Every adjustment and its evidence
each step between the raw polls and a published number, grouped by forecast, and the steps tested and not used
Every step between the raw polls and a published number is listed below: what it does, the values it uses, and the evidence for it. The standard evidence is a walk-forward comparison, over the past elections in the record, of forecasts made with the step and without it. Where a step cannot be switched off, or the comparison cannot score it, the entry says so and gives the evidence it does have.
Terms used on this page
- Walk-forward: each past election's forecast is reconstructed at a historical cutoff and scored against that election's result. Revised economic series and inferred poll-release dates can differ from the information actually available then; see the historical-data limitations.
- Log loss: a score of win probabilities that punishes a confident wrong call heavily; lower is better.
- Brier score: the average squared gap between a win probability and what happened (one for a win, zero for a loss); lower is better.
- Mean absolute error: how many points the forecast margin missed the result by, ignoring direction.
- Coverage: how often the result falls inside a forecast's range; a 90% range should hold the result about 90% of the time.
- Consistency correction: a change that makes the model do what it is specified to do; such a change is kept unless it makes forecasts measurably worse, beyond tolerances set before the test.
Senate and governor
- Which polls a forecast reads. A forecast reads only the polls that had been recorded and published by the moment it stands in. A historical poll without a recorded release date counts as released two days after its fieldwork ends. Only a poll's headline question is read, and a withdrawn or superseded poll is skipped. A poll stored in several versions counts once, in the first version stored. A firm's overlapping tracking waves of one question count once: the newest wave is kept, along with any wave that does not overlap a wave already kept.
Values: Release lag for a historical poll without a release date: two days after fieldwork ends. Each run records the rule it used.
Evidence: A rule about what could have been known, not an adjustment that can be switched off: reading a poll before its release would be leakage. The leakage tests plant a poll after the forecast's date and check that no forecast changes.
- What a statewide margin measures. Senate and governor forecasts, and their scores, measure the gap between the two leading candidates as shares of all votes cast, third parties included, the way results are reported. Polls enter with their undecided respondents included, and the scale law (below) carries the resulting margins onto that scale. House forecasts measure the margin on the two-party scale, the Democratic and Republican vote only, except a seat forecast from polls of its own contest, which uses that contest's two-candidate margin. House polls enter at their published margins, undecided respondents included, with constant corrections and no fitted slope.
Evidence: The definition of what is forecast, not a step that can be switched off. The definition's one alternative, sharing out undecided respondents in proportion before averaging, is less accurate on 748 past races.
- Campaign- and party-sponsored polls. A poll that a campaign, party committee, PAC or interest group pays an independent firm to conduct is kept; in Senate and governor races the sponsor's lean is left to the pollster's house effect. The in-house polling operations of campaigns and party committees (nine firms) are excluded by name.
Values: Sponsored polls kept; the excluded firms are listed, with their reasons, on the pollsters page.
Evidence: Statewide, leaving sponsored polls out improves pooled log loss and Brier score slightly but worsens the margin error; dropping sponsored polls is worse on log loss in three of five elections, in 2016 by more than the limit set in advance, and worse on Brier score in four of five. For the House, in the 2020 and 2022 poll-layer test, leaving sponsored polls out is less accurate on lean error and overall Brier score, though more accurate on the close-district Brier score; leaving them out also makes the seat totals less accurate over 2018-2024.
- Pollster house effects. Each pollster in a race gets its own lean, its house effect, estimated from that race's polls. The leans are constrained to average zero across the race's polls, each firm weighted by how many polls it released, so the trend tracks the average poll rather than any one firm. A lean every firm shares is not corrected here; the shared national error covers that shared lean.
Values: Size of house effects: half-normal prior with scale 2.5 points; weights are poll counts.
Evidence: Mixed on the 2006-2024 record: on the eve of the election forecasts are more accurate with it, and four weeks out they are slightly less accurate. Pooling house effects across races instead is less accurate.
- Design effect on poll sampling error. Every poll's sampling variance is multiplied by a design effect, the factor by which real polls are more or less precise than a simple random sample of the same size. The factor is measured from earlier elections' polls by the poll-dispersion law. An earlier cycle with nothing measured before it uses 1.6.
Values: 0.757 for 2026, measured on 2018-2024; 1.6 where nothing is measured.
