Methodology · Model changes
Every change to the model
Numbered hexes are model versions, releases that changed the model's arithmetic, with what shipped alongside beneath them. Whenever the model changes, every past day is recomputed, so a history chart always shows one model.
- FixCorrects how the model or its data handled a particular case.
Polls for a race's lone independent are corrected, not dropped
Oct 26A poll whose sponsor the feed tags independent now counts as that side's sponsored poll, lean removed, when the race has exactly one independent and the forecast places them on a side. Nebraska's Senate race gains two such polls.
Fix · Polling average · All offices
- FixCorrects how the model or its data handled a particular case.
Alaska governor polls match the general ballot
Oct 26Alaska's governor field now follows the Division of Elections' general ballot, which replaced a withdrawn finalist with the fifth-place ticket, and polls naming each ticket's lead candidate match it. The race gets a polling average.
Fix · Polling average · Governor
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Governor averages add The New York Times's polls
Oct 25Governor polls now come from The New York Times as well as VoteHub, adding polls in 30 of 36 races and the first polls in Arkansas, Idaho and Oklahoma. A poll both carry counts once.
Improvement · Polling average · Governor
- FixCorrects how the model or its data handled a particular case.
Swapped field dates are repaired when the survey confirms them
Oct 25A poll whose start and end dates arrive reversed is corrected when every other record of the same survey has the swapped dates. Otherwise it stays out of the averages.
Fix · Polling average · All offices
- FixCorrects how the model or its data handled a particular case.
Governor polls match joint tickets
Oct 22A poll naming only the top of a joint ticket now matches that ticket, which brings Ohio's governor polls into its average.
Fix · Polling average · Governor
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
A single feed label can be excluded
Oct 20Besides firms the ratings ban, Bellwether can leave one exact pollster label out of every average, for a firm with no rating of its own.
Improvement · Pollster ratings · All offices
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v15Oct 20
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
House district polls stay in longer
A district poll now counts for up to 240 days, discounted by age, instead of fading out by day 75. Many districts are polled only once or twice.
TestedBetter log loss on 2022 and 2024 House races.
Improvement · Polling average · House
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
A candidate-scandal term for House races
A nominee with a publicly known scandal loses about four points of margin from the day it became public.
TestedFitted on 51 contested races, 2014–2024.
Improvement · New input · House
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v14Oct 19
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Redrawn incumbents keep their overperformance
In states redrawn for 2026, a returning incumbent keeps their 2024 overperformance, tapered by how much their district changed.
TestedHeld in the 2022 redistricting.
Improvement · New input · House
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
House fundamentals refit on 2014–2024
Fitted on six cycles instead of two. Seat-level uncertainty rises from 5.0 to 6.0 points.
TestedTruer 80% intervals across six cycles.
Improvement · Refit · House
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v13Oct 17
- FixCorrects how the model or its data handled a particular case.
Appointed incumbents treated as open seats
An incumbent who did not win the seat's last election gets no incumbency credit, as in the historical fit.
Fix · New input · Senate · Governor
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Senate uncertainty sized on recent cycles
Senate fundamentals uncertainty now comes from the five latest cycles: 9.8 points, down from 13.8 since 1992.
TestedBetter held-out log loss, 1992–2024.
Improvement · Refit · Senate
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Polling error floors refit on size of miss
The floor on a polling average's error is fitted on how far averages missed, not only who won: 5.1 points for the House, 5.55 for the Senate.
TestedBrier 0.078 → 0.075 on 2022–2024 polling.
Improvement · Refit · House · Senate
- FixCorrects how the model or its data handled a particular case.
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v12Oct 14
- FixCorrects how the model or its data handled a particular case.
State lean measured against the nation
A state's lean is now its margin relative to the national margin in the same election, so the fit and the forecast read it on one scale.
TestedSenate even in backtests; governor slightly better.
