Tonight, against the forecast
As the calls come in, which did the model have right, and how surprised is it?
- Right — the side the model favoured won. Counted on every call, and separately on the calls it gave either side at least a 2% chance in, which are the only ones it could have got wrong.
- Surprise —
Brier = mean (p − outcome)²on the called races, besidemean p(1 − p), the Brier score the model expected on those same races. A calibrated model is surprised exactly as often as it said it would be. - Given the calls — each chamber re-read from only the simulations in which every called race went the way it was called: the page's own conditioning, with tonight's results as the pins.
The tug of war
Seat by seat, as the calls land: who has what, and where the model expects the rest to fall.
The country, tonight
Where the model's favourite has won, where it has not, and where every uncalled seat stands now.
- Solid — called, and the model favoured the winner.
- Ringed in gold — called against the model's favourite.
- Shaded — uncalled; the colour is the win probability re-read from tonight's returns, so it moves as results land: from the ballots themselves once 80% of a race is in (What the ballots say), from the conditioning below before that. Hover for the forecast beside it.
How the night is going
Against what the model expected this morning: is tonight a typical draw, or a tail, and which way has it been moving?
The record books
How far back do you have to go to find a night like this one?
- Seats — the Democratic total, against every past total that was a high-water mark going back from today.
- Midterm gain — tonight’s change from the seats held going in, against every past midterm gain that was a high-water mark; and beside it, the same for either party.
Predicting the rest
Given what has been called and counted, where does every uncalled race stand now?
μrest | obs = μrest + Σro (Σoo + τ²I)−1 (obs − μo)
- Observed — the counted margin in every race at least 20% in, and the sign alone in every called race without a count. A count's doubt grows with the share still out: about a quarter of a point when fully counted, ±3 at 93% in, ±6 at 87%, ±18 at 60%, ±31 at 30%.
- Where it stops being the right tool — the last few percent of a count. A margin at 97% in is not a measurement with an error bar, it is a lead in ballots, and the prior this conditioning still blends it with can pull a lead of under a point across zero. Chamber totals come from here; a mostly-counted race is read from its own ballots instead, in What the ballots say.
- Why early counts cannot stampede it — the gap between votes counted first and votes counted last is the same kind of gap everywhere in a state, and much of the country counts the same kind first. That shared early-count lean is a latent term in the conditioning, not independent noise, so three hundred half-counted races cannot together pin the national swing to their early-vote lean. As the counts complete the term vanishes and the count is the result.
- Moved — every uncalled race, through the covariance the model itself put between races: the shared national term, the regional term and the state term. A Florida district running five points left of its forecast moves the rest of Florida most, the South next, and everything a little.
- Why not just keep the matching simulations — that is what holding a call fixed does, and it is exact; but every close call halves the pool, and by mid-evening nothing survives. This conditions the joint as a Gaussian instead, and control is recounted from fresh draws of it.
Where tonight sits on the shared error
Are the races moving together, and which way?
shared error ≈ mean over mostly-counted races of (counted margin − forecast margin)
- The model's largest single error term is the one every race shares — the national swing, with a fitted standard deviation it states in points. The mean residual across mostly-counted races is a direct reading of it.
- Early returns are whichever votes a state counts first, so a race's margin is read only once most of its expected vote is in, and the mean is shown at two thresholds with the count behind each.
- To see the correlation matrix itself under tonight's calls, hold them fixed with the button above and open Shared error.
Every call so far
Most surprising first: the probability the model gave the eventual winner.
What the ballots say
For every uncalled race with a count: how many ballots are left, what they would have to run, and whether that is plausible.
final lead = counted lead + ballots still out × their lean
- Still out — AP's remainder, corrected by how far its expected totals have missed tonight at the same stage, measured from the races that have finished: the final count has landed below the expected total in most of them. Once AP pins a race at 99% the remainder is a placeholder, and the reading assumes 0.3–2% of the count is still to come in provisionals, cures and overseas ballots.
- Needs — the margin the trailing side must win the remainder by. The late ballots' lean tonight is measured for its spread about the running margin, not its direction: the lean of the ballots counted so far barely predicts the lean of the rest.
- Recount range — a lead under half a percent of a nearly complete count is a question of provisionals and recounts, and is labelled one rather than priced.
- Now — from the ballots once 80% of the vote is in; from the conditioning above before that. The map, the headline and the tug of war use the same number.
Bluer or redder than 2024
Compared with how the same places voted for president two years ago: how many moved toward the Democrats, how many toward the Republicans, and by how much?
shift = counted margin − 2024 presidential margin, same district or state
- Baseline — Harris minus Trump in 2024, on the 2026 district lines where a state redrew, and the statewide vote for Senate and governor races. It is the one number every race has. A midterm electorate is smaller than a presidential one, so a shift is partly who turned out and partly who changed their mind; Who turned out separates the two.
- Counted — a race enters once 90% of its expected vote is in. Before that a margin says more about which precincts report first than about the electorate.
- By votes cast — the median treats every race alike; the vote-weighted shift is the shift of the ballots themselves, which a big district moves more than a small one.
- Forecast expected — the same subtraction with the model's median margin in place of the count, so the two can be read side by side.
Who turned out
How much of the 2024 electorate came back, where it held up, where it fell away, and whether the places that lost the most voters are the ones that moved?
turnout = 2026 votes ÷ 2024 presidential votes, same district or state
- 2026 votes — the count once AP reports a race complete; AP's expected total while it is still being counted, which has run a little high tonight.