Evidence: Tested on the 2016-2024 record: forecasts are slightly more accurate with the measured value than with a fixed 1.6, and no election is worse by more than 0.0002 in log loss.
- Fundamentals prior on the race's starting level. The polling model's starting level for a race gets an informative prior: the state's previous result, for the same seat where possible, with a width set by how much such margins move between elections. The width for a given number of years between elections for the same seat is measured only when results from at least eight different states support it; otherwise the width for the nearest such interval that has eight or more states is used, scaled by the square root of the ratio between the two intervals.
Values: Volatility from elections before the target only; at least eight pairs of results from at least eight states per gap length.
Evidence: Tested on 284 past races: forecasts are more accurate with it in each of five elections. The eight-state minimum leaves accuracy unchanged on the 2016-2024 record.
- Width of the statewide prior, net of the national swing. The width of the starting prior, how far a seat's margin moves between elections, is measured after the national swing between those elections is taken out, as the chamber simulation measures it, so the swing is not counted twice. The prior is about 8% narrower than the raw width.
Evidence: Tested on the 2016-2024 record as a consistency correction: accuracy and range coverage are unchanged within the set tolerances.
- Statewide prior: the previous result net of its year's national mood. The starting prior of a Senate or governor race is the seat's previous result, and that result carries the national mood of the year it was won in. That year's national level is taken out and this year's estimated level put in, both read from one fit of earlier Senate (or governor) results on the national House vote, built only from elections before the one forecast. For 2026 this moves every Senate and governor prior toward the Democrats, Senate priors by 3.45 points on average and governor priors by 6.07, before the polls and the ground-up blend. The House is not changed.
Values: 2026: Senate level -1.47 + 0.787 × House margin; governor level -4.97 + 0.635 × House margin; national House environment +6.95 (Democratic).
Evidence: Tested on the 2016-2024 record: the published forecast's accuracy is unchanged within the set tolerances. The polling model alone is less accurate with it; the ground-up blend absorbs the difference.
- Results read at their November round in nominee runoffs. For contests decided in a runoff between party nominees (Georgia's Senate races of 2008, 2020 and 2022), the statewide data and the polling model's starting prior read the November round rather than the runoff. All-party runoffs keep their runoff.
Evidence: A correction of stored data, tested on the 2016-2024 record: better on every pooled measure for the statewide data, and unchanged within tolerance for the starting prior.
- Four polls to fit a trend. A race with four or more usable polls is fitted by the polling model. A race with one to three polls goes to the sparse-race blend. A race with none is not a poll forecast.
Values: Minimum four polls.
Evidence: Not tested on past elections as a threshold: below three polls the trend cannot be fitted at all. A minimum of three passes its own test, and the sparse-race blend now covers those races.
- Sparse-race blend. A race with one to three polls is forecast by combining the fundamentals prior with the average of its polls, weighting the prior and the poll average by their precision, rather than from the prior alone.
Values: Prior mean and spread; the poll average's spread as measured on past sparse races.
Evidence: Tested on the 2018-2024 record: forecasts are more accurate with it in every past election scored, both on the eve of the election and four weeks out. Such a race then takes the ground-up estimate as well, as every race does, so it counts a prior twice. Counting the prior once lowers the 2016-2024 log loss slightly (0.16400 against 0.16415) and the mean absolute error (4.56 against 4.57 points), but makes the Brier score slightly worse, and on the sparse races the 90% ranges, which already held too few results, hold fewer still (0.784 against 0.843). The change passes the rule registered before the change's test, but it is held back, to be made together with a correction of those ranges' width.
- Measured error of a sparse race's poll average. In the sparse-race blend, a handful of polls is weighted by how far such averages have actually landed from results in earlier elections, not by their sampling error alone. The national and race-specific parts of that error are taken out before the blend, because the forecast adds both afterwards.
Values: 9.37 points for 2026, pooled over earlier elections; sampling error alone would be about four points.
