Glossary
Watch the video on YouTube · All videos
What the words on this site mean here, with the numbers we actually use.
- Comparing forecast ranges
Can I compare all forecast uncertainty ranges directly?
Check the quantity (margins, seats or alignment bounds) and, for margins and seats, the coverage level.
Compare intervals using both their coverage level and the quantity they describe. Senate, governor and House race pages show 90% ranges for margins, except that a Senate or governor race forecast without polls shows an 80% range; state-legislature district pages show 80% margin ranges. Current chamber seat-count ranges use 80% coverage. Older state-legislature forecasts can show a labeled 90% seat range. For the same forecast, a higher coverage level includes more simulated outcomes. None of these intervals is a candidate's chance of winning; Senate alignment bounds describe different assumptions and are not coverage intervals.
Sources: Comparing forecast ranges · Forecast methods
- Election-day forecast
What is an Election Day forecast?
Our estimate of the result on November 3, 2026, allowing for what can still change.
Combines polls (for statewide races; for a House district, its own polls where there are any, weighed against the model's estimate), each district's or state's fundamentals (past results, demographics, economy, candidates, money) and the national environment, and widens the uncertainty for the time left before the election. This is the headline number on every page.
Sources: Election-day forecast · Forecast methods
- Nowcast
How does a nowcast differ from an Election Day forecast?
An estimate of where a race stands today, from its polls and, where polls are few, a starting estimate from past results. Not a prediction of November.
For a polled statewide race it is based on adjusted polling evidence: the polls, corrected for each pollster's measured lean and, where there are fewer than four, combined with the starting estimate from past results. Chamber nowcasts also include estimates based on past results and other non-poll data for seats without usable race polling; those seats are listed beside the total. The House has no nowcast, only an election-day forecast. A nowcast answers a different question from the election-day forecast, so the two are never averaged or shown as one number.
Sources: Nowcast · Forecast methods
- Tipping-point race
The race that delivers the seat that decides control, counted across the forecast's simulated elections.
In each simulated election the side that wins the chamber is given its seats from safest to narrowest; the seat that brings it to a majority decided control. A race's tipping-point share is how often it was that seat. It is not the same as being close: a close race in a chamber already decided rarely decides anything. The Senate and House pages list the ten highest shares; the same simulations drive You decide.
- Chance (win probability)
What does a 70% chance of winning mean?
How likely a side is to win: for the House and the chamber totals, the share of tens of thousands of simulated elections it wins; for a Senate or governor race, computed directly from the forecast's range of margins.
A 70% chance means the side wins about 7 times in 10 elections like this one -- and loses 3. It is not a vote share and not a margin. Over many races, things we call 70% should happen about 70% of the time; the Accuracy page checks that against past elections.
Sources: Chance (win probability) · Forecast methods
- 80% range
The span the result falls inside in 8 of 10 simulations.
One time in ten the result comes in above the range and one time in ten below it. A wide range means a genuinely uncertain outcome, not a vague model.
- Safe / Likely / Lean / Toss-up
What do Safe, Likely, Lean and Toss-up mean?
Labels for the favored side's chance, as shown to the nearest whole percent: Safe 95% or more, Likely 85-94%, Lean 60-84%, Toss-up below 60%.
The rating is a translation of the forecast's chance into words, computed the same way for every race; no one assigns it by judgment. Lean R means the forecast gives Republicans a chance of 60% to 84%, and so on. The cutoffs apply to the chance as it is shown, rounded to a whole percent, so a race at 94.6% shows as 95% and is rated Safe, and one at 84.6% shows as 85% and is rated Likely; a rating and the percent beside it never disagree.
Sources: Safe / Likely / Lean / Toss-up · Forecast methods
- Competitive / close
What makes an election competitive in FutureBallot?
A race or chamber whose favored side's chance, as shown, is below 85%: rated Lean or Toss-up.
One definition wherever the site says competitive or close without giving numbers: neither side's chance is 85% or more (the ratings above). Toss-up is narrower: the favored side's chance is below 60%. A page or article that needs a different band prints its numbers and its reason where it uses them.
Sources: Competitive / close · Forecast methods
- Research model
A forecast published before a person outside the project has checked its model; read its numbers as provisional.
A version loses this label when it is recorded as the standard model. For the House model that waits for a check by a person outside the project; the checks so far were made by AI model sessions. The House model's seat totals are tested on 2016-2024 (see Accuracy, which names the version on the site); while the outside check is outstanding the House forecast carries this label, and the House page says so beside the headline.
- Range for independents
Why can an independent create a range for Senate control?
Shown when an independent could win and FutureBallot has not established a firm commitment to caucus with either party.
Senate control depends on which party each senator organizes with. Where no firm caucus commitment has been found (Nebraska, Idaho, South Dakota and Montana in 2026; the Senate page gives the sources and search dates), the low end assumes the independent does not side with Democrats and the high end that they do. We do not guess a person's choice.
