2026 Midterm Election Polls and Predictions: The House Forecast, and the Senate

The national generic ballot has Democrats ahead by 6.7 points. That is one composite figure rather than six averages agreeing — and either way, it does not tell you who wins the House.

The reason is that the House is decided by six seats, not by six points. Below is what the average actually says, where presidential approval sits, what the seat arithmetic requires, what our own arithmetic makes of the seats, and why crossing from one to the other takes an assumption rather than a calculation.

What do the 2026 generic ballot polls say?

Democrats 48.2 to Republicans 41.5, a lead of 6.7 points. That is one composite figure, published by USPollingData and read on 14 August 2026, and it is the only generic-ballot number the seat arithmetic uses.

That composite is a weighted average of six aggregators — Decision Desk HQ, FiftyPlusOne, RealClearPolitics, Silver Bulletin, VoteHub and Race to the WH. The site that publishes it publishes the blend, not the six inputs.

One of the six is worth reading on its own. Silver Bulletin’s headline average, updated 14 August, is D+6.7 as published, and D+8.1 once it applies its likely-voter screen.

Hold on to that second number. A single aggregator’s own choice about who actually turns up moves its lead by 1.4 points — more than the distance between most published averages. And hold on to the 10.3%: the share no average can allocate is larger than the lead itself.

How many seats do Democrats need to win the House?

Six, on the current composition. This is the number that actually decides control, and it is smaller than the polling drama suggests.

Two things worth pinning down, and the first one matters more than the six.

Six is the gap to 218. It is not six seats taken from Republicans. Three of the four vacancies were Democratic-held — Eric Swalwell’s California seat, Sheila Cherfilus-McCormick’s in Florida, and the late David Scott’s in Georgia — against one Republican, Tony Gonzales’s in Texas. Fill each with the party that held it and Democrats stand on 215. Three is the number that has to come off the Republican side, and it is the number the ratings section below is really about.

Second, Republicans are not one seat from losing control. Kevin Kiley of California sits as an independent and caucuses with them, so they command 219 votes, not 218. And in the chamber as it stands, with four seats empty, a floor majority is 216 of the 431 members sitting. The 218 used throughout this article is the November threshold, when all 435 seats are filled again.

The “seats needed” figure you see elsewhere is often not six. USPollingData states the House as 222 Republicans to 213 Democrats and puts the requirement at five. Measured against the 2024 result — Republicans 220, Democrats 215 — it is three. None of those is a different calculation; they are different starting compositions, which is why the one above is taken from the Clerk of the House and printed in full.

Which seats are actually rated for each party?

211 rated Republican, 205 rated Democratic, and 19 the raters decline to call. These are the Cook Political Report’s district ratings as of 13 August 2026 — the day Cook last moved a district, shifting CO-03, OH-10 and OH-15 from Safe to Likely Republican — read the following day and aggregated by state.

Each tile is one state, drawn the same size because what is being counted is seats, not land. Blue means the delegation is rated majority Democratic, red majority Republican, grey an even split — Maine, Michigan, Minnesota and Pennsylvania are the four even ones.

Now put those ratings next to the majority line, and the race gets much clearer:

Democrats need 13 of the 19 toss-ups; Republicans need 7. That points the opposite way from the generic ballot, which is the tension in this election — but it is not the whole story either, because it assumes nothing outside the toss-ups moves. The section after next tests that assumption and finds it is the one doing the work.

The 19 toss-ups sit in 14 states, and four states carry more than one:

Which way are the raters moving?

One way. Every district the three of them moved in their most recent published update went toward the Democrats — eight changes, no exceptions.

Seven of the eight weaken a Republican-rated seat rather than strengthen a Democratic one: Cook took CO-03, OH-10 and OH-15 from Safe to Likely Republican, Sabato took OH-07 from Lean Republican to Toss-up on 10 August, and Inside Elections moved IA-02 and OH-07 from Likely to Lean Republican and IA-03 from Tilt Republican to Toss-up. The eighth, MI-08, went from Likely to Safe Democratic.

Two cautions before reading anything into it. A move is not a flip — Safe to Likely Republican is still Republican, and none of these eight seats changed hands on paper. And these are one update each; three updates are not a trend line.

One rater does publish its whole cycle. Sabato’s own change log carries 105 House rating changes since June 2025, and it also shows that dates on aggregator sites are not always the rater’s: the OH-07 move above is dated 13 August by the site we read the map from and 10 August by Sabato itself. We have used Sabato’s.

