How we grade the 2026 midterm forecasts

1. Who and what is graded

Four forecasters are graded on the same rules:

• Inside Elections, race-by-race expert ratings;
• FiftyPlusOne, a model that blends polls, fundamentals and expert ratings into win chances;
• Race to the WH, a model built on polls and fundamentals;
• DDHQ / The Hill, a model built mostly on fundamentals, with prediction-market prices at about a quarter of the weight and polls, as its own race data describes.

FiftyPlusOne lists Cook, Inside Elections and Sabato among its inputs, so FiftyPlusOne and Inside Elections are not independent of each other. DDHQ / The Hill is the only one of the three models that reads prediction markets.

There are 506 races on the ballot: 35 for the Senate (including special elections in Florida and Ohio), 36 for governor and all 435 House seats. FiftyPlusOne and DDHQ / The Hill publish no governor forecasts; those races show as "not rated" for them.

Also graded: ten more raters. Cook Political Report, Sabato's Crystal Ball, Split Ticket, Fox News, the Almanac of American Politics, Silver Bulletin, NBC News, RealClearPolitics, The Economist and Decision Desk HQ's own ratings. Their ratings come from the public rating tables on Wikipedia (CC BY-SA 4.0) and from 270toWin's per-outlet maps, as they stood at the final seal. When both list the same race for the same rater, the one with the later "as of" date is used, and on equal dates the 270toWin map. If the two copies differ in rating or side, the race carries a flag, and the rating graded is still the one chosen by date. They are graded exactly like the main four, by published rating, on the same battleground list, each over the battlegrounds it rated, in a table of their own.

2. Which forecast counts

Every collection run is sealed: its forecasts get one cryptographic root with independent timestamps, and the page shows the time of the latest seal. Two seals on November 3 freeze the record: 5:35 AM ET, before polls open, and 5:35 PM ET, 25 minutes before the first polls close. The evening seal is the one we grade. A note shows whether any forecaster changed a call during election day.

A forecaster's forecast for a race is the latest version on record at the evening seal, shown with its date. If a candidate withdraws or is replaced, the latest forecast before the seal still counts. Every single forecast, race by race, is part of the seal, so anyone can check one line without taking the whole record on trust.

3. Sides, candidates and special cases

• Ratings are read on one scale: Solid or Safe, Likely, Lean, Tilt and Toss-up. Inside Elections and Race to the WH use Tilt; Tilt is a call like Lean.
• Two sides per race: the party that holds the seat and its main challenger. The seat-holding party is the same for every forecaster.
• Independents. In Nebraska (Dan Osborn), Idaho (Todd Achilles), South Dakota (Brian Bengs) and Alaska's House seat (Bill Hill) the main challenger is an independent and no Democrat is on the ballot. Forecasters label these candidates differently, some as Democrats and some as independents, so we match them by name. In Montana a Democrat is on the ballot as well, so the Democrat and the independent (Seth Bodnar) are kept apart in three-way accuracy; in every two-side score the challenger's side is everyone but the Republican.
• California's 6th district. Kevin Kiley left the Republican Party in March 2026 and runs with no party against Democrat Richard Pan. No Republican is on that ballot, so a forecaster's "Republican" in this race means Kiley, graded as an independent holding the seat; a win by Pan counts as a flip.
• Same-party races are not graded. In a top-two primary state the two leading candidates advance whatever their party. In nine California districts both November candidates belong to one party (eight with two Democrats, one with two Republicans), so the winning party is known before the vote. These races are listed separately and left out of every score.
• Placeholder names such as "Generic Republican" or "A Democrat" are never shown as candidates.
• Where a model's chances for all candidates add up to slightly more or less than 100%, they are scaled to 100%.

4. Battleground races

A race is a battleground if at least one of the four forecasters rates it Lean, Tilt or Toss-up, or if any of the three models gives the favorite less than a 75% chance at the seal. Same-party races are never battlegrounds. At the latest seal, Nov 1, 5:10 AM ET, there are 98 battlegrounds: 14 for the Senate, 14 for governor and 70 in the House (100 at the first count, Oct 24).

The headline scores use battlegrounds only. A score across all 506 races would mostly reward calling safe seats.

5. How a call is scored

The rating call, the same for all four. A published rating other than Toss-up is a call for the side it names. Toss-up is no call, for everyone. A model that labels a race Toss-up gets no call even if its number leans 51 to 49.

Points: right 1, no call ½, wrong 0, over every battleground the forecaster rated. Next to the points: the right, wrong and no-call counts, the points share and the hit rate among the calls actually made. The table is sorted by points share, not by hit rate, so a Toss-up counts as half a point instead of dropping out of the count. A race a forecaster did not rate is not in its count at all, which is why the default view compares all four on the same races.

