How our election-night forecast works: one county's count and every race

Summary

On election night our forecast changes as votes are counted. A partial count from a single county moves not only that county's races but every race in the country: a little or a lot, depending on how far the count can be trusted and how much the races have in common. Each further count adds to the picture and narrows the range of outcomes, until the forecast rests far more on the votes than on the polls. You can test it yourself with a county count of your own.

This works because our pre-election forecast already describes how polls tend to be wrong together. Polls rarely miss one state alone. When they underrate a party, they usually underrate it nationally, in neighboring states, and among voters of the same kind. So when a county's count comes in above or below what we expected, we can ask which of those shared misses would explain it, and move every race that shares them.

The forecast works in three steps, repeated each time a new count arrives:

1. Read the count. Compare the county's count with what we expected there, and judge how much of the difference is real and how much is the noise of an early, partial count.
2. Update the shared misses. Combine every reading so far into one estimate of how the polls missed: nationally, by region, by state and among groups of voters. That estimate moves every race, counted or not.
3. Project each race and the chambers. Where a race has votes counted, project its final margin county by county. Then run our full simulation of the election again, thousands of times, for the chances of control of the House and the Senate.

Replayed on the 2024 presidential count, with counties reporting mail ballots first as many do, the projected final margin in each state was off by 1.7 points on average with half the vote counted. The running total at that point was off by 4.0 points, and our forecast before any votes were counted by 2.8. With 90% counted, the projection was off by 0.5 points.

Why early counts mislead

The running total on election night is not a small version of the final result. Three things get in the way.

Places report at different speeds. A state's small rural counties often finish early, and its big cities late, or the other way round. A state count that is 40% in may be 40% of one kind of place.

Kinds of vote are counted in turn. Most counties count one kind of ballot first. In many states mail ballots come first, and since 2020 they have leaned Democratic. In others, election-day votes come first, and they lean Republican. So the first fifth of a county's vote is often not a fifth of its voters. In 2020 and 2024 a county's mail vote and its election-day vote often differed by 20 points of margin or more.

Some places are simply different this year. A popular local candidate, a factory closing or a big new housing development can move one county with no wider meaning.

A good forecast has to separate the parts of a count that tell us about the country (the polls were off) from the parts that do not (this county counted its mail first, or has a local favorite). Everything that follows is a way of doing that carefully and in numbers.

What we expect before a vote is counted

Every count is judged against an expectation, so before polls close we set one for every race and every county.

Each race. Our published forecast gives each Senate, governor and House race an expected margin (the Democrat, or the main candidate opposing the Republican, minus the Republican) and a range around it.

How races miss together. In every simulated election our forecast draws a set of hidden errors that shift many races at once:

Shared error	How many	Typical size, points of margin	Which races it moves
National	1	2.9 statewide, 4.9 on the House vote	Every race
Region (census division)	9	2.6	Every race in the region
State	51 (with DC)	3.9	That state's Senate and governor races in full, its House races by half
Voter groups	3	2.5 together	Every race, by its share of white non-Hispanic voters, college graduates and Hispanic voters
The race itself	one per race	3 or more statewide, 4 for a House race	That race only

The sizes come from history: how far each state's presidential swing strayed from the national swing in every election since 2000, and how far the national polls missed.

Each race has a loading on each shared error: how many points one unit of that error moves its margin. A Texas Senate race loads on the national error, on the West South Central region, on Texas and, through the make-up of Texas's electorate, on each of the three voter groups. An Ohio race shares only the national error and the voter groups with it.

Each county. A county's expected margin is its 2024 presidential margin, moved by the swing we expect in that state's race:

\(\text{expected}_c = m^{2024}_c + \left(\mu_{\text{race}} - m^{2024}_{\text{state}}\right)\)

Here \(m\) is a margin and \(\mu\) is our forecast for the race, so the counties add up to our forecast for the state. A county's expected vote is its turnout in the last comparable election. Its loading is its state race's, except for the voter-group part, which comes from the county's own make-up (Census Bureau estimates of race and ethnicity, and the Department of Agriculture's county education figures). A county much whiter or more college-educated than its state therefore moves with places like it elsewhere, not only with its state.