Evidence: On sixty-seven past sparse races, using the measured error lowers the mean absolute error from 8.01 to 6.34 points and brings the 90% ranges close to their nominal coverage (0.925, from 0.791); the version using the measured error has worse log loss than the version without it. Across the full 2006-2024 backtest, the 80% and 90% ranges together miss their nominal coverage by 0.044 with the measured error against 0.020 without it on election eve, and by 0.065 against 0.046 four weeks out. Counting the national and race-specific parts once improves log loss (0.1630 against 0.1644) and Brier score (0.0504 against 0.0507) on 2016-2024 and is worse in no election, but one condition of the rule set before its test, on 90% range coverage, tied (0.949 both) and so was not met. The rule was amended after the results to use it. In these races the prior's error and the polls' error overlap (a correlation of 0.41 on the sparse races of 2020-2024); a version of the blend that allows for the overlap is scored against the published one on the 2026 results by a rule set before them.
- Shared national polling error. Every race's election-day spread includes an error shared by all races in an election, the size of past elections' average polling misses. The shared national error is fitted as a variance component over earlier elections, against the round of each result that the polls measured.
Values: 2.63 points for 2026.
Evidence: Fitted on results from 1998 through 2025. With the current fit, the published forecast is as accurate on the 2016-2024 record as with the earlier fit, and slightly better by log loss.
- Race-specific error. Every race's spread includes an error particular to that race, beyond its polls: the residual of the same variance-components fit as the national error.
Values: 4.93 points for 2026.
Evidence: Without it, the 90% ranges held 68% of results on 247 past races; with it they hold 90%.
- Drift to Election Day. A forecast made before Election Day is widened by how far races have historically moved between that point and the vote. The added variance is a × h^b, where h is the number of days left.
Values: 2026: a = 1.650, b = 0.669; about 3.9 points of standard deviation four weeks out.
Evidence: Fitted on the movement of past races' polling averages. The standard test scores the drift law four weeks out on 358 past races; the current coefficients are scored only on the eve of the election, where the term is near zero.
- Heavy-tailed election-day errors. The non-poll error terms are drawn from a Student-t distribution with five degrees of freedom rather than a normal, so large misses are more likely than a bell curve allows. A race's total error adds these to the polling model's own uncertainty, so the total is heavier-tailed than a normal but is not itself a Student-t with five degrees of freedom.
Values: Five degrees of freedom (a constant, not fitted).
Evidence: On the 2006-2024 record, forecasts are more accurate with it on average at both horizons, though not in every election; the value five is not fitted.
- Scale law (reference entry). The scale law is fitted on earlier elections' final polling averages against results, with an intercept plus a slope times the polling average. Only the slope is applied (the next entry); the intercept, a uniform shift, is not applied. This entry describes the fit as a whole; the fit's intercept and slope are judged separately.
Values: 2026: intercept -1.744, slope 1.070.
Evidence: As a whole, mixed on the 2006-2024 record: slightly less accurate on the eve of the election, slightly more accurate four weeks out.
- Scale law: slope. Moves the centre of every forecast away from a tie by the fitted slope, because final polling averages have been more compressed than results; the spread is not changed. For 2026 a 10-point lead becomes 10.7.
Values: Slope 1.070 for 2026 (fitted together with the intercept; a slope fitted without one would be 1.048).
Evidence: Tested on the 2006-2024 record: forecasts are more accurate with it in eight of ten past elections, both on the eve of the election and four weeks out.
- Ground-up blend (the published Senate and governor forecast). The election-day centre of every Senate or governor race between a Democrat and a Republican is averaged with a ground-up prediction, weighted by the precision of each. The ground-up prediction is a gradient-boosted model, a set of decision trees fitted in sequence, of the state's lean from its results, demographics, economy, candidates and money, plus the national House environment. The spread is kept, and the win probability is read from the same moved distribution. The blend first takes the scale law's intercept back out of the polling model's forecast.
Values: Prior variance from walk-forward residuals. The ground-up estimate keeps about a fifth of the weight against the polls, more where there are few polls.
Evidence: Tested on 2016-2022 and confirmed on 2024: forecasts are more accurate with it (Brier score 0.0585 against 0.0677 on 175 races), with most of the gain in 2020.
- Ground-up prior: spread by size of predicted margin. The ground-up prior's variance in the blend is measured separately for races predicted under ten points, ten to under twenty points, and twenty points or more, rather than as one pooled variance.
Values: Three classes: under ten, ten to under twenty, and twenty or more points.
Evidence: Tested on the 2016-2024 record: the prior alone is slightly more accurate, and its 90% ranges stay close to nominal coverage in every class.
