Get FutureBallot by email. Emails when FutureBallot publishes a new explainer or video, when each state's results are certified, when it reports how its forecasts did, and when the 2028 forecast begins.

Accuracy

How the 2026 forecast is doing, and how the model did in past elections: 258 scored races

2026, the first live election

Of 488 races, four are scored so far. A race is scored once its state certifies the count, the results page shows it certified, and it can be matched to the stored result rows; the rest are still being counted, awaiting certification, or held back.

Next certification due: Arkansas by November 18 (today).

The registered scoring is published once at least 95% of each office's sealed races are certified; until then this page is not that scoring.

Start here

2026 is the first live election. The historical scores are reconstructions from election cycles also studied during model design. Races in one cycle share errors; they are not independent validation trials. The table below shows exactly which deployed components each evidence record covers.

What has been tested · Current limitations and corrections

Forecaster scorecard: each forecaster compared, FutureBallot included, race by race on certified counts (four races scored so far).

How much independent evidence?

Number of election cycles scored: six; equal-election mean Brier score: 0.0429. Descriptive 95% election-cluster bootstrap range: 0.0219–0.0637. Whole elections, including their shared errors, are resampled together. The equal-election mean differs from a pooled race-weighted score. With few cycles, any such interval is descriptive, sensitive to the cycles available and not a clean prospective validation.

How much independent evidence?
ElectionScored racesBrier score
2016440.0851
2018680.0522
2020390.0448
2022630.0477
2024420.0272
202620.0004

Evidence by deployed component

Inspection, code availability and independent reproduction are different. The federal models are private; their explanations and outputs can be inspected. The state-legislature code is published. Full federal input bundles are not public. Versioned output manifest and reproduction boundaries · Available review materials.

The acceptance column distinguishes original criteria from amendments and limited validation. Every adjustment, original test rule and decision. A later amendment does not erase an original failure.

Evidence by deployed component
ComponentDeployed family/versionEvaluated populationAcceptance statusLimits
Statewide races
Deployed versionsingle_race_latent/0.8.7
Reconstructed Senate and governor elections, 2016–2024; shared election errors.Mixed original passes and retained amendments; first live election in 2026.Private code. Revised economic series and inferred historical poll release dates.
House districts
Deployed versionhouse_ground_up/2.13.0; house_district_prior/0.14.0
Historical district and seat-count evaluations; later components have separate records.Base gate plus amendments; the rater blend failed its original seat-error condition.The historical aggregate does not validate every live poll treatment or special contest.
Special candidate pairs
Deployed versionhouse_district_prior/0.14.0
Own-poll error components have historical tests; complete unusual candidate-pair forecasts do not.Limited validation; not an original pass for the complete special-case forecast.AK-AL and CA-6 use different estimators; their pages describe each actual specification.
Ranked-choice races
Deployed versionranked_choice_final_round/1.2.0
Separate final-round evaluation; not the standard two-candidate statewide score.Read the separate registered evaluation and current race's method.The available historical final-round sample limits generalization.
Chamber simulations
Deployed versionsenate_chamber/0.4.2; governor_chamber/0.4.1; house_ground_up/2.13.0
Component tests of shared errors, marginals and seat outcomes.Component evidence; historical race scores are not independent chamber-control trials.Few election cycles; independent alignment bounds do not model organizing votes.
State legislatures
Deployed versionstateleg_swing/0.11
Eligible regular single-member districts; qualifying chambers' tested subsets.Initial version retained despite an interval-coverage acceptance failure.Redistricting years, odd years and several district formats are outside that evaluation.
Ask
Deployed versiongemma-4-26b-a4b-it-maas; routing and training snapshot 959efd5e888a1959
Held-out question-classification evaluations, separate from election forecasting.Released with documented missed acceptance targets.Check the interpreted question; ambiguity should prompt clarification.
You decide
Deployed versionsenate run dde88370-2173-4f20-9590-9ee0167a59f6; house run d9446e9c-ba14-45cf-87d5-66e79157518f
Baseline reproduction and scenario arithmetic checks.Illustrative shifts and conditional simulations; no validated causal intervention model.Controls can overlap; calling races reduces the retained simulation sample.