Where we know how a county voted by method in 2024 (election day, early in person, mail, provisional), we also hold an expectation for each kind of vote. When a state's results split a county's count by kind, the votes counted are compared with what we expected of those kinds, and the votes still to come are projected at the margin expected of the kinds still out: election-day votes after the mail, say, not the county as a whole. Replaying 2024 in the seven states where we hold this, it cut the projection's error by about a fifth with 10 to 30% counted (2.4 points against 3.0). It does not shrink the range: a kind of vote's margin moves from one election to the next, and when we let it narrow the range as well, the 80% ranges held only 71 to 76% of results and close races were called wrongly.

Turning a partial count into evidence

A county's partial count becomes one number, the reading: its counted margin minus its expected margin, in points. A positive reading means the Democrats are doing better than we expected there. Each reading comes with a noise: how far it could differ from the true miss for reasons that tell us nothing about anywhere else.

\(y = m_{\text{counted}} - \text{expected}_c, \qquad \sigma^2 = \sigma_{\text{own}}^2 + \sigma_{\text{order}}^2 (1 - s)^2 + \sigma_{\text{sample}}^2\)

Here \(s\) is the share of the county's expected vote counted so far. The noise has three parts:

• The county's own swing, about 4 points. However complete the count, a county can move for local reasons.
• Count order, 14 points with nothing counted, shrinking to zero as the county finishes. This is the mail-first or election-day-first skew. Where we know which kind of vote a count is, and how that kind voted in 2024, the reading is taken against that kind's own baseline and this part falls to about a third of its size.
• Sampling, which matters only for the first few thousand votes. A first count of 2,000 votes is a few precincts, not the county.

Counties are pooled race by race. The counties in one statewide race are combined into one reading for that race: the margin of all the votes counted so far, against the margin we expected of the votes we expected to be in by now. Comparing it that way means a county that turns out more than usual moves the reading too, as it moves the result. Our first version compared each county with its own expectation and averaged the gaps, which misses turnout; on the 2026 counts that left every state's reading about 1.7 points short of its real gap, because Democratic counties turned out more than usual. The race's own error (its candidates, its campaign) is counted once for the pool, not once per county, because ten Texas counties all share the same Texas candidates. The count-order skew is also treated as shared: when one county counts mail first, its neighbors usually do too, so averaging many counties does not make it go away. Our 2024 replay showed why this matters. Treated as independent, the ranges were far too narrow and held the true result less than half the time where mail is counted first.

A worked example, with typical values. A large Texas county runs 4 points better for the Democrats than expected:

What has been counted	Noise (points)	Share of the reading passed to the Texas race
20%, mixed kinds of vote	12.3	19%
20%, known to be mail, against 2024's mail margin	6.4	48%
100%	5.0	59%
30 counties, all fully counted	3.1	79%

The share passed on is the variance the race's shared errors explain (about 36.6 points squared for a Texas statewide race) divided by that plus the noise squared. So the first row moves the Texas race by about 0.8 points, and the last by about 3.2. The share never reaches 100%, because the race's own error stays: no number of counties in one state can tell us whether Texas's Democratic candidate is simply stronger than the polls said, rather than the polls missing everywhere.

Counties that finish early. Many counties, often small rural ones, have counted every vote within an hour or two of polls closing. They give the forecast its strongest early readings. With the whole county in there is no count-order skew left, so the noise is only the county's own swing and the race's (5.0 points in the example above, against 12.3 for a county a fifth counted), and about 60% of what it shows passes to its state's races. A county is treated as complete when its count reaches the vote we expected there or the results say it is complete; one that has reported its election-day votes but not its mail ballots is read as partial.

The counties that finish first are not a fair sample of their state, and the forecast allows for that. Their share of the voter-group errors comes from their own make-up, so a run of mostly white, rural counties running better for the Republicans moves races in places like them most and suburban races less. Taken at face value, early counties mislead: the swing in each state's smallest fifth of counties missed the state's overall swing by 10.2 points on average between 2012 and 2016, 4.8 between 2016 and 2020 and 1.9 between 2020 and 2024. They say a lot about rural voters and less about a state as a whole.

