The method, printed in full
A forecast the reader cannot reproduce is a brand, not a measurement. Everything this desk uses to produce a probability is on this page: the inputs, every coefficient, the seed, and the things the arithmetic is blind to.
- 01
Start from something certified
Every seat begins at its state's certified 2024 presidential margin. Not a poll, not a memory of the state's reputation — the filed return. If the baseline is wrong, it is wrong in public and the source is one click away.
- 02
Carry it forward, imperfectly
Statewide candidates run a few points off the top of the ticket, so the presidential margin is carried at less than one-to-one. This single coefficient is the difference between a model and an assumption that a Senate race is a presidential rerun.
- 03
Apply the midterm drag
The president's party has lost ground in nearly every midterm since the war. That prior is applied as a fixed number of points against the president's party, printed below, and it is applied to both parties symmetrically whichever one holds the White House.
- 04
Adjust for who is actually running
Incumbency, open seats and appointed senators seeking their first full term each carry their own printed adjustment. Nothing else about a candidate enters the model — no fundraising, no endorsements, no vibe.
- 05
Draw the year twenty thousand times
Each run draws one national swing shared by every seat, plus an independent per-seat error. Correlated error is the reason a model can be wrong about thirty races at once, and any forecast that omits it will understate its own uncertainty badly.
- 06
Publish the whole distribution
The output is not a number, it is a shape: how often each seat total occurs, how often each seat casts the deciding vote, and how often each party ends up in control. A 52% call with a fat tail is a different object from a 52% coin flip and the reader is entitled to see which one this is.
If a number is not in this table, it is not in the model. Changing any value here produces a new version number, and the old one stays on the page with the date it changed.
Senate results track the presidential baseline of their state closely but not perfectly — statewide candidates still run a few points off the top of the ticket. The state's 2024 certified margin is carried at 0.78 rather than 1.0 so the model does not assume a Senate race is a presidential rerun.
In the 19 midterms since 1946 the president's party has lost national vote share almost every time. 5.6 points is the central tendency of that series, applied uniformly rather than tuned per state, because a uniform prior is the one a reader can check.
Name recognition, casework and money. Applied only where a sitting senator is actually running; retirements get nothing, which is why open seats move toward the state's baseline.
A senator who was appointed rather than elected carries a thinner version of incumbency. The bonus is reduced, not removed — they still hold the seat and the machine.
One draw per simulation, shared by all 35 seats. This is the correlation that makes a bad night a bad night everywhere at once — a model with independent seats produces a confidently wrong, far too narrow distribution.
Candidate quality, primaries, scandals, local drift. Independent per seat.
No incumbent means no track record, and an unresolved primary means the candidate is not yet a known object.
Every point of national swing has moved roughly 5.4 House seats across post-war midterms. The desk applies that rate to the same national-swing draw the Senate model uses — one environment, two chambers — rather than estimating 435 races against maps that are still in court.
The House boardA separate model from the Senate one, sharing its national-swing scale so the two boards cannot describe two different years. Coefficients are frozen; changing one is a new version, dated, printed here.
Governor's races track the presidential baseline about half as tightly as Senate races do. State offices are routinely won by the party that loses the same state federally, so a high carry-over would produce confident nonsense in exactly the states that decide this board.
The same post-war midterm penalty the Senate model applies, discounted because governorships are less nationalised. It is applied uniformly rather than tuned per state.
Larger than the Senate figure, and this is the best-evidenced asymmetry in the file: sitting governors are re-elected at a substantially higher rate than sitting senators, because voters grade an executive on potholes and storms rather than on the national argument.
The identical figure the Senate model uses, drawn once per simulation and shared across boards. One environment, two chambers and the mansions.
Considerably wider than the Senate's 5.5. Candidate quality, scandal and local record decide governorships more often than the national mood does, and the desk estimates none of those — so they are paid for honestly as variance.
Half this board is open because the incumbent is termed out. With no record on the ballot and unresolved primaries on both sides, the distribution is wider than a defended seat's.
- Only the twelve races the desk rates as live are modelled. The other twenty-four mansions are omitted rather than handed a fabricated near-certainty that would flatter the calibration score.
- Candidate quality is the dominant term in a governor's race and the model does not estimate it. It is carried as a wider error, which widens the interval instead of moving the centre.
- Michigan's live three-way risk is not modelled. A credible independent breaks the two-party assumption every probability on this board rests on, and the desk says so rather than absorbing it into noise.
- Maine runs ranked-choice generals. The model produces a two-party probability; a first-round lead there is not a result.
- No polling. When these races have a public polling record it enters as its own printed coefficient on a dated version bump.
- No polling is in the model. When these races have a public polling record, polling enters here as its own coefficient, with its own line in the table, on a dated version bump.
- The midterm drag is a single national number, not a per-state estimate. States where the president's party has historically bled faster than average are not being credited for it.
- Candidate quality is inside the per-seat error term rather than being estimated. A badly chosen nominee shows up as variance, not as a moved central estimate.
- The 2024 presidential baseline is two years old and cannot see any drift since. That staleness is one of the things the state error term is paying for.
- A generic-ballot margin outside four points that holds for a month, rather than a single survey with a friendly sample.
- Final maps in the mid-decade redistricting states landing more than five net seats away from the current draw.
- Special-election overperformance running consistently above the historical relationship with the eventual national margin — the earliest honest signal there is.
- A retirement in any seat currently rated Safe. Incumbents leave when they can count.
- Presidential approval crossing 45 or falling under 38 and staying there through Labor Day.
Structure: sitting members, class-2 Senate seats, 2026 gubernatorial cycle · retrieved 2026-08-11
Every probability produced by this method is locked on the record and written against the Standard.