Fix · Structure · Senate · Governor
- FixCorrects how the model or its data handled a particular case.
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v11Oct 14
- FixCorrects how the model or its data handled a particular case.
Ranked-choice polls read one candidate per side
Where a party has two candidates on a first-choice ballot, its leading one is read. A poll's final two-way round is preferred.
Fix · Polling average · House · Senate
- FixCorrects how the model or its data handled a particular case.
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v10Oct 13
- FixCorrects how the model or its data handled a particular case.
One-party fields are certain party outcomes
When certified ballot rules leave only one party eligible, as in some California top-two races, that party's win is certain.
Fix · Structure · All offices
- FixCorrects how the model or its data handled a particular case.
- FixCorrects how the model or its data handled a particular case.
Each survey counts once per matchup
Oct 13A survey's population cuts and ballot versions count as one poll: final round first, then likely voters, registered voters, adults.
Fix · Polling average · All offices
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Polling error floors refit for the fade
Oct 7Refit with the 45-to-75-day fade in place. The live effect was small.
Improvement · Refit · All offices
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Polls fade out instead of expiring
Oct 6A poll counts in full for 45 days, then fades to zero by day 75, rather than dropping out all at once.
TestedSteadier averages, as accurate or better.
Improvement · Polling average · All offices
- Public launchOct 5
Earlier changes were made while the site was in private preview.
- OtherA change to what is published or to the map, not to the method.
Only the independent model is published
Oct 3Public pages show the model that never sees market prices. The market blend is kept for internal comparison.
Other · Publishing · All offices
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v9Sep 27
- FixCorrects how the model or its data handled a particular case.
Generic-ballot polls matched across feeds
The two poll feeds label generic-ballot answers differently; both now read as the same matchup.
Fix · Polling average · All offices
- FixCorrects how the model or its data handled a particular case.
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Independents' internal polls counted
Sep 18A poll from an independent's campaign counts with its fitted lean removed, as party internals already did.
Improvement · Polling average · Senate
- FixCorrects how the model or its data handled a particular case.
Independents rated apart from their caucus
Sep 17Who an independent runs against is stored apart from the caucus they would join, so a seat can count as unaligned.
Fix · Structure · Senate
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Partnerships count as their own pollsters
Sep 16A partnership such as University of Houston/YouGov is its own pollster, and renamed firms are folded together.
Improvement · Pollster ratings · All offices
- FixCorrects how the model or its data handled a particular case.
Incumbency and money follow the nominee
Sep 15Incumbency and fundraising credit go to the settled nominee, not to the seat's party or every filer.
Fix · New input · House · Senate
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v8Sep 4
- OtherA change to what is published or to the map, not to the method.
Missouri moves to its restored 2022 districts
Missouri's House forecast uses the restored 2022 lines. Polls from before the ruling wait for boundary review.
Other · Geography · House
- OtherA change to what is published or to the map, not to the method.
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v6Aug 28
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
House fundamentals read fundraising
Each side's share of two-party FEC receipts is an input, which matters most in districts without polls.
TestedHeld-out uncertainty 5.3 → 5.0 points.
Improvement · New input · House
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
One polling average per matchup
Aug 28Each polled pairing gets its own average. Before a primary settles, pairings are weighted by how likely each is.
Improvement · Polling average · All offices
- FixCorrects how the model or its data handled a particular case.
Decided seats are counted, not simulated
Aug 23A seat already won, such as one with no opponent filed, counts for its winner in every run.
Fix · Structure · All offices
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Sponsored polls counted, lean removed
Aug 23Campaign and partisan polls count after their fitted lean toward the sponsor is removed: 6.4 points in the House, 5.6 for Senate and governor.
TestedMore districts averaged; 2024 House Brier 0.114 → 0.103.
Improvement · Polling average · All offices
- FixCorrects how the model or its data handled a particular case.
Averages follow a changed matchup
Aug 22The guard against sudden jumps in an average lets one through when the race's matchup changes.