Reading different ranges

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.

The record in brief

For the current statewide forecast model, the scored 2016–2026 record covers 258 governor and Senate races. Its forecasts missed the final vote margin by 4.53 points on average, ignoring direction. On the 258 forecasts now on file (2016–2026) where a recency- and size-weighted polling average could also be computed, the polling average's average absolute miss, ignoring direction, was 5.31 points, and the model's was 4.53 points. Its 90% ranges contained the actual result in 92.6% of forecasts. A well-calibrated 90% range contains it about 90% of the time, so these ranges have been somewhat wider than needed.

In the current statewide model's scored 2016–2026 record (258 scored forecasts), forecasts of the vote margin leaned 2.31 points toward Democrats on average. The current year's forecasts in it are scored against results certified since; scoring a result does not change the forecast. The lean was Democratic in each of its 6 scored cycles.

These are reconstructed historical forecasts, not forecasts published before those elections. Each cycle was forecast from data dated before it, with two known exceptions: state economic series are used in their later-revised form, and a historical poll with no recorded release date is assumed released two days after fieldwork ended. The model's design was also developed while studying these same elections. Races in the same election share polling errors, so six election cycles provide much less independent evidence than 258 separate tests would. 2026 is FutureBallot's first live election.

Technical details: run and model identifiers

Model single_race_latent/0.8.7, cycles 2016, 2018, 2020, 2022, 2024, 2026.

How to read the record

  • N is the number of forecasts scored: each race's election-eve forecast, one per race. Races in the same election share its polling error, so the elections, not the races, are the evidence (five of them).
  • Average error (mean absolute error) is how many points the forecast vote margin missed the actual margin by, ignoring direction.
  • Lean is the average signed error, in points of margin: positive means the forecast overstated the Democrat's margin, negative the Republican's.
  • 50%, 80% and 90% range columns show how often the actual result fell inside the forecast's 50%, 80% and 90% ranges. A well-calibrated 90% range contains the result about 90% of the time.
  • A ✗ marks a range column that is more than 5 percentage points away from its target in either direction: for the 90% range, below 85% or above 95%. That is an operational tolerance for flagging, not a statistical test: with five elections, whether a difference is noise depends on how many forecasts and elections stand behind it. On coverage, the test that decides whether a model change is made uses a different rule from this mark: among its other conditions, the promotion test refuses a change when a range covers less than its target by more than its tolerance. A range that covers more than its target is too wide rather than wrong, and the promotion test charges for it through the score instead.
  • Brier score and log loss score the win probabilities against who actually won; lower is better. Log loss punishes a confident wrong call more heavily.
  • AUC measures how well a forecast ranks races, so that likelier winners are ranked above likelier losers (1.0 is perfect, 0.5 is a coin toss). It does not measure whether a stated 70% chance comes true 70% of the time.
  • The head-to-head comparison with 538 covers one election (2018) and sixty-seven races; the model version it was run with is in the technical details under its heading. FutureBallot's Brier scores are lower against every 538 version shown; after correcting for the three comparisons only the difference against 538 lite is statistically distinguishable from chance. The model was developed with this election in view (it has been evaluated by many model versions), so this is not evidence that FutureBallot is more accurate.