- Ground-up prior on the all-vote scale. The ground-up prior is a two-party margin. Before the two-party ground-up prior is blended with the polling forecast on the all-vote scale, the prior is multiplied by the share of the state's previous general election for the office won by all Democratic and Republican candidates plus any other candidate who finished first or second, counting each candidate at the round where that candidate's share was highest.
Evidence: Tested on the 2016-2024 record as a consistency correction: accuracy is unchanged within the set tolerances (mean absolute error 0.05 points worse).
- Unpolled races published from the ground-up prior. A Senate or governor race between a Democrat and a Republican with no poll at all is published from the ground-up prior alone, with the prior's full measured error as its spread, and is labelled "fundamentals only, no polls".
Values: Spread: the prior's walk-forward error.
Evidence: Not tested on past elections as a switch: without it the race is unpublished, not forecast differently. The prior alone missed by 8.3 points on average on 2016-2022 and by 6.7 in 2024, and none of the eighty-seven seats the prior favoured by twenty points or more flipped.
- Senate and governor: scandal rule. After a verified scandal about a candidate themselves (a criminal charge or conviction, a specific misconduct allegation or official finding, or a documented false claim about their own record, reported by two outlets doing their own reporting), the race's margin moves 3.0 points against that candidate from the event's first date, only while the race has fewer than three polls taken after the event. Polls taken after the event already show most of its effect, so with three or more of them the margin is the polls' and the blend's alone. The spread is unchanged. A running mate's, relative's or staffer's trouble does not count. Events are listed by hand with their sources, and races forecast without polls take the shift too. One event is listed: the Colorado governor race (Victor Marx). Borderline cases are listed with the evidence that would make them count and are checked daily.
Values: 3.0 points, while the race has fewer than three polls whose fieldwork started after the event's first date; none after that. The events are listed by hand with their sources.
Evidence: Provisional. On seventeen past statewide scandal races the shift scores better on every measure, but the interval for its accuracy gain includes zero and the gain rests largely on one race (Missouri's 2012 Senate race). The statewide measurement was set up to report the result, not as a test to pass, so the rule has no statewide pass; it is used on the House test of the same shift, which passes on twelve House cases. On the eve of past elections, the polls taken after an event already carried most of its effect, which is why the shift stops at three of them. Both scandal shifts are scored again on the 2026 results.
- Alaska's ranked-choice races: the shape of the error. Alaska's Senate and governor races are forecast on their final ranked-choice round between the two strongest candidates. The polling part of each forecast's error is drawn with the heavy-tailed shape the other Senate and governor races use for their Election Day errors, and the error from ballots that transfer or run out differently from what respondents said is combined with it as a separate, independent error. The centre of the forecast does not depend on the shape.
Values: A Student-t shape with five degrees of freedom for the polling error; 5.05 points for the transfer error.
Evidence: With the same data, the heavy-tailed shape moves the Democratic candidates' Election Day chances by less than half a point: Mary Peltola's in the Senate race from 56.86% to 57.18%, and Jonathan Kreiss-Tomkins's in the governor race from 88.94% to 89.38%. The shape is used for consistency with the rest of the statewide model, under a rule set before the comparison that accepts such a change only when no chance moves by two points or more and no range end by half a point or more; the three past ranked-choice contests measure the transfer error alone. The two Alaska forecasts are compared with their certified 2026 results for description only; if either candidate in a forecast's final round is eliminated before that round, the contest is reported rather than scored.
House
- House: the corrected national environment. The House model's national environment is a corrected estimate (N3): the generic-ballot average with its mean past miss removed, blended with a fundamentals regression, rather than the generic ballot as it stands.
Values: Read from today's generic-ballot average.
Evidence: Tested on 2004-2024 nationally: the corrected estimate misses the national House vote by 2.16 points on average, against 3.01 for the raw generic ballot. Not tested on its own at the district level.
- Generic-ballot bias correction. Subtracts the generic ballot's average past miss before it enters the national environment, in every year. The same national term is in the ground-up prior of the Senate and governor forecasts, so the correction reaches them too. It rests on its own evidence, not on the evidence about the statewide polls' intercept, which is never applied: it is fitted on the generic ballot against the national House vote, and its miss in midterms alone is the same size as its pooled miss.
