In late 2020 and early 2021, this reporter wrote several stories focusing on the election predictions advanced by Thomas Miller, a data scientist at Northwestern University. I was intrigued by the highly original methodology Miller deployed in calling the trends, and outcomes, first in the presidential race, then for the two Georgia senatorial contests, where the surprise twin victories gave Democrats control of the upper chamber.
Now, Miller is back in the arena for the first time since offering those highly unconventional forecasts. He’s using a similar system to handicap the supposedly super-tight presidential election that will be decided just 48 days hence. His wildly out-of-the-mainstream call is sure to shock pollsters, pundits and Fortune‘s readers alike. But Miller’s view merits close attention for two basic reasons: First, it’s based on numbers-crunching that’s arguably a lot more scientific than the voter surveys almost always cited to chart the contest’s trajectory, and second, he achieved pinpoint accuracy four years ago.
In all three 2020 contests, Miller beat virtually every pollster, and modeler parsing multiple voter surveys. He missed the size of Biden’s win in the electoral college by just 12 votes, tagging every state for the correct column save Georgia. For the two senate runoffs, Miller refined his approach to sorting data on the Peach State, and scored again. A week before Election Day on December 6, 2020, the polls gave Republican David Perdue a wide lead over Democrat Jon Ossoff, and showed the GOP’s Kelly Loeffler in a dead heat versus opponent Raphael Warnock. By contrast, Miller’s numbers had Loeffler heading for a big loss, and Ossoff en route to a modest victory. Once again, the contrarian academic nailed it: Miller was just 0.2% short on Warnock’s 2.0% margin, and precisely on target in forecasting Ossoff’s 1.0% final bulge at the ballot box.
Harris vs. Trump in the polls and prediction markets
Miller’s approach vastly differs from the most of political prognostications by relying not on polls, but the prices established by Americans wagering their own dollars on the candidates they reckon are most likely to prevail. “Political betting sites are the best at predicting the wisdom of the crowd,” he told Fortune. He states that while polls tell you about the past, the odds on the betting sites map the future. “Polls are a snapshot of the recent past,” he adds. “They typically canvas small samples of 500 to 1,500 people. And the pollsters are asking respondents whom they’re planning to vote for at that moment, which may have changed a few days later when the results are posted. Most polls are around four or five days behind.”
That lag, he says, causes a lot of noise or variability that disguises the actual picture. Another problem with polls: Miller reckons that a better question than “Tell us which candidate you’ll vote for?” is “Which candidate do you expect to win?” And while the pollsters don’t pose that query, it’s just how the bettors are making a market.
For the 2020 Biden-Trump face-off, Miller deployed the pricing posted on the largest U.S. political betting site, PredictIt. He took the PredictIt odds in the 56 individual voting jurisdictions, tracked the hourly changes, and used his proprietary model to roll the data into daily odds that were extremely current given that folks were posting bets for one candidate or the other 24-7 on the site. For the Senate races, Miller took a different tack. He assembled a group of about 1,200 Georgians whom he lured by agreeing to pay them a few dollars to participate, and extra dollars if they named the contender most likely to win—not the necessarily one they planned to vote for, as well as predict the margin for victory. The method he developed, called a “prediction survey” taking the best parts of the polling and the betting market guided Miller to a near-perfect reading of the voting shares.
For the 2024 presidential race, Miller is once again relying on PredictIt. He praises the site for “having a more stable group of investors” than the populations screened by pollsters. “You have tens of thousands of people betting on the site at all times of the day,” he says. “The maximum contract is $850, and 37,000 ‘shares’ are traded on average every day.” Put simply, Miller regards PredictIt as a highly liquid market resembling an exchange for, say small-cap stocks or high yield bonds that gathers huge numbers of buyers and sellers. “Financial markets are forward-looking, and incorporate information instantaneously, and PredictIt provides the same benefits,” says Miller.
