How the Forecast Works

The short version

The model takes each competitive race, blends three independent reads on it (the ratings, the polling, and the state’s underlying partisan lean), adjusts for the national mood, and then simulates the entire map 10,000 times. A party’s odds are simply the share of those simulations it wins. Nothing here favors a party by design. The inputs are public, the weights follow from how reliable each input is, and the same recipe runs for a seat in a deep blue state and a seat in a deep red one.

How the forecast works

1. Ratings

For each race the model starts with the published competitiveness ratings (Cook Political Report and Sabato’s Crystal Ball, where available). A rating like “Lean R” or “Toss-Up” is translated into an expected vote margin using a lookup that was trained on real Senate outcomes from 2014 through 2024. In other words, “Lean R” means whatever “Lean R” races actually did over the last decade, not a number someone picked by feel.

2. Polling, weighted by age

Where a race has polling, the model folds it in. Fresh polls count for more than stale ones. The reliability of a poll degrades as it ages, so an old survey is widened toward uncertainty and pulled back toward the rating-based estimate. A race with no recent polling leans more heavily on its rating and its baseline (below). A race with fresh, plentiful polling lets the polls do more of the talking.

Only some polls count. Since 29 September 2026 the model reads a race poll only when it is cited to the pollster’s own release page. A poll cited to an aggregator, a news report, a prediction market, a campaign or party committee, or a page about a different office is left out, and the forecast page counts each kind. No polling aggregator feeds a race. The pollsters we name, and the credit each one asks for, are in where each input comes from.

There is a hard floor under that. Poll age is counted in whole days, and a poll we can date is dropped once it reaches 46 days old. Up to that point the age penalty applies and the poll keeps a reduced say; past it the poll is dropped, the race runs on its rating and its baseline, and the race is marked as carrying no current polling rather than being handed a number that a survey from six weeks ago helped set. The line is hard, not a steep discount: a poll on its 45th day still counts, and on the next day it does not.

Some of the polling we hold arrives with no usable date on it. A missing date is not evidence that a poll is fresh, and we will not invent one, so those polls are dropped as well and the race runs on its rating and its baseline. The reason is plainer than it looks: a poll we cannot date can never reach the cutoff above, however old it really is, so exempting it would put a share of our polling permanently outside a rule we publish. Each run keeps its own record of the two kinds of exclusion separately, the stale and the undated, because the first is a poll that aged out and the second is a gap in what our sources gave us. That record is not yet something you can browse race by race on this site, and until it is, the count above is what we can show you.

Added 24 August 2026. Until this change an old poll was widened toward uncertainty but never removed, so a spring survey still helped set a race’s number in August. We measured the change before shipping it, by replaying the model against the facts as they stood on 24 August 2026: 119 of the 470 modeled races lost their polling to the cutoff, one of them a House district whose stored poll date read October 2016. In that replay the Senate topline did not move and the House probability of a Republican majority moved by less than a point. Those are the numbers of a test replay on one day’s data, not of the run you are reading.

Updated 25 August 2026. The day after that cutoff shipped we stopped exempting the undated polls, and this one was not a small change. Replaying the model against the facts as they stood on 25 August 2026, 72 more races lost their polling, the count of races the model could use polling for fell from 122 to 50 out of 470, and the House moved from a Republican majority probability of 25 percent to 44 percent, which carried its label from Lean D to Toss-up. The Senate did not move at all, because only two of the 72 were Senate races. We are telling you this rather than letting the number move quietly: the shift is a change in what the model was willing to trust, not something a voter did. What those 72 polls had in common was that we could not date them; several claimed margins their districts make very hard to believe, and one stored date read 2026-05-00, which is not a day. We are not claiming the polls that survived are sound. Those are the numbers of a test replay on one day’s data, not of the run you are reading.

3. Partisan baseline

Every state has an underlying lean that does not move much year to year. The model captures it from the 2024 presidential margin, expressed relative to the country as a whole: The Downballot’s results by district for House races, and the Federal Election Commission’s official results for Senate races. This is the floor under a race. It keeps an unpolled seat in a lopsided state from drifting on thin evidence, and it gives the model a sane prior when the other two signals are quiet.