How one county moves every race

Before any vote is counted, each shared error is believed to be zero, give or take its typical size. Every reading says something about a weighted sum of them: its loadings times the errors. Putting all the readings together, the most likely value of every shared error, and how sure we are of it, come out of one calculation. This is standard Bayesian updating of a normal prior, written in its information form:

\(\Lambda = I + \sum_i \frac{h_i h_i^{\top}}{\sigma_i^2}, \qquad \eta = \sum_i \frac{h_i\, y_i}{\sigma_i^2}, \qquad \hat{\theta} = \Lambda^{-1} \eta, \qquad \operatorname{cov}(\theta) = \Lambda^{-1}\)

Here \(h\) is a reading's list of loadings, \(y\) its value and \(\sigma\) its noise; \(\theta\) is the list of 64 shared errors, measured in units of their typical size. Any race's expected move is then its own loadings times the estimated errors, \(h \cdot \hat\theta\), with an uncertainty of \(\sqrt{h^\top \Lambda^{-1} h}\).

Who moves, and by how much. The split between national, regional and state errors is not set by hand: it follows from the loadings. A reading from Texas loads on the national error, the West South Central region, Texas and the voter groups. With one reading the forecast cannot tell which of these is responsible, so it shares the move among them in proportion to their size. Taking the worked example's county, 20% counted and 4 points better for the Democrats:

Race	What it shares with the Texas county	Typical move
Texas statewide	everything but the county's own swing	about 0.8 points
A Texas House district	the same, with half the state part	a little less
Oklahoma	national, region, voter groups	about 0.3 points
Ohio	national, voter groups	about 0.2 points

Why more counties make it sharper. Two things happen as counts arrive:

• Readings that agree add up. Each one adds its information to \(\Lambda\), so the uncertainty in every shared error shrinks. A dozen states all running 2 points better for the Democrats pin down a national miss that no single state could.
• Readings that disagree separate the errors. Texas running ahead of expectations and Ohio behind cannot both be the national error, so the forecast assigns each to its own state. Counties with different make-ups (a heavily Hispanic county and a mostly white rural one in the same state, say) separate the voter-group errors from the regional ones.

This is the idea at the heart of the method: a small sample from a single county says a little about every race, and each county added says more, until the shared errors are known well enough that races with nothing counted are forecast mainly from the votes elsewhere. A county's evidence is also passed to the House districts that lie in it.

Projecting each race's final result

For a race with votes of its own counted, the final margin is built county by county: votes already counted stay as counted, and each county's remaining votes are projected.

The race's swing so far. Each county's counted margin is compared with its expectation. The differences are averaged, weighted by votes counted times the share of the county counted, so a county a tenth of the way in counts for much less than a finished one. The average is then shrunk toward zero according to how sure the pre-election forecast was. A race the polls were confident about needs more evidence to move:

\(\delta = \bar{g}\, \frac{\sigma_{\text{forecast}}^2}{\sigma_{\text{forecast}}^2 + \sigma_{\text{obs}}^2}, \qquad \sigma_{\text{obs}}^2 = \tau^2 \frac{\sum w^2}{(\sum w)^2} + \left(10 (1 - \bar{f})\right)^2\)

Here \(\bar g\) is the weighted average gap, \(w\) the weights, \(\tau\) = 6 points is how far one county's swing typically strays from its race's, and the last term is the count-order skew shared by the race's counties: 10 points with nothing counted, shrinking as the counted counties fill up (\(\bar f\) is their average share counted).