What it does establish is direction, and it matters later: the raters are still in the process of downgrading Republican seats. A ratings snapshot is a lagging measure of an environment that is still moving, and the arithmetic further down this page treats that snapshot as settled.

What is our own seat projection for the House?

Nine ways of running the same arithmetic give Democrats between 214 and 221 seats, and three of the nine reach a majority. Two choices decide which — which rater you believe, and where you draw the line around “in play” — and in eight of the nine cells the polling is not one of them. That is the finding, and it is worth more than a number would have been.

The rule has two steps. Take a rater’s ratings as given, then allocate whatever counts as “in play” in proportion to the two-party share of the generic ballot — Democrats’ 48.2 against Republicans’ 41.5, or 53.7%. Run it against all three raters who publish a House ladder, at three definitions of in play:

Democratic seats out of 435; 218 is a majority, and bold marks the cells that reach it. Inside Elections uses a five-step ladder with a “Tilt” rung between Toss-up and Lean; folding Tilt into the Lean row so the three fit one grid is our alignment, not theirs.

The rater matters more than the data

Look along the top row. On the narrowest, most cautious reading — only genuine coin-flips are in play — Cook gives Republicans 220, Sabato gives Democrats a bare majority at 218, and Inside Elections gives Republicans 221. Same rule, same national number, three ladders read within a week of each other. The spread is only four seats, Democrats 214 to 218 — but 218 is inside it, so those four seats are the difference between the two answers a reader actually cares about.

For most of the grid the generic ballot is not doing the work either. Eight of the nine cells would need the national lead to move somewhere between 6.4 and 41 points before their winner changed, and it currently reads D+6.7. The exception is the bottom-right cell: Inside Elections at its widest sits 0.3 points from flipping, close enough that any rounding decides it. Everywhere else, the answer was settled before the polling was consulted.

The rule can be worse than the ratings it starts from

Sabato’s column does not move in one direction. Toss-ups only gives Democrats 218; widening to Lean takes them down to 216; widening again puts them back to 221.

That is not a quirk, it is the method failing in public. Sabato rates 13 Lean Democratic seats against 7 Lean Republican, so Democrats are favoured in 65% of that rung — and our rule overwrites that with the national figure of 53.7%, handing back fewer seats than the rater’s own judgement gave them. A rule that replaces a district-level assessment with one national number can take away seats that the rater says you are winning. It is the clearest illustration on this page of what a national average cannot do.

Why the models still go further than any of our nine

Because they weight every competitive seat instead of rounding it to whoever is ahead. That is the whole disagreement, and across our nine cells it is worth five to twenty seats — and still five to seventeen measured only against the widest row from each rater.

Cook’s ladder has 50 competitive seats outside the toss-ups: 29 rated Republican (20 Likely, 9 Lean) against 21 rated Democratic (11 Likely, 10 Lean). A model does not record a Lean Republican seat as Republican. It records something like a 65% Republican chance and books the other 35% to the Democrats. Add those fractions across 29 seats and you get a double-digit number our arithmetic scores as zero — and because Republicans hold eight more seats in that exposed band than Democrats, what the rounding discards falls almost entirely one way.

There is a third difference, and the arrows further up this page are the evidence for it. A ratings snapshot lags the environment, and we treat it as final. Rating agencies move a district only once they are convinced, because being early is punished more than being late — so in a shifting environment their labels trail it. Cook, Sabato and Inside Elections were all still downgrading Republican seats in their latest updates. A model reads the same underlying data and does not have to wait for a label to change; our rule cannot move a seat that a rater has not already moved.

There is a fourth thing our rule cannot see. A national margin is not distributed evenly: piling up extra votes in districts a party already wins raises the national number without moving a single seat. So a generic-ballot share is not a seat share even in principle, and any rule that converts one into the other — ours included — is making an assumption rather than a measurement.

Expressed as the share of Cook’s 69 competitive seats each forecast gives Democrats, the argument compresses to a single number:

We say 53.7%, they say 61 to 73% — and all three raters’ widest rows land within half a point of each other, just below 53.7%: Cook 53.6, Sabato 53.3, Inside Elections 53.1, each measured on its own competitive band of 69, 75 and 64 seats. That clustering is not agreement between the raters; it is the rule reproducing the national figure it was given. Our version assumes competitive seats fall in proportion to the national vote. The models assume the party with the wind behind it takes them at a higher rate than its vote share, which is what usually happens, because those seats sit on the tipping point where a small national swing changes a disproportionate number of hands. Correcting for that would mean putting a win probability on each rung, and none of the three raters publishes one, so we have not invented any.