Every figure in this record is walk-forward: each cycle was forecast from what was estimable before it (with the two data-vintage exceptions above), and scored against the certified result. The model was never fitted to a cycle's own results, but its design was chosen while studying these same past cycles, so none of them is a clean test: the first clean test of the current models is November 2026, scored from forecasts published before it. Read the record with that in mind. The same past elections have also been used to test many proposed changes (most written down before they were run; the list is below), and every test run on them weakens what they can show. The test that promotes a model change treats each election as one unit, so five elections can give only weak evidence for any change. The statewide record is scored without the scandal rule; the statewide record, scored with the rule as published, gives the same values on all seven summary figures compared here (average error, lean, the shares of results inside the 50%, 80% and 90% ranges, the Brier score and the number of wrong calls, not including log loss) as it does without the rule. The rule moves a race's election-day margin 3.0 points against a candidate after a verified scandal about the candidate (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), but only while the poll count since the event is below three. Of this record's races, nine have an event before election eve in the list of past events used to test the rule, a list mixing events the rule counts with events it does not. In seven of them, three or more polls were taken after the event, so no shift applies whatever the event's kind; the other two are an illness (Minnesota Senate, 2020) and a debate remark (Illinois Senate, 2016), kinds of event the rule does not count. Scored with those two moved as well, the lean is 2.33 points toward the Democrats instead of 2.31, the share of results inside the 50% ranges is 47.3% instead of 46.9% and the share of results inside the 80% ranges is 82.4% instead of 82.8%; the average error, share of results inside the 90% ranges and number of wrong calls are unchanged. The rule's own test, on seventeen past races, is on the adjustments page. Where a historical statewide poll's release date is not recorded, the record assumes a release two days after its fieldwork ended, though most polls were released one to a few days after fieldwork (median two days). Re-fitting using the actual release dates found for 97.5% of the polls that matter raises the average error by about 0.05 points and changes three of 225 winner calls (measured on an earlier race population; the analysis is not published on this site). The current statewide model's national economic figures are those first published before each past election (with income before 2010 from the 2010 release); only the state economic series are used in their later-revised form.

The record

Forecast
ForecastNAvg. errorLean50% range80% range90% range
This model2584.53+2.3146.9%82.9%92.6%
A recency- and size-weighted polling average, same forecasts2585.31+2.86———

On the same 258 forecasts, the model misses by 0.78 points less than the polls it is built from, and leans 2.31 points toward the Democrats where the polling average leans 2.86. Computed November 17, 2026 from the same scored forecasts.

Brier score 0.0511, log loss 0.1622 on the same 258 forecasts.

Against the people who do this for a living

Two numbers, because they mean different things. The wrong-call count covers only races a rater actually called — a rater who rates more races as toss-ups has fewer chances to be wrong, so the column is not comparable across raters on its own. AUC uses the full ordinal ranking including the toss-ups. For this comparison the model's call is whichever side the model gives a chance above 50%, so the model calls every race, including those the site rates Toss-up (the favored side's chance below 60%). Each race is scored as the Republican against the other candidate, so a rating for an independent, such as Angus King in Maine, is a call on that candidate. The model version each comparison was run with is in its technical details.

Senate races

The races: 114 of the 124 Senate races of 2018–2024 in the record above. Left out: eight of the record's races, all in a state that held two Senate contests in the same year (Minnesota (2018), Mississippi (2018), Georgia (2020), Oklahoma (2022) and Nebraska (2024)), whose ratings cannot be matched to one contest by state name; and two contests between two candidates of one party, California (2018) and Alaska (2022), where a rating for a party does not say which of them wins. On the 114 races compared the model's favorite lost in six; the wrong-call list below has eight Senate races from 2018–2024, the other two being Georgia (2020) and Alaska (2022), outside this set. Raters who rated fewer years are compared on the same races in the years they rated.