Values: 2.62 points toward the Democrat on average, 1998-2024; 2.70 points toward the Democrat in seven midterms.
Evidence: Not tested on its own under the standard test. Neither no correction nor an alternative of the same kind does better on the elections still available for testing.
- House: fundamentals in the national environment. Blends the bias-corrected generic ballot with a regression on the midterm penalty, presidential approval, income growth and the previous House vote, each weighted by its past error.
Values: Regression coefficients refitted on 1960 to two years before the election.
Evidence: Tested on 2004-2024 nationally: the blend misses the national House vote by 2.16 points on average, against 2.44 for the corrected generic ballot alone. Not tested on its own at the district level.
- House: national uncertainty. Every simulated House election draws one national environment around the corrected estimate. The national environment's spread is the national estimator's walk-forward error at the end of the campaign, plus the movement still to come and the extra measurement error of today's generic-ballot average beyond what past end-of-campaign averages had.
Values: End-of-campaign error 2.75 points, before the movement to come and today's measurement error are added.
Evidence: Part of the House model's tested package; not tested on its own.
- House: scandal rule. After a verified scandal about a House nominee (a criminal charge or conviction, a specific misconduct allegation, or a documented false claim about their own record, reported by two outlets doing their own reporting), the district's margin moves 3.0 points against that nominee from the event's first date; the spread is unchanged. Events are listed by hand with their sources. Party splits are not part of the rule.
Values: 3.0 points, persistent; the events are listed by hand with their sources.
Evidence: Tested on 2018-2024, all twelve scandal candidates finished below the forecast, and the shift makes forecasts more accurate. The House scandal shift is scored again on the 2026 results.
- House: shared state error. Districts in the same state share an error, measured from walk-forward residuals about each election's mean. The election's own lean is drawn with the national error that every district shares. State and district errors are drawn from a Student-t distribution with four degrees of freedom.
Values: Walk-forward residuals from 2014 to two years before the election.
Evidence: Tested on the House record: accuracy and range coverage are unchanged within the set tolerances.
- House: district error by race type. A district's own error is sized by its type: an incumbent on the same lines, an open seat, or new lines.
Values: Incumbent on the same lines 7.25 points, open seat 7.86, new lines 8.11.
Evidence: Part of the House model's tested package; not tested on its own.
- House: district response to the national swing. The district model takes the national swing as an input, so each district's response to a wave is learned rather than uniform.
Values: Learned by the district model.
Evidence: Part of the House model's tested package; not tested on its own.
- House: seats without a two-party contest. A district without both a Democrat and a Republican on the ballot is given, without simulation, to the one major party with a nominee on record. A district where neither side has a nominee on record, or where both sides do, is simulated from its lean. A seat forecast from its own polls (Alaska at-large, California's 6th) is never given away this way.
Evidence: Not tested on past elections: the House test assigned these seats from their actual winners, so there is no forecast to score.
- House: district poll update. Moves a polled district's lean toward its own polls: each pollster's latest poll, corrected for its sponsor class's measured lean, combined with the model's estimate by precision, with an error shared across the election.
Values: Sponsor-class leans measured on 2018-2024 results.
Evidence: Tested on 2020, 2022 and 2024 with polls as of four weeks out: lean error, Brier score and seat error are all better than the model alone. Reasonable variants of the sponsor correction move the House total by up to about six seats.
- House: rater blend. Moves each rated district's lean part of the way toward what the average rating of the professional race raters has meant in past elections, by a weight fitted on past elections.
Values: Weight about 15%, refitted on each House run; category offsets from 2018-2024.
Evidence: On 2018-2024 it improves the district probabilities (Brier score 0.0331 against 0.0353) but makes the seat-total error worse (11.76 seats against 11.18, all of it from 2022). The rule set before the test required both, so the blend fails it. The rule was amended after the results, to accept the higher seat error for the better district probabilities, and the cost is stated on the Accuracy page. The 2022 cost has a known cause: a rating is a call about the margin in its own year, and the blend reads it as an average lean pooled over years. A second form that reads each rating as a margin in its own year is scored against the published blend, and against no blend, on the 2026 House result by a rule set before that result.
- House: the three core raters. The rater average uses Cook, Sabato's Crystal Ball and Inside Elections only.