- Harris and Trump districts — the same ratio over the districts each carried in 2024. If one side's voters stayed home, it shows here before it shows in any margin.
- Against 2022 — the last midterm's House vote by state, where every district in the state is in and none went uncontested. The presidential comparison says who came back; the midterm comparison says whether this was a big midterm.
- Turnout against shift — a district that kept its voters and still moved is a change of mind; one that lost half of them is a change of electorate.
Where it moved
Every district and state, shaded by how far it moved from its 2024 presidential vote.
- Blue — bluer than the place voted for president in 2024; red — redder. Full colour at thirty points. Grey is a race under 90% counted. Hover for the two margins and the turnout.
P(control) = runs won ÷ 20,000, all 506 races drawn together in each run.- Whether these probabilities have come true (calibration): Can you trust it.
Every seat, in order
Where does the majority fall, and which seats sit on it?
- Order — median margin, safest Democratic to safest Republican. Hatched: not on the ballot.
- Colour —
P(D wins); amber where an independent is favoured. - Outlined — where control changes hands. Independents sitting out: where Democrats first out-number Republicans. Free to choose: where each party reaches its number, D from the left and R from the right; a split between two outlines leaves the independents deciding.
Seat totals
How many seats, and not just who wins?
Both chambers at once
Can one party take both?
One mark per outcome: a run's House and Senate totals. Each quadrant is one way the night ends.
Seats that could change hands
Which seats are most likely to flip, in any chamber?
P(flip) = P(the party holding the seat loses it)
- Expected flips —
Σ P(flip); the 80% range is counted from the simulations. - Net —
Σ P(flip → D) − Σ P(flip → R). Independents' gains are listed apart. - Redrawn — new lines for 2026, so part of that flip is the map.
What to watch
Which races is this resting on?
- House, Senate — ranked by tipping-point share: the share of 20,000 runs in which the race casts the majority-making vote.
- Governors — no majority, so ranked by distance from 50%.
The House, by district
Where are the seats, and how much does the model know about each one?
- Suppress uncertain colour — grey where the model knows least (uncertainty suppression).
- Play simulations — one simulated election night per frame.
Every race
What does the whole field look like, race by race?
One dot per race at its median margin, Democrats left. Click any dot or row for its working.
What moved it
Why is this number different from last time?
Each cause is priced by re-running the forecast with only that input changed. The causes sum to each card's movement.
Which races moved
The chamber shifted — where did that come from?
The most decisive seats by tipping-point share, on one shared scale. Zero is a tie.
Forecast history
How has this forecast moved over time?
One point per published run. Solid: same method at both ends. Dashed: the method changed, so the step is partly the model.
What is still to come
How much more will the model learn before November?
The national polling
What happens to the forecast if the polls are off?
The generic ballot is the largest single input, and it leans:
- Correction — 2.74 pts toward Democrats, subtracted; overstated in 10 of 13 past cycles.
- Leverage — 13 pts of House control per point it is wrong.
The slider re-runs the whole forecast at another assumed lean.
Shared error
When the model is wrong about one race, is it wrong about the others?
Usually yes. Each race's error is four terms, three of them shared:
σtotal = √(σnat² + σreg² + σstate² + σidio²)
- σnat — a national miss, shared by every race in the country.
- σreg — shared across the region.
- σstate — shared across the state. Worth 3.39 points.
- σidio — this contest alone. Student-t, 5 degrees of freedom.
The House map
How far is the map from the country it is drawn on?
lean = district margin − national margin, 2024 presidential
The 218th district sits 3.26 points right of the country: Democrats need to win the national vote by about that much to carry it.
The state borders
And how far are the Senate and the governorships?
The same 2024 lean, scored by state. Unlike district lines, state borders are not redrawn.
The model, end to end
What happens between a poll and a seat count?
- Two estimates of each margin — a corrected poll average, and a prior built without polls. Every race has a prior; most have no polls.
- Blended by how much polling there is, then simulated with errors shared nationally, by region and by state, and counted.
Each box links to the section that shows what was fitted for it.
Pollster ratings
How much does each pollster count, and which way is it corrected?
poll′ = margin − hpollster − s · sponsor
Polling error
How wrong is a poll, and how much of that can more polls fix?
- One poll
√(σ_nat² + σ_reg² + V_race(d) + V_house(d) + V_noise(d)) - Floor
√(σ_nat² + σ_reg² + V_race(d)). No number of polls goes below it. - d: days between fieldwork and the election. Shaded: past the corpus, extrapolated.
Sponsored polls
Do polls paid for by a party lean toward it?
One race, step by step
How does a single seat get its number?
Every race goes through the same seven steps. Pick one and step through.
Checks against past elections
Has this model been right before?
- Refitted without the scored election — pollster ratings, house effects and the polling error curve.
- Fitted with it — the prior's width, the state error term, elasticity and the generic-ballot correction, each pooled over the same cycles the House replays score. Those replays are partly in sample, and read slightly kinder than a clean hold-out would.
What could be wrong
What does this model say is wrong with it?
- Alarms — recomputed every run; they come and go with the data.
- Limitations — structural; they clear only when the model changes.
What the model is working from
How much of this is polling, and how much is inference?
Mostly inference: almost nobody polls a safe district. An unpolled race runs on its prior, with a wider error bar.
Payload diagnostics
Is the page showing what the engine actually computed?
Every total here is recounted in your browser from who won each race in each of 20,000 draws. These checks confirm it matches the engine.