Senate: Comparison with race raters
RaterCyclesRater wrongRated toss-upModel right on thoseRater AUCModel AUCVerdict · model − rater [95% CI], p corrected for every rater tested
Cook Political Report2018–20240 / 922216 / 220.98140.9898tie · +0.008 [-0.003, +0.020], corrected p 1.000
5382018–20245 / 10775 / 70.98910.9898tie · +0.001 [-0.008, +0.008], corrected p 1.000
Inside Elections2018–20244 / 10775 / 70.98880.9898tie · +0.001 [-0.004, +0.006], corrected p 1.000
RealClearPolitics2018–20240 / 833125 / 310.96370.9898model · +0.026 [+0.009, +0.047], corrected p 0.014
Sabato's Crystal Ball2018–20246 / 11400 / 00.98420.9898tie · +0.006 [-0.002, +0.016], corrected p 1.000
Fox News2018, 2022, 20240 / 73139 / 130.98890.9883tie · -0.001 [-0.013, +0.010], corrected p 1.000
Decision Desk HQ2020, 2022, 20242 / 7875 / 70.98890.9917tie · +0.003 [-0.006, +0.011], corrected p 1.000
Politico2018, 2020, 20220 / 701510 / 150.98440.9883tie · +0.004 [-0.007, +0.016], corrected p 1.000
Daily Kos2018, 20200 / 4895 / 90.98880.9912tie · +0.003 [-0.009, +0.016], corrected p 1.000
The Economist2020, 20223 / 5333 / 30.98140.9910tie · +0.010 [+0.000, +0.028], corrected p 1.000
CNN20180 / 2364 / 60.97780.9833tie · +0.006 [-0.037, +0.050], corrected p 1.000
Election Daily20242 / 2900 / 00.99040.9952tie · +0.005 [-0.007, +0.026], corrected p 1.000
New York Times20180 / 2097 / 90.94440.9833tie · +0.039 [-0.003, +0.105], corrected p 1.000
CBS News20220 / 2354 / 50.98981.0000tie · +0.010 [+0.000, +0.040], corrected p 1.000
Technical details: run and model identifiers

Model single_race_latent/0.8.7.

Governor races

The races: all eighty-eight governor races of 2018–2024 in the record above. On them the model's favorite lost in six, the same six governor races of 2018–2024 as in the wrong-call list below. Raters who rated fewer years are compared on the same races in the years they rated.

Governor: Comparison with race raters
RaterCyclesRater wrongRated toss-upModel right on thoseRater AUCModel AUCVerdict · model − rater [95% CI], p corrected for every rater tested
Cook Political Report2018–20240 / 691913 / 190.97650.9911tie · +0.015 [+0.002, +0.033], corrected p 0.189
Inside Elections2018–20242 / 78106 / 100.98480.9911tie · +0.006 [-0.004, +0.021], corrected p 1.000
RealClearPolitics2018–20240 / 632519 / 250.95970.9911model · +0.031 [+0.011, +0.058], corrected p 0.018
Sabato's Crystal Ball2018–20246 / 8711 / 10.97720.9911tie · +0.014 [+0.003, +0.030], corrected p 0.112
Politico2018, 2020, 20220 / 63148 / 140.98370.9885tie · +0.005 [-0.006, +0.018], corrected p 1.000
5382018, 20222 / 6163 / 60.98750.9848tie · -0.003 [-0.012, +0.005], corrected p 1.000
Fox News2018, 20220 / 53148 / 140.97860.9848tie · +0.006 [-0.008, +0.024], corrected p 1.000
Daily Kos2018, 20200 / 3785 / 80.98380.9858tie · +0.002 [-0.013, +0.020], corrected p 1.000
Election Daily2022, 20242 / 4300 / 00.99570.9978tie · +0.002 [-0.003, +0.011], corrected p 1.000

Ratings are read from Wikipedia's transcription of each rater's published table, not from the rater. Where the two disagree this is the one that is wrong. Concurrent same-state contests and special elections are dropped rather than joined by state name alone. A rating is ordinal and is never converted to a probability, so no Brier or log loss is reported against a rater. Each call is used as a cited fact, under its rater's name, and Wikipedia's compilation is CC BY-SA 4.0; the basis for this use is on the Data sources page (DATA_SOURCES §2b).

Technical details: run and model identifiers

Model single_race_latent/0.8.7.

The House model, election by election

The House model, shown election by election, is fitted only on earlier elections and scored on each of 2016–2024. The figures are district win probabilities (Brier score, lower is better), the districts whose favorite lost, the share of districts whose result fell inside their 90% range (a 90% range should hold the result nine times in ten), and the Democratic seat total against its 90% range. The district ranges' width is set from earlier scored elections only; 2016, the first election in this record, uses the unadjusted width. The published model's scandal rule moves a district's margin 3.0 points against the nominee a listed scandal is about, however many polls follow it. The scoring records the past-scandal list's earliest and latest listed cycles as 2018–2024; that span alone does not mean every cycle inside it is covered. The rule moves a district's margin in twelve recorded House district contests, and no district in 2016 is moved because the list does not cover that cycle. Scored without district polls. The published model adds changes this record does not include: California's 6th District forecast from its own polls; district polls discounted by their age; and Alaska's at-large seat and California's 6th District given a range taken from how far averages of a House contest's own polls have missed the result in past elections, allowing for an occasional larger miss.