The default view compares all four on the same races: every Senate and House battleground. Governors, forecast by Inside Elections and Race to the WH, have their own tab. A view of all races is shown too, with its different counts.

Two extra records for the three models:

• Forced call: whoever the model gives more than 50% is its pick, with no abstaining. It reads right, wrong and tie; an exact 50–50 is a tie, so the three numbers always add up to the races rated.
• Common 60% line: a pick only where the model gives one side at least 60%, scored with the same points (right 1, no pick ½, wrong 0) on the same races. Inside Elections is scored here by its ratings, Tilt included.

The models use different cut-offs for their own labels (FiftyPlusOne starts Lean at 60%, Race to the WH calls anything under 65% a Toss-up), which is why the forced call and the common line sit next to the rating call.

6. How the chances are scored

• Probability accuracy: the gap between the chance a model gave the winner and 100%, squared and averaged over every called race the model forecast (all chambers, battleground or not) and, separately, over battlegrounds. Lower is better.
• Probability skill vs a 50/50 coin: one minus the battleground accuracy divided by 0.25, the score of saying 50–50 in every race. 0% means no better than a coin, and below zero means worse. On the scoreboard it is computed on the races of the tab shown, so the default view compares the three models on the same Senate and House battlegrounds; a forecaster's own page uses every battleground it rated, which for Race to the WH includes governors.
• Three-way accuracy: for Democrat, Republican and independent, after matching candidates by name, the squared gap between each stated chance and the result (100% for the winner, 0 for the others), added up for the race and averaged over every called race the model forecast. 0 is perfect, 2 is the worst.
• Calibration: every called race the three models forecast, all chambers. Each race has two sides here, the seat-holding party and everyone else, so the favorite's chance is always at least 50%. Races are grouped by that chance (50–64%, 65–79%, 80–94%, 95–100%), and each group shows how often the favorite won, with a 95% range.
• Ratings held: for all four, over every called race, how often each rating level held, as "k of N".
• Flips: over every called race, the races where a forecaster's rating named the challenger, against races the challenger actually won.
• Seats against expectations: for the seat-holding party, the seats it actually kept minus the sum of the chances the model gave it, over every called race the model forecast.

7. Disagreements, late moves and surprises

• Disagreements: races where forecasters named different sides, or a battleground where some named a side and others said Toss-up. Races where they named opposite winners are counted apart.
• Late moves (on the page: forecast changes): any change of rating, of favorite, or of a model's lean across 50% between two sealed runs, from October 21 to the seal, in any race, battleground or not. A change that comes with a switch of the file we read for that outlet is left out of the count. Each move is dated by the first sealed record that shows it and marked as a move toward the Democrats or toward the Republicans: a move away from the Republican counts as toward the Democrats, also where the challenger is an independent. Chances on the page are rounded, so a rating can change while the number shown stays the same. After the results, each move is marked as having helped or hurt. Inside Elections' October changes are listed by its own change dates, apart from that count.
• Surprises: battlegrounds where every forecaster that made a call got it wrong.

8. Results

• A race counts once a winner is called; certification is not awaited. Calls come from Decision Desk HQ's published race results and from a results log where every entry names its source and time. The time of a call is the time it first appears in our sealed record.
• Until a race is called it stays out of every score, in both the numerator and the denominator. The page shows how many battlegrounds are called and lists the rest with the reason.
• Runoffs: Georgia's Senate and governor races go to a runoff on December 1 if no one passes 50%; Louisiana's six House districts hold an open primary on November 3 and a runoff on December 12 under the same rule. These count once decided. Louisiana's Senate race is a regular November 3 race.
• Ranked-choice races in Maine and Alaska count after the final round.
• If an early call and the certified result differ, the certified result stands and the change goes into the corrections log.

9. What the page shows

• Before the vote: how each forecaster splits the races between the parties and Toss-up, the chamber totals if every favorite wins (the Senate total includes the 65 seats not on the ballot), the races where they disagree with each forecaster's rating, the latest forecast changes, and the time of the latest seal.
• After the vote: the scoreboard, a page for every battleground and every forecaster, calibration, disagreements, surprises, the races still waiting for a result, and the table for the ten other raters.
• Each forecast belongs to its outlet and is credited with a link to it. FiftyPlusOne: "Data from FiftyPlusOne". Wikipedia rating tables are used under CC BY-SA 4.0.
• Times on the page are Eastern Time.
• Corrections log: any change to the grading code and any entry in the results log is a dated line showing the old and new versions.