4. The national environment

Races do not move independently. When the national mood shifts toward one party, it tends to shift in many places at once. The model carries that shared mood as a single national number and applies it across the board, then lets each race vary around it. That is why a small national swing can move a lot of seats together, and why the forecast does not treat the map as a set of unrelated coin flips.

That national number starts from the generic congressional ballot (the “which party would you vote for” polling average): the Average row of Wikipedia’s compiled generic-ballot table, which is licensed CC BY-SA 4.0. We take that one row and none of the page’s text. We do not feed it in raw, and here is the part we want a skeptical reader to see plainly: the generic ballot has typically overstated Democrats in recent cycles. So we shrink the raw number toward zero and apply a Republican-leaning correction sized to that pattern, a fixed slope-shrink plus a Republican-direction haircut (both disclosed, with their exact values, in the model parameters below) before the model uses it. If the raw average reads Democrats plus seven, the model works from closer to Democrats plus four. We would rather haircut a figure that flatters one side than run the raw number and let a known bias sit uncorrected.

We also widen the uncertainty on that national number the further we are from Election Day, on a fixed schedule set in advance. Six months out, the honest error bar on the national environment is wide (the generic ballot can be off by five points even on the eve of an election, and more than that this early). As November approaches, the band narrows on a pre-committed timetable, not because anyone decided the race got clearer. A forecast that quietly grew more confident as one party’s position improved would be telling you a story; the schedule keeps it honest.

5. Fusion, weighted by precision

For each race the model now holds three estimates (rating, poll, baseline), each with its own uncertainty. It combines them by trusting the more certain ones more. A confident, fresh poll outweighs a vague rating. A noisy poll defers to the baseline. The result is one combined estimate per race, with an honest error bar attached. No signal is allowed to dominate just because it is loud. It has to be precise to carry weight.

The three have to be measuring the same thing before they can be combined. Ratings and the partisan baseline describe how a seat leans compared with the country. A poll does not: it measures the race as it stands, national mood included. So the model subtracts the national number from the poll before combining, and adds it back once, in the simulation. Get that wrong and the national mood is counted twice for every polled race, which is exactly what this forecast did until 26 August 2026.

6. The simulation

With a combined estimate and error bar for every race, the model runs the whole map 10,000 times. Each run draws one shared national shock (so the races move together, as they do in real life) plus its own local noise for each seat. Every run produces a full seat count. Stack up all the runs and you get a distribution: not a single number, but the range of plausible outcomes and how often each one shows up.

7. Seats and probabilities

From that pile of simulations the model reports the median seat split (the middle outcome), an 80 percent range (the band that contains four out of five simulations), and the probability of a majority (the share of simulations in which a party reaches the threshold: 51 in the Senate, 218 in the House; a 50-50 Senate goes to the Vice President's party, currently Republican, so Republicans control it at 50). Joint control of Congress is counted from the same simulations, not multiplied together after the fact, because the two chambers tend to move in the same direction in a given year.

How to read the forecast page

The page answers one question per chamber with two instruments, then shows what the answer rests on. In the order it appears:

• Who controls each chamber. Two rows on one seat axis. The solid row is the main model, the published forecast. The dashed row is the structural comparison model. Each row states its median seats and its chance of a Republican majority in coarse words, and each carries its own run date. The main model’s row also prints its 80 percent seat range, a whisker on the axis with both ends stated; the comparison row prints none. The range is the model’s own interval and its accuracy is untested, as the calibration review above explains.
• The gap between the two models. The distance between their chances of a Republican majority, in points, printed beside the simulation noise that limits how finely two runs can be compared and beside the age gap between the two runs. A gap larger than the noise is not thereby meaningful, and a gap inside it is a limit of the measurement, not a finding that the models agree. When the comparison run is more than 14 days older than the main run, the page withholds the gap and says why.
• What the model cannot see. How many races carry a poll the model can use, how many had a poll the gate removed and why, and how many had none. The counts come from what the run itself recorded.
• How it moved. The change since the prior run, when the two runs share a model version. When they do not, the page says there is no comparable prior run rather than inventing a move.
• What if the mood shifts. The seat counts under a few pinned national environments (say, a Democratic-leaning year, a neutral year, and a Republican-leaning year). This is the only control on the page. It shows how sensitive the map is to the one thing nobody can poll yet: the mood on election day.
• Every race. A published rating beside the model’s own claim for each race, with a dashed bar for how uncertain the model is about that race, and a line saying what its number rests on. A solid badge is somebody else’s published rating, credited to its rater. An outlined badge is ours.