House model record by election
ElectionBrierDistricts whose favorite lostDistrict 90% ranges holding the resultDemocratic seats, expectedActual90% rangeRange held it
20160.02801298%199194182–218yes
20180.04442489%217235202–236yes
20200.03511989%243222227–260no
20220.03081898%202213187–217yes
20240.02291298%206215190–220yes
Technical details: run and model identifiers

Record of house_ground_up/2.13.0; published house_ground_up/2.13.0; scored 2026-11-02 by scripts/score_house_published.py.

House districts and the raters

From 2016 to 2024 the House model's favorite was the opposite of the raters' election-eve average of at least Lean in only 7 races, and the raters' side won all 7; that is too few races to test against chance. Measured with a test written before it was run, within each election year, in 123 competitive House districts with polls in 2020–2024 the result moved about 74% of the way from the poll-based estimate toward the raters' view (95% range 53–91%): the raters know something about close districts that the polls miss. Each rated district's lean is moved part of the way toward what the raters' average rating has meant in past elections: 15% of the way in the current fit (17% in the 2024 test), a share fitted only on earlier elections. Tested on 2016–2024, it made the district probabilities more accurate in 2018, 2020 and 2024 and less accurate in 2022. It made the House seat total less accurate, by 0.6 seats on average over 2016–2022. The test written before the blend was run required both a lower Brier score and a lower seat error; the blend lowers the Brier score but raises the seat error, so the blend fails that requirement. The rule was amended after the results were seen, to accept the higher seat error for the better district probabilities, and the blend is in use under the amended rule. That share is the same in every district. A version with a much larger share for close races fails its test, which was written before the larger-share blend was run: following the raters that far also follows them when they miss together, and in 2022 the larger-share version misses the seat total by twenty-six seats, where the published blend misses it by nine. Two limits: the average is over Cook, Sabato and Inside Elections only, as in the history, but the weight was fitted on their election-eve ratings while today's input is their current view, weeks out; and the test scored the blend without district polls, so the published model (blend plus polls) is tested as a combination separately: with polls, Brier 0.1115 against 0.1214 without on the polled districts of 2020–2024, seat error 12.2 against 13.5.

Head to head with 538, 2018

Technical details: run and model identifiers

Model single_race_latent/0.8.7.

Head to head with 538
VersionN538 BrierModel Brier538 wrongModel wrongp (corrected)
538 lite670.07150.0531750.03
538 classic670.06690.0531750.09
538 deluxe670.06130.0531550.23

Read the p column before the Brier column. The p values are corrected for the three comparisons (Holm) and come from resampling whole states. Only the difference against 538 lite is statistically distinguishable, on one election of 67 races. That election has been used to evaluate many versions of this model, so a lower score on it is not out-of-sample evidence, and reporting the Brier difference as a win would claim more than the data supports.

2018 is the only cycle 538 published a per-race review for, and it is one of this model's better cycles (low error, almost no lean), so one cycle flatters it. Cook and RealClearPolitics are excluded by this project's source rules; Sabato and Silver Bulletin have no registered license-clean bulk source. 538 'deluxe' includes Cook's and Sabato's ratings as inputs, which is the nearest available comparison to them.

Where the error is

By office

By office
By officeNAvg. errorLean50% range80% range90% range
governor1014.44+1.6549.5%84.2%93.1%
senate1574.59+2.7445.2%82.2%92.4%

By region

By region
By regionNAvg. errorLean50% range80% range90% range
MW674.77+3.6846.3%77.6%92.5%
NE494.70+1.6340.8% ✗81.6%87.8%
S755.05+3.3641.3% ✗84.0%90.7%
W673.59+0.2758.2% ✗88.1% ✗98.5% ✗

The regional spread is mostly polling density and the national lean, not geography: the fitted regional term is 0.00 and cannot detect a regional sd below about 2 points.