Sources: Range for independents · Forecast methods
- Simple average
How is a simple polling average calculated?
The plain mean margin of a race's polls in the 30 days up to its newest poll.
Every poll counts equally, whoever ran it or paid for it. Shown so the adjustment in the adjusted average is visible.
Sources: Simple average · Forecast methods
- Adjusted average
What is an adjusted polling average?
The polls weighted by recency and size and corrected for pollster lean.
This is the polling input to the nowcast. For a Senate or governor race the nowcast also stretches the lead by the factor described on the methodology page and, with one to three polls, combines the average with the starting estimate from past results. Each firm's lean is estimated relative to the other firms polling the same race, shrunk toward zero and centred so the firms' leans average zero; a lean shared by every firm cannot be seen this way and is carried as extra uncertainty, not corrected. Statewide, a poll paid for by a party or campaign gets no separate discount: the poll's lean shows only through its pollster's estimated lean.
Sources: Adjusted average · Forecast methods
- Partisan sponsor
What is a partisan-sponsored poll?
A poll paid for by a party, campaign or party-aligned group.
Such polls are published more often when they are good news for the sponsor, so on average they lean toward them. They are kept, shown and flagged. In House districts the forecast removes the measured lean of each sponsor class (Democratic, Republican, nonpartisan); statewide they are corrected only through their pollster's measured lean.
Sources: Partisan sponsor · Forecast methods
- National environment
How the country as a whole is leaning this year, in points.
Estimated from the generic congressional ballot, presidential approval and the economy, and checked against past midterms. It moves every race together, which is why seats are not independent coin flips.
- Model change
What does a model change mean for forecast comparisons?
A move in a forecast that includes a change in our method.
When the method changes, a race's forecast can move without new information, so part of the move may not be news from the race. Those moves are listed separately so they are not mistaken for news, and the changes of method are dated on the change log.
Sources: Model change · Forecast methods
- Ground-up estimate
A forecast of a Senate or governor race built without its polls.
A model fitted on past elections predicts the state's margin from its past results, demographics and economy, the candidates' incumbency, experience and money, and the national environment. The election-day forecast averages it with the poll-based forecast, each weighted by how accurate it has been; a race with no polls is forecast from it alone. The ground-up estimate is made only for a race between a Democrat and a Republican.
- Lean (of a district)
How much more Democratic or Republican a district votes than the country.
The district's margin minus the national margin. The House model predicts each district's lean and adds the national environment to it.
- Latent opinion
The underlying support in a race that the polls measure with error.
The trend model treats each poll as a noisy reading of one line of support that moves a little each day, and estimates that line, its trend, from all the polls together.
- House effect
How far a pollster's results lean, on average, compared with the other firms polling the same race.
Measured for each pollster in each race and removed before its polls are combined, so one firm leaning differently does not tilt the average; a firm with few polls has its lean held close to zero. A house effect is relative: it cannot correct a bias that every firm shares.
- Design effect
How much a poll's real sampling error differs from what its sample size alone suggests.
A poll's sampling error is multiplied by a factor measured on past polls, because weighting and the way a sample is drawn change how precise it really is.
- Generic ballot
A national poll question asking which party's candidate a voter would choose for the House, without naming the candidates.
Its average, corrected by its average past miss, is the main input to the national environment.
- Heavy-tailed (Student-t)
An error distribution in which large misses are more likely than under a bell curve.
The forecasts draw their error terms from a Student-t distribution rather than a normal one, because polling misses as large as some past ones would be near-impossible under a bell curve. The fewer its degrees of freedom, the heavier its tails.
- Walk-forward testing
Reconstructing forecasts for past dates, then scoring them against election results.
The information cutoff aims to exclude records not yet public. Historical reconstructions have documented exceptions: some economic series were revised later, and missing poll publication dates are inferred from fieldwork. Accuracy describes these limits; retrospective tests are distinct from a prospective live record.
- Brier score
The average squared gap between a forecast's chance and what happened.
What happened counts as one for a win and zero for a loss, so a 70% chance that comes true scores 0.09 and one that does not scores 0.49. Lower is better.
- Seat error
How far a forecast's expected seat total is from the seats actually won.
For the House, the gap between the Democrats' expected seat total and the seats they won, whichever way it falls, averaged over the elections tested. Lower is better.
- Log loss
A score of chances that punishes a confident wrong forecast heavily.
The negative logarithm of the chance the forecast gave to what happened, averaged over forecasts. A wrong forecast made with near certainty scores very badly. Lower is better.
- Interval coverage
How often results fall inside a forecast's range.
A 90% range should hold the result about nine times in ten; coverage well below that means the ranges are too narrow, well above it that they are wider than needed.
Election questions
How is a chance of winning different from a predicted vote share?
A win probability describes how likely a candidate or party is to win. A vote share describes the fraction of votes a candidate receives. A 70% chance of winning does not mean a forecast of 70% of the vote.
Sources: Win probability