We Built Our Own Election Forecast. Then We Asked the AIs to Call the Race.

Today we're launching the Rational Ground 2026 Forecast for members: every Senate, House and governor race, rerun every day, with the work shown.

Every two years, the same thing happens.

A handful of outlets publish a number. Democrats 71%. Republicans 58%. The number gets quoted on cable, and then it gets quoted as fact. Almost nobody asks how it was made.

Then election night comes, and half the time the number was wrong in a way nobody warned you about.

I don't have a problem with forecasts. I have a problem with forecasts you can't check.

So we built one you can.

What it is

The Rational Ground 2026 Forecast covers all 506 races on the November ballot: 35 Senate seats, 36 governors and all 435 House districts.

Every day, the model takes in the newest polls, campaign money and district maps. Then it plays out the election 20,000 times. If Democrats hold the Senate in 13,000 of those runs, their chance is 65%. That's all a forecast is: a count of how often something happens when you play it out.

Every race shows what produced its number: the polls behind the average and how much each one counts, the money each campaign has raised, every outside group that just started spending, and how the district has voted before.

And the full method is written down on the page, in plain English, with every setting the model uses.

How it works

Five ingredients go in.

The polls. As of today, 866 polls on individual races. Some pollsters lean one way, year after year, so we measure each pollster's lean against everyone else's this year and adjust for it. A poll from a firm that always runs three points hot gets cooled off.

The national mood. There are another 473 "generic ballot" polls asking which party people will vote for in Congress. That's the tide that lifts or sinks every boat. Right now it sits at Democrats +6.8.

Each place's lean. How a state or district voted before: 70% the 2024 presidential race and 15% each the 2018 and 2022 midterms. We didn't pick those weights out of the air. We tested them against what actually happened in 2018 and 2022 and kept the mix that called those years best.

Incumbents. Sitting members get a small bonus, plus part of their own track record. A governor who keeps winning in a state that votes the other way, like Phil Scott in Vermont, has earned that.

Money and maps. We pull campaign money straight from the Federal Election Commission's filings and flag every surge in outside spending, because the big groups often know something before the polls do. And 173 House districts have new lines since 2024. Texas, California, Florida and others redrew mid-decade. We score every seat on its new map, not its old one.

What it says today

As of October 8:

• Senate: Democrats have a 65% chance of control. Republicans 30%. About 5% of the time, neither party gets a majority on its own.

• House: Democrats 91%.

• Governors: Democrats are expected to hold about 26 or 27 of the 50 seats.

I didn’t build this to tell you what you want to hear. I built it to tell you what the numbers say, and to show you exactly why.

The reason is mostly that national mood. A generic ballot of Democrats +6.8 is a strong year for the party out of the White House. If that number moves, everything moves.

The Senate runs through five states. In the races most likely to decide control, the tipping point is Texas, Ohio, Alaska, Michigan or Maine. Three more are close to a coin flip: Kansas, Iowa and Nebraska.

These numbers will change. That's the point. The page updates every day, and members can watch it move.

The experiment: the AI Anchor Desk

This is the part I'm most excited about.

In 2024, two economists, Van Pham and Scott Cunningham, found something strange. When they asked ChatGPT directly to predict the future, it was mediocre. When they asked it to write a story set in the future, a scene where someone looks back and reports what happened, its predictions got noticeably better. I wrote about their paper at the time on Hello AI.

So we're putting that to a real test.

Every day, we ask Claude and ChatGPT to write the election-night broadcast for every major race. A made-up network, the Continental News Network. A made-up anchor, Dana Whitfield. "We now have a projection to make in the race for U.S. Senate in North Carolina…" And then the AI writes who wins.

We run each scene many times and count. If ChatGPT calls the race for one candidate in 7 scenes out of 10, that's a 70% chance in ChatGPT's eyes.

Each AI writes two versions:

• Blind: just the race, the candidates and the date. Whatever the AI "believes" on its own.

• Briefed: we also hand it our polling average, the latest polls and the money. But never our odds, so it can't just copy us.

Here's why this is a fair fight. In the original study, the trick worked best when the answer was already somewhere in the AI's training data. November 3, 2026 isn't in anyone's training data. Nobody, human or machine, knows how this ends.

On the members page, one switch lets you see our model alone, or our model with the AIs mixed in. On election night, we'll know whether the machines saw something we didn't.

What members get

• The full forecast: every race, updated every day

• A page for each race, with the polls, the money, the history and the AI scenes

• "Money moves": every outside group that suddenly starts, or stops, spending in a race

• The new district maps, and which party each one helps

• The full method, in plain English

It lives behind the member wall at members.rationalground.com/elections.

If you've been reading Rational Ground for free, thank you. If you want to watch this race the way we do, and you want this work to keep going, this is the moment.

Become a member: $56 for your first year instead of $70 (offer ends Nov. 4)

They'll tell you who's going to win. We'll show you the math.

— Justin