Don't count Donald Trump out. His odds have improved substantially since bottoming after he blew the first debate and contracted COVID-19. That upward momentum is still in place, and if it continues at the same pace, the race could be too close to call by election eve.

Those are the findings of Thomas Miller, a Northwestern University data scientist who created a new methodology tailored to predicting the outcome of this presidential election. "Trump's picked up since he got back on the trail doing his thing," Miller told Fortune. "This is a very competitive race that's tightening, but the media hasn't picked up on it. You still hear lots of talk on TV about a landslide, with Biden taking Texas. That's unlikely."

Miller, director of the Northwestern data science program, created similar computer models to forecast the outcome of professional baseball games based on simulating individual matchups between batters and pitchers—every single at-bat is modeled, he says. "I play every game in its entirety before it happens, see who wins, and that's how I establish the odds. Then I match those odds with those on the betting sites, and if there’s a major divergence, that's the time to bet. I found that out of 15 games, maybe one or two are worth betting on."

In August he started crunching data on the election. But instead of using the two most common tools—weighing the polling numbers or using such historical measures as how the economy is likely to shape the outcome—he based his predictions on the odds posted on America's most prominent political gaming site, PredictIt. For Miller, the "prices" that political gamblers place on the candidates are a far more reliable guide than the preferences that respondents state in pollsters’ telephone surveys.

"These betting markets provide the best odds because when people put their own money into something, you can be sure that's what they believe," says Miller. "When Trump's posted price is 40 cents and Biden's is 60 cents, and you think Biden's chances are better than 0.6, you'll bet on Biden. That's the beauty of an efficient political marketplace." For Miller, the political futures sites best project "the wisdom of the crowd."

Miller notes that the surveys ask people whom they'll vote for, meaning whom they'd like to win: “PredictIt effectively asks, ‘Whom do you believe will win?’ That's the better question." He adds that polling results reflect opinions collected days before, and not the gaffe the candidate committed yesterday, or the big shift in support in the hours after a good or bad debate. "My results are out five days before the polls come out," he adds. "I was five days in advance knowing the first presidential debate was won by Biden." By the time they appear, polling results are already out of date.

Miller relies on PredictIt because the site takes bets on 56 separate jurisdictions awarded electoral votes: the 48 "winner-take-all" states, and seven districts in Nebraska and Maine, as well as the District of Columbia. Hence, he's able to deploy real-time, minute-by-minute data from what are effectively 56 separate political markets. He notes that the PredictIt platform is highly liquid, since it boasts 100,000 bettors, and that it's restricted to U.S. residents, many of whom are knowledgeable about the on-the-ground, bellwether trends in their counties and states.

The methodology, trademarked the Virtual Tout, updates the forecasts every hour on his website. (Or at Twitter at @virtualtout.) Miller distills the shifting odds in PredictIt’s mainly state-by-state markets into two principal data points: first, the number of electoral votes that the model forecasts Biden will win, and second, the former vice president's chance of victory, shown as a percentage. Miller’s site shows a curve representing the trend in electoral votes. The curve rises when the PredictIt odds from the state betting windows, filtered through Miller's methodology, move in Biden's favor. The trajectory descends, as now, when Trump's fortunes across the 56 markets improve.

Superimposed on the graph are important events, many of which cause an immediate spike or cratering in the candidates' electoral count and odds of winning. In less than an hour, the methodology registers Biden's bump from Trump's angry outburst on 60 Minutes or Michelle Obama's heartfelt video message. "Sometimes I haven't been listening to the news, and I see the time series move dramatically, then I turn on the news to see what made it happen," says Miller.

Let's trace the events that in the first phase measured by Miller brought a big uptick in Biden's fortunes. Miller launched the site in early August and found that Biden held an edge of 80% to 85% to Trump's 20% to 15% during the entire month that encompassed both campaigns. By early September, Biden was holding that lead, and the model gave him around 310 electoral votes, 40 more than needed for victory. Starting with the first debate on Sept. 29, things got even worse for the President. "He blew the first debate," says Miller. "A week later, he's diagnosed with COVID-19. I could see Biden's count jump by four votes within an hour of the announcement." That reminded Americans of the plague threatening their health and the economy. On Oct. 6, as Trump languished in Walter Reed hospital, Biden's numbers hit their high-water mark of 326 electoral votes and a 94% probability of taking the White House. A Biden landslide appeared to be building.

Then, the momentum shifted. The start of Trump's comeback coincided with the vice presidential debate on Oct. 7. But in a comical episode that also shows the sensitivity of Miller's model to passing changes in voters' mood, Trump suffered a blip following the Saturday Night Live skit four days later. The SNL parody of the VP debate showed the famous fly sticking to Pence's head, and Biden, who had been watching the face-off with wife Jill, jumping into a time machine to be transformed into the offending insect, recalling the metamorphosis in the movie The Fly. "From around midnight to 5 a.m., Biden gained a few votes," says Miller. "It must have been the SNL skit. Six million people viewed it on YouTube. What else would move the needle in the middle of the night?"

The fly didn't sting for long. On Oct. 12, Trump returned to barnstorming, hosting a signature raucous rally in Stanford, Fla. "The downward slope for Biden began," says Miller. "Trump's back on the trail. He's recovered and acts, and looks, reborn." The dueling town halls on Oct. 5 gave Biden a brief lift that quickly faded. "The second debate was regarded as a draw, but Trump's momentum continued," says Miller. Then Trump suffered what looked like a serious setback from his obstreperous performance in his 60 Minutes interview with Lesley Stahl. "Women viewers perceived that he was insulting a respected female journalist," says Miller. "It's as if he was saying, 'I don't care about you, you're a woman.'" Biden gained around five electoral votes and benefited for two days. Then Trump's gains got right back on the track they'd jumped after the Stahl debacle.

Miller says that when the Trump resurgence started, he expected it to be short-lived and thought that Biden would resume the 80%-plus odds and over 310 electoral votes he'd enjoyed from August through September. On Oct. 28, Trump's chances improved as Biden’s slipped to 76.4% and 242 electoral votes. By mid-afternoon the next day, Biden had gained a bit to 78.7%. Miller still puts Biden's odds of victory at 3-to-1, down from 4-to-1 a few weeks ago. "I would say the Democrats have the lead, but the Republican ticket is catching up, and this is by no means a landslide," he says. "Trump can absolutely win, and everyone should be aware of that."