By how much the race was polled

By how much the race was polled
By how much the race was polledNAvg. errorLean50% range80% range90% range
1-4526.97+3.6340.4% ✗73.1% ✗84.6% ✗
4-10584.59+3.0351.7%84.5%93.1%
10-30953.75+1.5149.5%83.2%94.7%
30 or more533.48+1.6843.4% ✗90.6% ✗96.2% ✗

The sparse column is where the model is weakest. Races with one to three polls go to the sparse blend because the polling model is fitted only where a race has four or more usable polls, a threshold that is not itself tested (a minimum of three passes its own test). Those races' poll averages are weighted against the fundamentals prior in inverse proportion to their measured error, because such averages have missed results by far more than their sample sizes imply. National generic-ballot polls taken within three days of each other by different firms, once each firm's house effect is taken out, scatter around each other a little less than their sample sizes imply (a design effect below one), so the larger miss in sparse races includes an error that every poll of a race shares.

By how close the race was

By how close the race was
By how close the race wasNAvg. errorLean50% range80% range90% range
0-3383.88+0.5339.5% ✗86.8% ✗97.4% ✗
3-8433.53+2.0044.2% ✗88.4% ✗95.3% ✗
8-20934.03+2.0252.7%82.8%95.7% ✗
20+845.89+3.6045.2%78.6%85.7%

Error and lean are both largest where the race is least close. The obvious reading — that the forecast is compressed toward the middle — is wrong and was tested: compression would make the lean negative where the Democrat is ahead, and it is positive in seven of the eight cells of that analysis (margin band by favored party, not the tables above). What this column is showing is the partisan gradient described below, seen through a column that happens to correlate with it (the analysis is not published on this site).

By cycle

By cycle
By cycleNAvg. errorLean50% range80% range90% range
2016445.28+3.8236.4% ✗79.5%86.4%
2018683.23+0.7660.3% ✗92.6% ✗100.0% ✗
2020396.30+5.5320.5% ✗64.1% ✗84.6% ✗
2022634.84+0.5346.0%76.2%90.5%
2024423.79+2.9361.9% ✗97.6% ✗97.6% ✗
202623.58+2.4850.0%100.0% ✗100.0% ✗

Each row counts the forecasts scored by this page's model, one per race. The scorecard counts every certified race that any registered model scored, the House model's included; the House model also scores House districts that this page's model never scores, so the scorecard's 2026 figure is higher than this table's, and part of that difference remains however many statewide races this model scores.

What is still wrong with it

This section is not an appendix: a track record that reports only its successes is selective. Each item below says what it was measured on.

  • It has leaned toward the Democrats in every scored election. In the current statewide model's scored 2016–2026 record (258 scored forecasts), forecasts of the vote margin leaned 2.31 points toward Democrats on average. The current year's forecasts in it are scored against results certified since; scoring a result does not change the forecast. The lean was Democratic in each of its 6 scored cycles. The Senate and governor forecasts apply no correction for that measured lean. In three evaluations registered in advance, on 2018-2024 races with earlier versions of the statewide model, a forecast centered on the historical lean scores worse on the registered measure, log loss, because the lean in past elections has not predicted the lean in the next one. Subtracting the average historical polling error from race polls is no better than leaving the polls uncorrected: the correction's gain in average error is within noise. The scale law, fitted on past elections as a uniform shift plus a stretch away from a tie, has two parts. The scale law's uniform shift, which for 2026 would move every Senate and governor forecast toward Republicans, is not applied: in past elections the shift improves log loss in presidential years but worsens it in midterms taken together, and 2026 is a midterm. FutureBallot scored the same 234 Senate and governor races of 2016-2024 on the eve of the election with and without the shift: the shift improves log loss in 2018 by 0.005 but worsens it in 2022 by 0.041. For those election-eve forecasts, pooled over all 234 races, the version without the shift has better log loss and Brier score, a worse average absolute error (4.36 points against 4.22) and a larger signed error toward Democrats (2.21 points against 1.21). The scale law's stretch away from a tie is applied, and the stretch favors neither party. One correction toward Republicans does reach the Senate and governor forecasts, indirectly: the national term in their ground-up estimate subtracts the generic ballot's average past miss toward Democrats, as the House forecast does; that generic-ballot correction has not had the standard test on its own. The lean above is measured with that correction in place.