Fix · Polling average · All offices
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Silver Bulletin's 2026 pollster ratings adopted
Aug 22Ratings enter through each firm's historical lean and its ban status. The quality score fitted at zero weight.
Improvement · Pollster ratings · All offices
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v5Aug 17
- OtherA change to what is published or to the map, not to the method.
Polling averages keep their uncertainty
Each day of a polling average's history stores its uncertainty, so race pages can draw a band around the line.
Other · Publishing · All offices
- OtherA change to what is published or to the map, not to the method.
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v4Aug 17
- FixCorrects how the model or its data handled a particular case.
Polls matched across feeds by candidates
A poll carried by both feeds is matched on the candidates it offered, so it counts once.
Fix · Polling average · All offices
- FixCorrects how the model or its data handled a particular case.
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v3Aug 16
- OtherA change to what is published or to the map, not to the method.
Chamber dots sampled by percentile
The hundred runs drawn on a chamber page are one per percentile of the simulation, so their split matches the published probability.
Other · Publishing · All offices
- OtherA change to what is published or to the map, not to the method.
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v2Aug 16
- OtherA change to what is published or to the map, not to the method.
Senate history records its ballot reading
Each recomputed day of the Senate forecast stores the generic-ballot reading it used, falling back to the House average as the live forecast does.
Other · Publishing · Senate
- OtherA change to what is published or to the map, not to the method.
- Model versionA release that changed how the forecast is calculated. What shipped with it is listed inside.
Model v1Aug 16
- FixCorrects how the model or its data handled a particular case.
House national anchor refit
The House anchor is fitted on the same presidential lean it is applied to.
Tested2024 Brier 0.027 → 0.021.
Fix · Refit · House
- FixCorrects how the model or its data handled a particular case.
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
House effects shrink toward each firm's history
Aug 12A pollster's lean this cycle is shrunk toward its historical lean, not toward zero.
TestedBetter 2024 House and Senate Brier.
Improvement · Pollster ratings · All offices
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Governor forecasts added
Aug 11Every 2026 governor race is forecast, with fundamentals and a polling floor fitted on past governor races.
Tested140 held-out races: Brier 0.159 → 0.108.
Improvement · Structure · Governor
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Poll weights fitted on 2,533 past polls
Aug 9Sample size, age, house effects and an error floor are fitted on FiveThirtyEight's House and Senate archive.
Tested2022–2024 Brier 0.107 → 0.081.
Improvement · Polling average · House · Senate
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Senate fundamentals fitted on 2016–2024
Aug 9Fitted values replace hand-set ones. The window was chosen by testing on races back to 1992.
Tested536 held-out races: Brier 0.114 → 0.100.
Improvement · Refit · Senate
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Environment uncertainty grows with time left
Aug 9The generic ballot is read as a forecast of November, with uncertainty that grows with the days remaining.
Improvement · Structure · House
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
House fundamentals fitted on past results
Aug 8Lean, incumbency and overperformance are fitted on 788 contested 2022 and 2024 races, each year's national environment estimated apart.
TestedFit on 2022, scored on 2024: error 3.8 → 3.3 points.
Improvement · Refit · House
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
An independent model, apart from markets
Aug 8Each race gets a forecast that never sees market prices. No blend is more confident than its most confident input.
Improvement · Structure · All offices
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Correlated errors in the simulation
Aug 8Each run draws a national, a state and a race error, so seats in one state move together.
Improvement · Structure · All offices
- ImprovementA new input, method or fit, adopted because it tested better or covers more.
Estimates combined as margins
Aug 8Inputs are combined as margins weighted by precision, and disagreement widens the result.
Improvement · Structure · All offices
- FixCorrects how the model or its data handled a particular case.
Duplicate polls counted once
Aug 8A poll carried by both feeds counts once, and a firm spelled two ways counts as one.
Fix · Polling average · All offices