COYNTOSS by @achasecoyne

METHODOLOGY · UPDATED SEPTEMBER 5, 2026

How the models work

How state fundamentals and polling become race estimates, with current limits.

SENATE MODEL · v0.52

Senate methodology

The Senate model estimates the two-party margin in every race. Positive numbers favor Democrats; negative numbers favor Republicans.

  1. Set a state baseline

    We start with how each state voted for president compared with the country as a whole. The baseline uses the last three elections, with 75% of the weight on 2024.

    B = 0.10L2016 + 0.15L2020 + 0.75L2024

  2. Adjust for the national environment

    Next, we use Silver Bulletin’s generic-ballot average to account for the national mood. A one-point change in that average moves a Senate race by 0.75 points before state polls are added.

    E = GCB × 0.75

  3. Add candidate fundamentals

    Elected incumbents get two points. Appointed senators do not; we treat those races as open seats. The candidate adjustment comes from prior election results after subtracting what the state, national environment and incumbency already explain. We shrink those scores toward zero and cap them. Version 0.52 also lets a small amount of especially strong candidate evidence survive the final poll blend, but never more than one point.

    F = B + E + I + C + X

  4. Weight the polls

    Recent polls count more than old ones, large polls count more than small ones and likely-voter polls count more than surveys of adults. We also adjust for pollster quality and keep one firm from gaining too much influence by releasing the same survey several ways. The recency half-life is 45 days. Party-sponsored polls are excluded.

    α = 0.85 × W / (W + 1.5)

  5. Blend fundamentals and polling

    We then combine the fundamentals estimate with the polling average. Polls gain influence as useful polling accumulates, but they can never make up more than 85% of the forecast. Even a well-polled race keeps some connection to its political baseline.

    M = (1 − α)F + αP̄

  6. Publish the race rating

    The margin and the rating are separate outputs. The usual bands are Toss-up inside 3.5 points, Lean from 3.5 to 7, Likely from 7 to 15 and Solid beyond 15. We allow documented exceptions when a race’s rules or candidate field make the raw two-party number misleading. Major independents are modeled as independents, not quietly renamed Democrats.

Current limits and weighting rules

Candidate history is useful, but it is also easy to overrate. We first remove the part of a past result that can already be explained by the state, the national environment, incumbency and any independent allowance. A candidate is normally capped at three points in either direction. The wider six-point cap requires at least two consistent results from before the forecast. After polling is blended in, a candidate who hit that wider cap can keep another 0.93 points. A candidate with at least three consistent statewide results can keep one point if the modeled strength is at least three points. The post-blend adjustment is capped at one point.

Independents are not interchangeable. If the same candidate has already run statewide as an independent, we use half of that weighted overperformance and subtract any candidate-strength credit already counted. If there is no such history, the temporary three-point allowance applies only when the current Senate race has polling. An unpolled independent gets no bonus. Neither does a routine third-party candidate.

Sample size enters through a square-root adjustment, capped from 0.75 to 1.25. Likely-voter polls get full electorate weight, registered-voter polls get 0.90 and adult polls get 0.75. We down-weight related releases from the same firm. Pollster-quality multipliers run from 0.85 to 1.15.

The map shows margins. The control percentages come from a separate simulation and should still be read as beta output. The model does not yet correct for pollster house effects, and some candidate judgments still require a documented manual decision.

Sources: Generic ballot: Silver Bulletin. Presidential baseline: 2016, 2020, and 2024 state results relative to the national result. Pollster-quality reference: Silver Bulletin predictive plus-minus table.

GOVERNOR MODEL · v0.0.1 EXPERIMENTAL

Governor methodology

Experimental model. We have not finished backtesting this model. Its numbers can still change as that work is completed.

Governor races are less tied to national partisanship than Senate races. This model therefore gives more weight to polls, candidates and outside race ratings.

  1. Build a simpler baseline

    We combine 65% of the state’s partisan lean with 55% of the national generic-ballot margin. The weights are deliberately smaller than a full partisan baseline because governor races often break from federal voting patterns.

    Simple = 0.65 × lean + 0.55 × GCB

  2. Account for candidates and incumbents

    Candidate experience can move a race by as much as three points. Incumbency is worth 2.5 points. Verified elected incumbents also receive an electoral-history adjustment, while challenger quality is capped at two points.

  3. Translate expert ratings

    We translate Cook and Sabato ratings into margins: zero for Toss-up, five points for Lean, 10 for Likely and 18 for Safe or Solid. We then average the two estimates.

  4. Blend polling and estimate uncertainty

    Governor polls have a 30-day half-life. They gain influence as useful polling accumulates, up to a 90% ceiling. We estimate win probabilities with 100,000 correlated t(7) simulations, which allow for more extreme outcomes than a normal distribution.

Validation and limitations

We selected the incumbent-history rule using elections before 2014, then tested it on 254 usable governor races from 2004 through 2022. We could not reconstruct complete historical polling and expert-rating series, so the poll blend and uncertainty settings remain provisional.

The model allows more uncertainty than the congressional forecasts. Race-level standard deviation is five points when polling is strong, six with medium polling, 7.5 with little polling and 12 when there are no polls.

BETA · 2026 U.S. HOUSE

House map methodology

The House page reads the district estimates already calculated in the workbook; it does not rerun the model in your browser. It matches the districts, probability-inputs, polling-summary and polls tabs using the standard state-and-district ID.

Most district lines come from the Census Bureau’s 2026 TIGER/Line files for the 120th Congress. Missouri is the exception: the site uses the 119th-Congress map recorded in the boundary manifest. Census labels at-large districts as 00; the public map calls them district 01.

House source: Open the House forecast map. The workbook and boundary record are beta and subject to source/status review before publication.

LOCALIZED FORECAST · v1

How statewide margins become local estimates

The statewide margin is the forecast. The local maps show what that forecast would look like across counties—or Alaska House districts—if places moved together from their 2024 presidential vote.

  1. Start with 2024 presidential results

    For each county or district, we calculate the Democratic share of the 2024 Democratic-plus-Republican vote. Third-party votes are left out rather than assigned to either party.

  2. Apply one statewide log-odds shift

    We apply the same shift to every place while preserving its 2024 relationship to the state. Using log-odds keeps every estimate between 0% and 100%, even in counties where one party received almost all the vote.

    local log-odds = 2024 local log-odds + statewide shift

  3. Match the statewide forecast

    We solve for the shift that makes the vote-weighted local estimates equal the statewide COYNTOSS margin. A map is not published unless every geographic ID also matches the corresponding election result.

What the local map does not claim

The 2024 vote is used to weight the map. It is not a prediction of 2026 turnout, and the map does not assume that a Senate or governor candidate will rebuild the presidential coalition county by county. The local colors are one way of distributing the statewide forecast, not 50 separate county models wearing a trench coat.

Alaska is shown through its 40 state House districts because that is how its presidential results are reported. Kalawao County, Hawaii, is omitted because the source does not report it separately. We do not invent votes to make either file look tidier.

Connecticut and Massachusetts also have town maps. They use official 2024 results and current Census municipal boundaries. Other town views will remain off until the election and boundary files match cleanly.

Sources: Boundaries are from the U.S. Census Bureau’s 2024 cartographic files. County and Alaska House district returns are from Tony McGovern’s documented 2024 compilation. Town returns come from the Connecticut Secretary of the State and Massachusetts Secretary of the Commonwealth. COYNTOSS checks every geographic identifier before publication.