Miller stresses that his approach is based on the results from over 60 years of presidential elections.

I became convinced that Miller’s got a better system while following his framework and predictions during the 2020 election cycle. He called the presidential race for Biden within 12 electoral votes, and in the Georgia senatorial runoffs that followed, posited that Democrats Raphael Warnock and Jon Ossoff would both win when the polls still showed them trailing their GOP opponents by substantial margins.

On the Harris-Trump race, I’ve posted three stories between Sept. 18 and Oct. 3, and in all of them, Miller’s methodology showed Kamala Harris holding a big lead, while the polls displayed a dead heat, and the media almost invariably characterized the contest as too close to call. In fact, the earliest piece stated that, by Miller’s metrics, the vice president could be headed for a landslide approaching President Lyndon B. Johnson’s blowout victory in 1964. Even in the Oct. 3 update, Harris—as of the previous day—was leading the former president by a big bulge of 66 electoral votes, 302 to 236.

In the past week, the contest has taken yet another abrupt turn—this time in Trump’s favor

In a battle that already charts like the craziest of rides on the waviest of roller coasters, Donald Trump over the week through midday on Oct. 9 has staged a sharp comeback. To Miller, that’s proved something of a surprise, since in the 20 days or so following the presidential debate, the race had seemingly entered a steady state where Harris consistently held the lead. But as of midnight on Oct. 8, the Miller system gives a tiny edge to Trump: 270, versus 268 for Harris. “This is the first time in history that a race has gone from a toss-up to a landslide for one party; then to a toss-up then to [a] near landslide for the other party, then back to a toss-up,” Miller told Fortune. “The dynamic of this election is that after one candidate jumps way out in front, the race always works its way back to even. That’s where it stands now. I follow this minute by minute, and the results keep toggling back and forth around the 270 electoral votes needed to win.”

It’s important to briefly review the Miller model that worked so well in 2020. It’s governed by two ruling principles. The first relies on the probabilities expressed by bettors who’ll reap big gains if the candidate they’re wagering will win, not the one they plan to vote for, prevails on Nov. 5. To get those odds, Miller uses the prices posted on the most established political betting site, PredictIt, which on average trades a gigantic 35,000 shares a day, bought or sold on the two candidates. Miller eliminates the votes for third party contenders, and makes the Trump and Harris prices equivalent to 100% of the total.

He then deploys his methodology to translate the probabilities into popular vote shares for each candidate. That task requires a number of adjustments. For example, a 50% share on PredictIt for Trump equates to less than 50% of the likely popular vote for the GOP ticket, in part because the wagerers are mainly men, and often regular sports gamblers, who lean Republican. On the other hand, a Republican ticket can win at well under half the popular vote, another factor Miller has calibrated in his number crunching.

The second ruling guideline: The projected popular vote percentages closely track the share of electoral votes each candidate receives. That’s the data scientist’s takeaway from studying every election since 1960. In the past few days, the PredictIt odds have narrowed from favoring Harris by several points to virtually even, accounting for the electoral lead that’s now seesawing from one candidate to the other, within a narrow band of a few electoral votes (EVs).

Following the careening course is an exercise in electoral whiplash

On his homepage, the Virtual Tout, Miller posts a graph showing the Democrat EV share at the close of each day, overlaid by big events that appear to have moved voters. The graph resembles a giant “S” laid on its side. In early June, a Biden-Trump face-off looked like an even match. That was the first “toss-up” moment. Following their debate on June 27, Trump took a huge lead, and by the time of the Republican convention in mid-July, Trump had a commanding total of well over 300 EVs. That was the “first landslide” phase, and the only one for Trump. The Republican standard-bearer maintained a double-digit EV lead until his disastrous appearance before the National Association of Black Journalists on July 31, where he falsely claimed that Harris had misled voters about her race.

That day, the Democrats took the lead for the first time. But Trump bounced back fast, almost leveling the count by Aug. 5. That’s the “second toss-up” point. The next day, Harris secured the nomination and named Tim Walz as her running mate. Biden’s withdrawal gave the Dems a big boost, and sent Harris climbing; she reached a summit of just over 350 electoral votes on Aug. 11. That milestone marked the apex of the “second landslide” span, this time tagging Harris as the big victor just a month after Trump looked like an easy winner.

Once again, Trump closed the gap, getting within seven EVs of Harris on Sept. 9, the day before the debate. That’s the “third toss-up” point. Famously, Trump’s disappointing performance onstage in Philadelphia sent loads of EVs into the Harris column. This time, the uptrend peaked on Sept. 20 at 337 Harris votes to Trump’s 201. The chart had careened to the “third landslide” zone, the second time in six weeks that Harris looked headed for a walkaway.

In late September, the gulf gradually narrowed. Miller believes that JD Vance did a bit better than Walz in their Oct. 1 debate. But Vance’s good performance, he says, did virtually nothing to improve Trump’s chances. Two days later, Harris still held a significant edge of 58 EVs at 298 to 240. Trump’s big surge started on Oct. 4. In the four days that followed, he vaulted to 270 and into the tiniest of leads, while Harris dropped by 30 EVs to 268.

Miller sees a deadlock until Nov. 5, not a continued movement toward Trump

I ask Miller if he sees a wave building for Trump, given the momentum he’s shown in the past few days. “I don’t think so,” he replies. “Trump’s still playing to his base. I haven’t seen him changing his pitch to reach out to more voters like a normal politician.” So what accounts for the sudden jump in GOP numbers? “It may be because of fears of widespread warfare in the Middle East, and that American troops will join the conflict,” he notes. “The Biden regime is pursuing a traditional foreign policy of supporting allies militarily, leading to the view that could happen. Trump is more of an isolationist, and that appeals to a lot of people who think of Afghanistan and Iraq and say, ‘Why are we trying to do this? Why get involved in foreign wars?’ A lot of people across the spectrum are not interested in our going to war.”

The pattern of huge surges that keep deflating, bringing the race back again and again into even balance, symbolizes America’s current predicament. “We’re an incredibly divided nation,” says Miller. “The most likely trend over the next 26 days to Nov. 5 is that the forecast keeps going back to a toss-up. The next month will be a crazy time.”