Each county's remaining votes are projected at its expected margin, plus the race's swing \(\delta\), plus part of how far the county itself is running from that:

\(m_{\text{rest}} = \text{expected}_c + \delta + k \left( m_c - \text{expected}_c - \delta \right), \qquad k = \frac{\tau^2}{\tau^2 + \left(14 (1 - f_c)\right)^2}\)

The share \(k\) is small when only a sliver of the county is in, because the first votes are usually one kind and say little about the rest. It is 22% with a fifth counted, 42% at half and 95% at nine-tenths. So a county whose first mail count runs 20 points ahead of expectations is not assumed to stay 20 points ahead, but a county 90% counted that is running ahead almost certainly will.

The projected final margin is the counted margin in votes plus the projected margin on the remaining votes, divided by the total expected. A county already past its expected vote is assumed to have a little more still to come (3%), since late ballots keep arriving for days. Where a race comes in only as a statewide total, with no county figures, the vote still to come is taken from the source's own estimate of the share counted (the Associated Press's, for most races), not from our expected turnout.

The range around it combines five parts, each scaled by how much is left to count:

Part	What it covers
The race's swing	how sure we are of \(\delta\), times the share of the vote still to come
Each county's own swing	\(\tau\) = 6 points on each county's remaining votes, less the part already seen
Turnout	each county's remaining vote is uncertain by about a third of itself; this matters where that county leans differently from the race
Count order	the skew on the votes counted in counties that are only partly in
A floor	at least half a point until every vote is in, and a quarter of a point once every precinct is: provisional and cured ballots, the canvass and the estimate of the vote still to come can all be off, so a 0.6-point lead with 99% in is about a 90% chance, not a call

The chance of each side winning comes from the projected margin and its range, assuming the error is normally distributed.

Every candidate's expected result. Alongside the chance, the forecast gives every candidate's expected final vote and share: the final total is the count so far plus the projected remainder, candidates outside the two main sides keep their share of the count, and the two main sides share the rest so that their margin is the projection, each with an 80% range. On 7 November 2026, with 96% of Mississippi's Senate vote counted and Scott Colom 1,936 votes ahead, it expected about 991,700 votes in all: Colom 485,470 (49.0%) and Cindy Hyde-Smith 485,345 (48.9%), a 51% chance for Colom.

Races decided by a runoff or ranked choices. Where a party has several candidates on one ballot and no one has half the vote (Louisiana's House seats go to a runoff; Alaska and Maine count ranked choices), the race is read as all the Democrats' votes against all the Republicans'. So the candidate leading the count can be the one less likely to win: in Alaska's Senate race Mary Peltola led the first round, but with a second Republican's votes expected to pass to Dan Sullivan, the forecast gave Sullivan a 60% chance.

Races counted only as a statewide total. Where a state publishes no county figures, the race is read as one unit and the vote still to come is taken from the source's own estimate of the share counted (the Associated Press's, for most races), not from our expected turnout. Our first version gave such a race a quarter of its count still to come once the count passed our expected turnout. On election night that turned a 0.2-point Democratic lead in Mississippi, with 96% counted, into a projected Republican win.

Races with nothing counted yet take their move from the shared errors, leaving out the race's own readings so no count is used twice. A simpler carried swing is also applied: half the vote-weighted gap between counted races' projections and our forecast, phased in as more races are counted.

Control of the House and Senate

The chances of control come from the same simulation as our pre-election forecast: 5,000 simulated elections, each drawing its own set of shared errors and its own race-by-race luck. On the night two things change:

• The shared errors are drawn around what the counts say. Instead of being centered on zero, they are drawn from the estimate above, with its narrower uncertainty. As counts come in, the simulated elections agree with each other more and more about how the polls missed.
• Races with votes counted are drawn from their live projection, not from the pre-election forecast.

Because the shared errors are drawn once per simulated election and applied to every race in it, races keep moving together on the night, just as they do before it. If the counts so far suggest the polls underrated the Republicans by 2 points, every uncounted race leans that way in every simulated election, and the chance of a sweep of the close races rises accordingly.

The simulation uses the same random draws as the published forecast, so with no votes counted the election-night forecast equals the pre-election forecast exactly. Any difference you see on the night comes from votes.

