A prediction model built by Thomas Miller, a data scientist and professor at Northwestern University, continues to project that Democrats will flip five seats in the Senate elections, even as Republican chances have risen modestly in recent days.

The model powers Virtualtout.io, a predictions site whose record has drawn attention before. In January of 2021, Miller's approach posited—contrary to the polling at the time—that Democrats would win both of the two Senate races in Georgia. It scored again by calling a Trump triumph in 2024.

For the current Senate contests, Miller's template once again starts with the “prices” posted on betting sites. He relies on the two platforms that host the most trading, Kalshi and Polymarket—prediction markets where traders buy and sell contracts tied to election outcomes—and calibrates the prices from both sources so that they reflect percentage probabilities of victory or defeat in each of the 35 races.

He then applies multiple time series analysis that adjusts for “underlying factors” that tend to drive all the election “markets.” The overriding force by far, he says, is whatever Donald Trump has just done or said.

“The states don't move independently of one another,” Miller told Fortune. “If Trump does something voters don't like, you'll see the Republican numbers drop in all the states.”

On Virtualtout.io, a live “ticker tape” for all 35 races runs across the top of the screen. It shows the odds of a Democratic or Republican win, updated every 10 minutes, along with the change over the past 24 hours. The data is naturally far more current than polling, which often reflects voters' preferences from several days or more before it is posted.

Miller's numbers are a snapshot of today's probabilities, not a prediction of what will happen on Nov. 3. But they do show Democrats in a strong position, as of now, to flip seats in five states—an outcome that would radically change the count from today's 53-47 in favor of the GOP to 52-48 for the Democrats—and, with it, majority control of the chamber.

Virtualtout went live on Oct. 23, but Miller has collected numbers starting from Sept. 17. On that date, his models showed a 34.8% chance that Republicans would keep control. As of early morning on Oct. 26, that figure had dropped to 23.3%.

The Republican side, however, is doing a bit better; its odds have risen two points from an “all-time” low of 21.3% on Oct. 24. The interplay among the races is also notable, with three states showing somewhat better chances for Republicans than previous numbers: Texas, Maine, and Michigan—a seat the Republicans had hoped to flip, but which still favors Democratic candidate Abdul El-Sayed.

Starting where GOP chances look remote and haven't brightened: North Carolina is effectively out of reach, with Miller's model giving Republicans a 4.5% chance of winning. Ohio was a near toss-up six weeks ago; now incumbent Jon Husted has only a 32.5% shot versus Sherrod Brown. Alaska is also looking dim—the GOP's chances opened Miller's derby at 33% and now stand at 24.8%.

Texas also started as near-even before Republican odds began to fall, but in recent days they've improved, rising from 34.2% on Oct. 24 to 36.4% on Oct. 26. The outlook in Maine has improved a bit as well for the Republicans, going from 32% for Republican Susan Collins on Sept. 17 to today's 37.9%. And in Michigan, the GOP's odds dropped to a low of 26.4% on Oct. 5, but have since risen to nearly 34%.

Miller's numbers are highly fluid, and Republicans in his model still harbor roughly one-in-three or better chances of winning in four states. It is by no means a done deal for the Democrats. The slightly higher numbers for Republicans could mark a trend—an election that looked lacking in suspense could show plenty in the closing days before Election Day on Nov. 3.

This story was originally featured on Fortune.com.