Miller, it gives the best estimate of the three. On Jan. 2, Ossoff’s Prediction Survey lead was 0.4 percentage points, rising to 0.86 points the following day. “That’s still close to a toss-up, with a slight edge to Ossoff,” says Miller.

The Loeffler-Warnock race is a different story. “The race was dead even until the end of December, according to the Prediction Survey,” says Miller. But by Jan. 3, Warnock had established a lead of 1.32 points. “It’s still close,” notes Miller. “But we think Warnock will win, though by nothing like what our poll, and most polls, are showing.” He has “a lot more trust” that Warnock will prevail than he does that Ossoff will.

Winning both means the Senate goes Democratic, giving the Democrats control of both houses of Congress and potentially granting the Biden administration far more power to enact its agenda. That outcome is closer to a toss-up than either of the two races individually, says Miller. “Both of these races are competitive, much more so than the polls are showing,” he notes.

In a sense, Miller developed this methodology in response to his biggest, and virtually sole, mistake in calling the presidential election. “The betting odds I was relying on were consistently calling Georgia for the Republicans,” he says. “North Carolina was more of a toss-up than Georgia! We missed it completely.” Miller thinks that if he’d run the Prediction Survey then, he would have seen that Georgia was actually a toss-up, based on asking folks who watch the local news, talk to their neighbors, and see big shifts, just as they start developing, regarding whom they think will win and rewarding them if they are right. “I saw a Republican state. If I’d been talking to people who knew what was happening I would have been in a much better position to predict Georgia,” he concludes.

For this writer, Miller’s methodology provides the best real-time view of one of the most dramatic political cliff-hangers in recent history.