Guarding against bad data

A single bad number can do a lot of damage to a forecast that lets every count move every race. A county reported twice, a typo in a feed, or a precinct's votes filed under the wrong county could, unchecked, move the whole map.

Each reading is checked against all the others. Before a reading is used, the forecast asks what every other reading predicts for that county. A reading far outside that prediction is down-weighted rather than trusted in full. The weighting follows a Student-t distribution with 4 degrees of freedom, which behaves like the normal distribution for readings close to expectation but gives far less weight to extreme ones, and it is repeated a few times until it settles. In a rehearsal before the 2026 midterms, a single mismatched Vermont count took the chance of Senate control we showed from 91% to 29%. With the check in place, the same count barely moves it.

Counts that cannot be placed in a county are added to the race's total as counted, but do not become readings. When one source's total is stored over another's county figures, the county figures are matched to the candidates by name; where that cannot be checked and the race is close, they are set aside rather than risk reading them the wrong way round. On 7 November 2026 such a mix-up had Minnesota's Senate race, 10 points apart in the count, projected as level.

People stay in charge. The forecast gives chances. It never calls a race. A race is called only by a person, once polls have closed there and the forecast's chance has passed a high threshold.

Testing it on the 2024 count

The best test would be a real election night, replayed from the counts as they actually arrived. Those are hard to come by: the live feeds are not kept in public archives. So we replay the night from the final 2024 presidential result instead:

• The forecast is made to miss. Before the count, each state's forecast is set to the true 2024 result plus a typical forecast error (drawn at random, with a spread of 3.5 points), and each county's expectation follows from it and from the county's 2020 result. Expected turnout is the true vote with a 6% error.
• Counties report in a realistic order. They start and finish at random paces. In the seven states where we hold each county's 2024 vote by method, each county counts its mail ballots first, then early votes, then election-day votes, then provisional ballots, so the running count drifts as real counts do. Elsewhere each county's count is an even mix of its voters.
• The forecast reads the night. At 30, 50, 70 and 90% of the state's vote counted, the counties counted so far are read as evidence and every county's remaining vote is projected, exactly as on a real night.

Counted	Projection, error	Running total, error	Forecast alone, error	80% range held the result	Typical range
30%	3.0	6.1	2.8	85%	±6.4
50%	1.7	4.0	2.8	82%	±3.0
70%	1.2	2.4	2.8	81%	±1.7
90%	0.5	0.9	2.8	82%	±0.8

Average error in points of margin, states where mail is counted first, 84 replays.

At every stage the projection beats the running total by a wide margin. From half counted on, it also clearly beats the pre-election forecast, which is the point: the votes take over from the polls. At 30% counted, when counties are reporting mail first, the projection is no better than the forecast alone, because such a skewed early count tells the forecast little, and it rightly trusts it little.

The 80% ranges are about right where mail is counted first, the closer copy of a real night. In the even-mix states they are too cautious: they held the result 92 to 99% of the time, roughly twice as wide as they need to be. We have left them that way on purpose. No real county counts a perfectly mixed sample of its voters, so the even-mix replay flatters the forecast, and a range that is a little too wide is a smaller failing on election night than one that is too narrow.

Ideas we tested and did not adopt

Lopsided and small counties. A county that votes 85-15 has less room to swing than a 50-50 one, and small counties swing more for their own reasons. On county presidential swings from 2012 to 2024 both were true: allowing for lopsidedness cut the error in the most lopsided counties by about a fifth. But built into the forecast and replayed on 2016, 2020 and 2024 from the election before, the projected final margin of each state was no better (the average error changed by less than a tenth of a point at every stage), and early estimates from small counties got slightly worse. Lopsided counties are mostly small and carry little of a state's vote, and where they swung unusually, as rural counties did in 2016 and 2020, the voter-group errors already account for it.

Projecting each county's remaining vote by its own make-up. Instead of one swing for a whole race, each county's remaining vote could move by its own share of white, graduate and Hispanic voters. Where counties count an even mix of ballots this was much better early in the count (off by 1.2 points instead of 1.8 with a tenth counted). Where they count mail first it was badly worse, because the mail lean runs with the same make-ups and was read as a swing among those voters. Versions that guarded against that did no better overall than the forecast as it stands, so it keeps its caution about count order.