    Under that lean, as a scenario and not a correction: every race shifted by the lean the model's own scored record shows, then re-simulated.

    The forecast under the measured lean
    ChamberPublished: expected seats, chance of controlShift (points)Under the shift
    Governors28 Democratic governors, Democrats more than 25 80%, Republicans more than 25 12%, a twenty-five each split or no party above twenty-five 9% (figures are rounded, so the three outcomes shown may not add to 100%)2.3 toward Republicans26 Democratic governors, Democrats more than 25 55%, Democrats 25 or fewer 45%
    Senate52 Democratic seats, Democrats 73–76%, Republicans 24–27%2.3 toward Republicans50 Democratic seats, Republicans 52–55%, Democrats 45–48%

    A check on how much the answer depends on that lean, not a prediction of which way the polls will miss in 2026.

  • The 50% range is the one furthest from its target. In the current record it contained the result 46.9% of the time against a target of 50% (too narrow).
  • The current record covers six election cycles. Effects that operate on a whole election have six observations, not 258. Anything reported per cycle is noisier than its decimal places suggest.

Measured on an earlier model's ten-cycle record

The items below come from the record of an earlier version of the statewide model, scored over 435 races in ten cycles (2006–2024). They are not re-measured on the current model and should not be combined with the figures above.

  • There is a partisan gradient it does not correct. The polls overstate the Democrat more in redder states, in proportion to how red the state is — negative in each of the 9 cycles with enough races to measure it, coefficient of variation 0.52, independent of the level. In the reddest quarter of states the lean is 5.9 points toward the Democrats whether the model has the Democrat or the Republican ahead — which side is favored makes no difference, only how red the state is. Correcting it improves average error from 5.46 to 5.24 and 90% coverage from 0.887 to 0.906. It was pre-registered under log loss, which it does not improve, and the gate refused it (q = 0.685). Switching the metric after seeing the result would make the pre-registration worthless, so the gradient stays uncorrected, and the finding is marked as suggestive only; 2026 is the finding's declared confirmatory election.
  • 2014 was the worst cycle and is not fully explained. A lean of 6.50 points toward the Democrats, 90% coverage 0.723 against a target of 0.90.
  • The regional term is 0.00 and that is a bound, not a measurement. The estimator cannot detect a regional standard deviation below about 2 points, so "no regional structure" means "none large enough for this design to see".

The public failure log records every defect found in this system, what it cost, and the check now intended to stop it recurring.

Every wrong call in the record

In twenty of the 258 scored 2016–2026 forecasts, the side the model favored lost. Each one is listed below: the race, the favorite and the chance the forecast gave it, and the candidate who won. These are the same scored forecasts as the record above and the reliability table below. The favorite is the candidate given more than 50%.

Every wrong call in the scored 2016–2026 record: twenty races whose favorite lost
CycleRaceFavoriteFavorite's chanceWinnerWinner's margin (points)
2016Indiana GovernorJohn R. Gregg/Christina Hale (D)63%Eric Holcomb/Suzanne Crouch (R)6.0
2016Missouri GovernorChris Koster (D)72%Eric Greitens (R)5.6
2016New Hampshire GovernorColin Van Ostern (D)61%Chris Sununu (R)2.3
2016New Hampshire SenateKelly Ayotte (R)51%Maggie Hassan (D)0.1
2016Pennsylvania SenateKatie McGinty (D)69%Pat Toomey (R)1.4
2016Wisconsin SenateRuss Feingold (D)86%Ron Johnson (R)3.4
2018Florida GovernorAndrew Gillum (D)78%Ron DeSantis (R)0.4
2018Florida SenateBill Nelson (D)81%Rick Scott (R)0.1
2018Indiana SenateJoe Donnelly (D)52%Mike Braun (R)5.9
2018Kansas GovernorKris Kobach (R)56%Laura Kelly (D)5.1
2018Nevada GovernorAdam Laxalt (R)59%Steve Sisolak (D)4.1
2020Georgia SenateDavid Perdue (R)68%Jon Ossoff (D)1.2
2020Maine SenateSara Gideon (D)69%Susan Collins (R)8.6
2020North Carolina SenateCal Cunningham (D)68%Thom Tillis (R)1.8
2022Alaska SenateKelly Tshibaka (R)51%Lisa Murkowski (R)7.4
2022Arizona GovernorKari Lake (R)74%Katie Hobbs (D)0.7
2022Kansas GovernorDerek Schmidt (R)53%Laura Kelly (D)2.2
2022Nevada SenateAdam Laxalt (R)58%Catherine Cortez Masto (D)0.8
2022Wisconsin GovernorTim Michels (R)54%Tony Evers (D)3.4
2024Pennsylvania SenateBob Casey Jr. (D)76%Dave McCormick (R)0.2