Where it can be wrong

A model is only as good as its inputs and its assumptions. Ratings can lag a fast-moving race. Polling can be sparse, herded, or systematically off, as it has been in recent cycles. The partisan baseline assumes the recent past still describes the present, which is exactly the assumption a realignment would break. The national-environment estimate is itself uncertain, and the correlation between chambers is a modeling choice, not a measured fact. The forecast reports ranges and probabilities precisely so that none of this is papered over. A 70 percent favorite loses about three times in ten. That is not the model failing. That is the model telling you the truth about how much it does not know.

Two models, read together, never blended

Beside the forecast above runs a second, structural model. It reads a district or state partisan lean, the national generic ballot, and incumbency, and it reads nothing else. No polls. No race ratings. It exists to disagree with the forecast, because a model built from different inputs disagreeing is information, and two models built from the same inputs agreeing is not.

The forecast page now shows both on one seat axis, and prints the distance between their chances of a Republican majority as the page’s hinge: the honest measure of how much the forecast depends on ratings and polls. The two are never averaged, ensembled or reconciled, and no blended number exists anywhere on the page. The comparison model is labeled as a comparison wherever its number appears, and it is not calibrated for publication.

Read the gap with its limits beside it. Each run draws 10,000 simulations, so two runs can differ by a fraction of a point from sampling noise alone, and the page prints that noise next to the gap. The two runs are also made at different times, and the comparison model is not rerun every night, so part of any gap can be the age of the older run. When the comparison run is more than 14 days older than the main run, the page withholds the gap and says why instead of printing a distance that mostly measures staleness.

The comparison has already turned up two defects in our own data. When the second model first ran, it disagreed with the forecast most sharply in a set of districts where the published forecast was leaning on polls carrying no date at all. Chasing that down found both problems: polls were being stored without recording which party led, and undated polls were never being aged out. The forecast above no longer uses either, and the two dated notes further up this page record what changed.

Dated reading, 25 August 2026. Against the forecast run of that day the two models were close on the House, both near a 2 in 5 chance of a Republican majority, and far apart on the Senate, the published model near a 2 in 5 chance and the structural model near a 3 in 4 chance. Two things drove the Senate gap and it is worth keeping them apart: the published forecast reads polls and race ratings that the structural model does not, and at the time the two models divided up uncertainty differently. The gap was not a clean measurement of either one on its own. That reading predates the variance rebuild described above and is kept as history, not as a standing score; the live figures are on the forecast page, each beside the date of the run it came from.

The second model is not calibrated. Its accuracy has not been tested against past cycles, no reliability curve has been run on it, and nothing about it changes the numbers in the published forecast. It is a check on our own work, published because a check nobody can see is not a check.

One thing the check found is worth stating plainly, because it was a criticism of the forecast above rather than of the second model. In the run compared then, every one of the 470 races came out with no independent uncertainty of its own, which is the subtraction problem described earlier. That has now been fixed. Each race carries an independent error bar sized by what is actually known about it, so a seat with a current poll, a seat the raters call competitive, and a landslide seat no longer read alike. Every race’s interval widened and none narrowed. A second defect surfaced alongside it: the forecast was fusing a poll, which measures a race including the national mood, with two signals that already have the national mood removed, and then adding the national mood again on top. That pushed every polled race toward the party the national number favored, by about one and three-quarter points. Both are fixed. Together they moved the Senate from Lean D back to Toss-up, which is a change in what the model says produced by fixing how it counts rather than by anything a candidate did. The House stayed at Toss-up.

What it is not

It is not a bet, an endorsement, or a reason to tune out. The point of putting the reasoning on the page (what each model reads, what it cannot see, and the gap between them) is that a forecast you can inspect is a forecast you can hold accountable. Watch whether the races resolve the way the chances said they would. That is the test, and we are keeping score.

← Back to the forecast