Values: Cook, Sabato's Crystal Ball, Inside Elections.
Evidence: Tested on 2018-2024 against averaging every rater: slightly more accurate (Brier score 0.0329 against 0.0331; seat error 11.66 against 11.76).
- House: district polls weighed against the blended lean. When a district has polls, they are combined with the rater-blended lean using that blended lean's own measured error.
Values: Prior variance from the blended model's walk-forward residuals.
Evidence: Tested on past House elections: mean seat error 12.16 against 13.47 for the blend alone, and Brier score 0.112 against 0.121.
- House: district polls discounted by age. Each district poll counts for less the further it was taken from Election Day: full weight within sixty days, falling steadily to half at 210 days, and a quarter beyond. An older poll never counts for more than a newer one. The schedule is a judgement that polls taken further from Election Day should count less, not a fitted value; polls within sixty days count equally.
Values: Weight one up to sixty days, falling linearly to 0.5 at 210 days, 0.25 beyond (fixed).
Evidence: The rule set before the backtest says to use the age discount unless it is clearly harmful; the discount does not show the clear improvement the other House changes must show. Tested on the 2016-2024 record, the age discount was not clearly harmful compared with equal weighting (squared error 0.58 lower, 95% interval -2.70 to +1.10), and the discount was better at 120 days out, at ninety days out and at sixty days out, and neutral in the final month. Measured on district polls of 2004-2024, a poll's error does grow with its age, but by less than the schedule assumes: as a weight against a poll taken on the eve, about 0.79 at 120 days and 0.68 at 210 days, where the schedule gives 0.80 and 0.50. Weights from that measurement are scored against the schedule, and against equal weights, on the 2026 House result by a rule set before that result.
- House: spread of the seats forecast from their own polls. Two House seats are forecast from polls of their own two candidates, because neither is a contest between a Democrat and a Republican: Alaska's at-large seat (an independent against a Republican, decided by ranked choice) and California's sixth district (a Democrat against a candidate with no party preference). Their spread is how far such poll averages have landed from results in earlier House elections, drawn with the same heavy-tailed shape as the rest of the House model. For the ranked-choice seat, the error from ballots that transfer or run out differently from what respondents said is added.
Values: 8.41 points, measured on 644 earlier polled House races; a Student-t shape with four degrees of freedom.
Evidence: On 406 polled House races of 2018-2024, each scored with its own election held out, the 90% ranges cover 0.901 of results and the 80% ranges 0.768, both within the limits set before the test, and the log loss and Brier score are better than with the House model's full error, whose 90% ranges covered 0.983. The test used races already used to choose the size, so it is a second look rather than a fresh one. The spread is scored again on the 2026 results.
- House and statewide estimate: one religion census for every election. The share of each area's residents in each religious group comes from the same edition of the U.S. Religion Census, the 2010 study, for every election from 2012 to 2026, so the House model and the ground-up estimate for Senate and governor races learn and predict on one measurement. The 2020 study counts some groups on another basis, so mixing the two editions would read a change in how the census counts as a change in the electorate.
Values: The 2010 study for every election.
Evidence: On 2024, the one past election where the choice of edition matters, the House and statewide forecasts score the same with either edition, within limits set before the comparison (House Brier score 0.01946 with the 2010 study for every election, against 0.01982), so the consistent input is used.
- Training money dated by when each filing was received. The House model and the ground-up estimate for Senate and governor races learn from past campaign money as it stood before each election: outside, party and PAC spending counts only if the Federal Election Commission had received the filing that reports it by the cycle's cutoff. Filings from before mid-2015 carry no date in their identifying number, so each is dated by the receipt date the Commission records for its filing number, and a filing the Commission does not list is not counted.
Values: Filings received by the twentieth of September of the election year count; spending dated before October of that year.
Evidence: Dating the 2012 and 2014 filings this way lowers the Democratic-side outside money counted for those years from about $35 million each to $0.6 million and $2.2 million. On past elections the House record improves (Brier score 0.0362 against 0.0368, seat error 11.9 against 12.2) and the Senate and governor record is no worse, within limits set before the comparison.