Is Harris ahead of Trump? This model says yes
For gauging Biden-Trump rematch, Miller developed a “generalized, linear model” based on results from the most recent sixteen presidential elections starting in 1960. Starting with that year’s Kennedy-Nixon race, every presidential race has been decided by 538 electoral votes. “The model’s based on a lot of historical data,” says Miller. It shows that the daily pricing on PredicitIt translates closely into the share of the popular vote favoring each candidate. Say on a certain day, bettors give candidate A a quote of 52 cents, amounting to 52% chance of winning. By Miller’s reckoning, those 52% odds should also mean that at that moment, the best forecast holds that 52% of likely voters plan to cast their ballots for candidate A.
Each day, Miller recorded the prices posted at midnight on a graph where he overlayed the crucial events shaping the election. Titled, “Effects of Current and Campaign Events on Forecasted Electoral College Votes,” it’s posted on his website, The Virtual Tout (virtualtout.com). Overall, the chart tracks the number of electoral votes the PredictIt odds are showing. It’s shaped like a giant “U,” resembling a steep ski slope that bottoms then rises into a giant mountain.
Well before the Biden-Trump debate, the Republican was far ahead; on Miller’s chart, he stood on top of the ski slope. Trump looked to command almost 400 of the 538 electoral votes in mid-June. After the onstage match on June 27, Trump got even more dominant, and he peaked in the days framing and during the Republican convention at a share of almost 500. The Democrats’ odds rose sharply after Biden withdrew on July 21, and by the time Harris secured the nomination on August 6, the Democrats were solidly in the lead at around 325 electoral votes. They kept gaining over the following two weeks, hitting 400 in a stunning reversal by the time their convention ended on August 22.
Then, Trump staged a comeback. In the days before the September 10 Trump-Harris debate, Harris was still ahead, but Trump had nearly caught up. “At that point, the race was essentially a tossup,” observes Miller. “The forecast for the Democrats was 288.” It was the onstage battle in Philadelphia that wrecked the 78-year old former POTUS, according to the Miller numbers. Within a day after the candidates left the podium, Harris had jumped to exactly over 400 electoral votes. The Harris endorsement from Taylor Swift, secured the day of the debate, probably helped sink Trump’s chances, according to Miller. Since then, Harris has maintained for 400-plus vote total.
A second chart on The Virtual Tout shows how the share of the popular vote influences the electoral tally over those 16 elections. Keep in mind that Miller is using the PredictIt prices as equivalent to the popular vote: A candidate whose price is 51% on Election Day is likely gets 51% of all ballots cast. He points out that Democrats traditionally must secure an advantage of more than two points or more to win, and that Republicans can prevail with as little was roughly 48%.
As of September 16, PredictIt is showing a price of 55 cents for Harris, and 45 cents for Trump, the reverse of the scenario before Biden’s departure. Once again, those odds translate in 55% of the popular vote for the Democrat according to Miller’s model. If the situation persists, Trump faces an absolute rout. “It would be somewhere between the defeats of Barry Goldwater by Lyndon Johnson in 1964, and Bob Dole by Bill Clinton in 1996,” says Miller. “We’re talking about a blowout where Harris gets over 400 electoral votes and wins Michigan, Pennsylvania, North Carolina and every other swing state.”
Miller notes that at least in recent history, America’s never witnessed a reversal of fortune remotely as dramatic as this one. “It’s gone from a drastic landslide in Trump’s direction to a drastic landslide for Harris,” he marvels. The distance is now so great that only another epic swing would bring Trump back into contention, and Miller predicts that right now, it looks like Harris will win big on November 5. As a coda, he recalls a slogan the Johnson campaign used to bash Goldwater: “In Your Guts You Know He’s Nuts.” Miller’s markets-based analysis posits that the people betting their own money are right in predicting that by the time the candidates left the stage on September 10, millions of voters likely to back Donald Trump abandoned the ex-President, starting the shock waves that could cause an avalanche for Harris that as of now, few see building.