What does Trump’s approval rating say about the midterms?

It is the input our arithmetic ignores entirely and the models weight heavily. Presidential approval is the structural variable in midterm forecasting, and Trump’s has been falling all year.

Pew Research Center, which asks the question the same way on the same panel each time, has it at 34% in its 20–26 April survey of 5,103 adults — the lowest of his second term. That is down from 37% in January, itself down from 40% the previous autumn.

Read across sources rather than within one and the same problem appears as with the generic ballot. USPollingData’s composite, read on 14 August, has approval at 38.2% against 58.8% disapproval — a net of −20.6. Decision Desk HQ’s July model cited 42% approve to 55% disapprove as one of its drivers. Pew, the composite and DDHQ are eight points apart on the same quantity, and each is internally consistent.

What matters here is not the level but that all three are far under water, and that this is exactly the sort of variable a district model uses and our rule does not. Our arithmetic sees one number — the generic-ballot share — and nothing else. A model that also weights approval, candidate quality, fundraising and past district results has four more reasons than we do to expect Lean and Likely Republican seats to fall, which is where the eleven-to-twenty-seat gap in the chart above comes from.

What about the Senate in 2026?

A separate problem, with a narrower path. Republicans hold the Senate 53-47 — a caucus figure. The official party breakdown is 53 Republicans, 45 Democrats and two independents, and both independents caucus with the Democrats. Award every seat to the party it is rated for and Democrats reach 47 before the toss-ups, so their only route to 51 is to sweep all four of them — and one of the four, Michigan, is a seat they are defending rather than chasing. Lose a seat they currently hold and sweeping the four is no longer enough: 46 plus four is 50, so they would need a fifth win from outside the toss-up list.

Cook and Sabato name the same four: Alaska, Maine, Michigan and Ohio. Inside Elections is stricter and calls only Michigan a toss-up, which is why it ends the cycle with more Republican seats than the other two.

The forecast models split here in a way they do not on the House. Decision Desk HQ’s July model had both parties averaging 50 seats and still made Republicans favourites, because a 50-50 chamber is a Republican one while the Vice President breaks ties. The Economist’s median is also 50-50 and leans Republican; Race to the WH leans the other way, and reported that new fundraising data had improved the Democrats’ position in Alaska, Georgia and New Hampshire. On the House the models agree and we are the outlier. On the Senate they do not agree with each other, which is a reason to treat it as genuinely open rather than as leaning either way. That is a harder ask than a six-seat House gain, because the Senate map is fixed by which states happen to be up. We set out the four seats, all three raters and the polling in each in our 2026 Senate control guide . Texas is not one of the four — Cook, Sabato and Inside Elections all had it Lean Republican in early August — but it has moved enough to get its own page .

What could still change before November?

The things that are not in any of these numbers. Three are worth watching specifically.

• Vacancies and special elections. These move the arithmetic in one direction, not both. California’s 14th holds its runoff on 18 August and Georgia’s 13th on 25 August, and both are Democrat-versus-Democrat contests — so within eleven days Democrats go from 212 to 214 and the gap above falls from six to four. Texas’s 23rd is Republican-held with no date set.
• Candidate filing and retirements. Open seats behave differently from incumbent-held ones, and the ratings agencies move districts when incumbents leave.
• The gap between polls and raters. In several races this cycle, the polling average and the professional ratings point different ways. When that happens, the raters are weighting things a national poll does not measure — how a district has actually voted, and who is on the ballot.

What this forecast cannot tell you

Who will control either chamber. This page reports what one composite average says, what presidential approval says, what the seat arithmetic requires, and what three raters’ ladders produce under three different definitions of which seats are in play. Those nine results run from 214 to 221 and straddle the majority line, so they are not a call. The district-level models all make Democrats favourites for the House and we have printed their seat totals beside ours rather than burying the disagreement; on the Senate those same models do not agree with each other. This page does not price any of it — no betting or prediction-market odds appear here.

It also cannot tell you that a lead in August survives to November. The generic ballot in mid-August is a measurement of now, taken before most voters are paying attention, and its value is as a baseline you can check against later — not as a result.