The winner's margin is the gap between the two candidates as a share of all votes cast, the scale the forecast is made on.

When the forecast said X%, how often did it happen?

Each scored forecast, grouped by its favorite's chance, with how often that favorite won. A well-calibrated forecast wins about as often as it says. Read it with care: races in the same election share that year's polling error, so each band rests on a handful of elections (the last column), not on independent races, and a band can miss by several points by chance.

In the 90-97% band favorites won more often than stated, by more than 5 points: there it was not sure enough.

When the forecast said X%, how often did it happen?
Favorite's stated chanceForecastsAverage stated chanceFavorite wonElections
50-60%1855%56%4
60-70%1766%65%5
70-80%1676%75%5
80-90%2186%90%5
90-97%3995%100%5
97% and up14799%100%6

Races the record does not score

sixty-nine races are not scored, for one of the reasons listed. A race where the result shows the forecast's two candidates were not the race (a third candidate won, or the two took under 80% of the vote) is outside what the forecast was asked; a race with no result for one of the two was forecast, but the record has no result to score for that pair. A race with no certified result on file is awaiting one: either the state has not certified the count, or the certified rows are a gap in the data. Both are listed here so the record is not read as covering them.

  • No result for one of the two contenders: AL governor 2026, AL senate 2026, AR governor 2026, AR senate 2026, AZ governor 2026, CO governor 2026, CO senate 2026, CT governor 2026, DE senate 2026, FL governor 2026, GA governor 2026, GA senate 2026, HI governor 2026, IA governor 2026, IA senate 2026, IL governor 2026, IL senate 2026, KS governor 2026, KS senate 2026, KY senate 2026, LA senate 2026, MA governor 2026, MA senate 2026, MD governor 2026, ME governor 2026, ME senate 2026, MI governor 2026, MI senate 2026, MN governor 2026, MN senate 2026, MS senate 2026, MT senate 2026, NH governor 2026, NH senate 2026, NJ senate 2026, NM governor 2026, NM senate 2026, NV governor 2026, NY governor 2026, OH governor 2026, OH senate 2026, OK governor 2026, OK senate 2026, OR governor 2026, OR senate 2026, PA governor 2026, RI governor 2026, RI senate 2026, SC governor 2026, SC senate 2026, SD governor 2026, SD senate 2026, TN governor 2026, TN senate 2026, VA senate 2026, VT governor 2026, WI governor 2026, WV senate 2026, WY governor 2026, WY senate 2026
  • No certified result on file: CA governor 2026, FL senate 2026, NC senate 2026, NE governor 2026, NE senate 2026, TX governor 2026, TX senate 2026
  • The two scored contenders took under 80% of the vote: AK senate 2016, AK governor 2022

Every test run on these past elections

In all, ninety-eight written-up tests and checks of the forecast models and their data are listed, most of them registered before they were run; many are scored on the same past elections as the record above, and others on elections outside it, including elections before 2016, elections in 2025 and state legislative contests. That reuse is why the record is development evidence, not proof: November 2026, scored from forecasts published before it, is the first clean test. The list of tests.