Chambers
- Chamber simulation: correlated errors. Senate and governor control probabilities draw one shared national error, with regional and office terms, per simulated election across all races. The shared national error is drawn wider than on a race page because the error's size is itself estimated from a limited number of past elections. The shared national error's distribution is a Student-t whose scale is σ_nat, the shared national error's standard deviation, times the square root of (one plus one divided by n), with degrees of freedom equal to n minus one, where n is the number of elections. The shared national error's tail weight is estimated from past elections; each race's own error is a Student-t with five degrees of freedom. The model's measured lean is shown as a sensitivity beside the headline and never applied to it.
Values: Estimated from history at run time. For 2026, seventeen elections give a national error about 10% wider in standard deviation, and the parameter that sets how often extreme misses occur is close to a normal distribution.
Evidence: Not testable by the per-election rule: one chamber outcome per election gives five to ten scored points. The chamber model passes its own test, and race-level numbers are unaffected.
- Chamber simulations keep each race's published forecast. After the Senate and governor control simulations draw their joint margins, each race's draws are mapped by rank onto that race's own published election-day distribution. The dependence between races is kept, and a race's chance inside the chamber equals its chance on its own page. Seats with no stored distribution, filled from fundamentals, keep their draws.
Evidence: Tested at each election eve 2016-2024: every race within one point of its published chance, the chamber test passed, and the log score slightly better, in both chambers.
National environment
- Generic ballot: each pollster's lean removed before averaging. Before the national generic-ballot polls are averaged, each firm's lean is measured against the other firms' polls taken at the same time, shrunk toward zero when the firm has few polls, and centred so the typical firm counts as zero. A firm that polls every week at a lean then no longer pulls the average toward itself. The industry-wide miss is left to the generic-ballot bias correction.
Values: Shrinkage k/(k plus four) for a firm with k polls (a firm with four polls keeps half its lean); consensus over the sixty days before each poll, that firm left out.
Evidence: Tested on 2004-2024: the corrected average's squared error is lower (7.84 against 8.82), and it is better in eight of eleven elections a month out and on the eve.
Tested and not used
These are not part of any published forecast. Each is listed with the evidence against it, so a reader can see what the forecast does not do.
- Likely-voter adjustment. No adjustment moves one sample population toward another: likely-voter, registered-voter and adult samples enter alike. When a release reports several populations, one version is kept: likely voters, else registered voters, else adults. A firm's typical lean, including any that comes from the population it samples, is absorbed by its house effect.
Evidence: A likely-voter adjustment, which moves registered-voter and adult polls by the gap measured where a poll is reported both ways, was worse on log loss but better on margin error on the 2020-2024 record, and it improved log loss in only one of three elections, so it is not applied.
- Scale law: intercept. The scale law's intercept, a uniform shift of every Senate and governor forecast toward one party (1.744 points toward the Republican for 2026), is never applied, in any year.
Evidence: In the election-eve backtests of 2016-2024 (234 races), the forecasts without it are more accurate taken together: log loss 0.1691 against 0.1733 and Brier score 0.0524 against 0.0546, with a slightly larger average miss (4.36 points against 4.22). By kind of year, leaving it out helps in midterms and costs in presidential years; the rule is the same in both.
- Mean polling-bias correction. Subtracting the average past polling miss from every forecast is not done.
Evidence: Out of sample, its gain is within noise, and it fails its log-loss test three times; the same question is answered by the scale law's intercept, which is also not applied.
- Range widening and fading shifts after major events. Widening a race's range after a confirmed severe event, and a centre shift that fades as new polls arrive, are not applied. The persistent scandal shift for Senate and governor races is a separate entry.
Evidence: Both fail their own tests on past events.
- Combined sponsor and firm records. A combined record such as "NBC News/Marist" is read as its own pollster, not pooled with the firm that fielded it.
Evidence: Reading such records as their fielding firm makes Senate and governor forecasts slightly worse on 2006-2024 and changes nothing in the House.
- Turnout and enthusiasm. No forecast estimates who will turn out in November or how enthusiastic each side is. Turnout enters only indirectly: through the polls, whose likely-voter screens are the pollsters' own turnout models; through past results, demographics and the national environment; and through an error band measured on past elections whose turnouts varied.
Evidence: Two turnout-based House models are less accurate than the published one or no better than noise. Both count Hispanic voters inside other groups too, so they rule out those two designs, not turnout modelling in general.
Every dated change to the method, with its evidence, is on the change log: How the model has changed.