Limits, and what would make it better

• Not knowing what kind of vote a count is. A county's first 20% may be all mail ballots. Read as a mix, that looks like a Democratic surge. The forecast allows for this, but such a reading is weak until the county is more than half in. Where the results say which kind of vote a batch is, the reading is far stronger.
• A wrong baseline. A county whose expected vote is badly off (rapid growth, a boundary change, a collapse in turnout) reads as a swing that is not there. The check against the other readings catches a single odd county, but not a whole state with an out-of-date baseline.
• A shift the shared errors cannot describe. A swing among a group the forecast does not model, younger voters for example, is spread among the nearest ones and may show up in the wrong places.
• Independent and third-party candidates. Where the main challenger is an independent, the race is treated as moving like the Democrat would, with extra uncertainty. That is an approximation.
• The first few counts. With a handful of counties in, a national move rests on one or two states. The chances move, but most of the pre-election uncertainty is still there. Calls do not follow from it.

What would make it better, in order of value: the kind of vote in every count from every state (the forecast already uses it where a state's results split the count, but many do not); kind-of-vote baselines from more than one election, so that the range can safely narrow when the kind is known; precinct baselines, which turn each partial county count into many small like-with-like comparisons; more voter groups, such as age and how urban a place is; and a real night to test against, which keeping snapshots of the 2026 counts as they arrived will give us.

How this compares with the TV networks' decision desks

The Associated Press and the major networks each run a decision desk: a team of statisticians and political analysts who decide when a race can be called. Their methods differ from ours in three main ways.

• They have survey data on the night. The networks commission exit polls and the AP runs its own large survey of voters. Those give them an estimate of each state's result at the moment polls close, before any votes are counted. We start from the pre-election polls instead.
• They work at the precinct level. Their models compare each reporting precinct, or a chosen sample of precincts, with its own past results, and they receive the count in finer detail than a public results page shows. A precinct is small and usually one kind of place, so each comparison is cleaner than a partly counted county.
• People make the call. A model's estimate is one input. Analysts also weigh which places are still out, how each state counts its votes and how far the count has to go, and they wait until polls have closed in a state.

What our forecast does that a single race call does not is tie every race together. A county counted in Texas moves our chances in Oklahoma, in Ohio and in the fight for control of the Senate, all at once and all through the same shared errors used in our pre-election forecast. The New York Times's live "needle", first shown in 2016, is the public tool closest to ours: it sets an expected result for every place before the night, compares each count with it and simulates the possible outcomes, but it works precinct by precinct where we work county by county.

Try it yourself

Pick a race and a county, say how much of its vote has been counted and enter the totals so far, and see what that one count does to the Senate, the House and every race in the country: test our election-night forecast.

Glossary

• Margin. The Democrat's share of the vote minus the Republican's, in percentage points. Positive means the Democrat leads.
• Expected margin. What we expect a race or a county to show before any votes are counted, from our pre-election forecast.
• Reading. A count's margin minus its expected margin: how far a place is running from what we expected.
• Noise. How far a reading could be from the true miss for reasons that say nothing about anywhere else: local swings, the order of the count, and chance.
• Shared error. A hidden amount by which the polls miss many races at once: nationally, in a region, in a state or among a group of voters.
• Loading. How many points one unit of a shared error moves a race or a county.
• Count order. The skew in a partial count caused by one kind of vote (mail, early or election day) being counted before the others.
• Projection. Our estimate of a race's final margin once every vote is counted.
• 80% range. The range we expect to hold the final margin four times out of five.
• Simulation. Running the election thousands of times with different draws of the shared errors and each race's own luck, to give the chance of each outcome.

How to cite

Polling Forecast (2026). How our election-night forecast works: one county's count and every race. Retrieved 9 November 2026 from https://www.pollingforecast.com/research/us-election-night-forecast. Live figures change as new polls arrive; the date above is when the text was last revised.