COYNTOSS by @achasecoyne

METHODOLOGY · UPDATED SEPTEMBER 24, 2026

How the models work

Here is the recipe behind the numbers: the inputs, how the models combine them and where important blind spots remain.

SENATE MODEL · v0.53

Senate methodology

The Senate model estimates the two-party margin in each race. Positive numbers favor Democrats; negative numbers favor Republicans. The margin is an estimate, not a promise about what any one state will do.

  1. Set a state baseline

    Each state’s presidential vote is compared with the national result. The baseline gives 10% weight to 2016, 15% to 2020 and 75% to 2024—a long enough view to steady one unusual election, with the newest result doing most of the work.

    B = 0.10L2016 + 0.15L2020 + 0.75L2024

  2. Adjust for the national environment

    Silver Bulletin’s generic-ballot average supplies the national mood. Its effect is reduced: a one-point change in the ballot average moves a Senate race by 0.75 points before state polling enters the calculation.

    E = GCB × 0.75

  3. Add candidate fundamentals

    Elected incumbents receive two points; an appointed senator is treated as running for an open seat. Candidate strength comes from past results after accounting for the state baseline, national environment, incumbency and independent allowance. The remaining score is shrunk toward zero and capped. Version 0.53 allows unusually strong candidate evidence to carry through the final poll blend, up to one point.

    F = B + E + I + C + X

  4. Weight the polls

    Recent surveys count more than old ones, larger samples count more than smaller ones, and likely-voter polls count more than adult surveys. Pollster quality is also considered, and multiple versions of the same poll do not give a firm extra influence. Polls have a 45-day recency half-life; party-sponsored surveys are excluded.

    α = 0.85 × W / (W + 1.5)

  5. Blend fundamentals and polling

    The fundamentals estimate is blended with the polling average. Polls earn more influence as useful surveys accumulate, but their share tops out at 85%. Even a race with plenty of polling keeps some connection to its political baseline.

    M = (1 − α)F + αP̄

  6. Publish the race rating

    The margin and rating answer different questions. Ratings are Toss-up inside 3.5 points, Lean from 3.5 to 7, Likely from 7 to 15 and Solid beyond 15. Exceptions are documented when election rules or the candidate field make a two-party margin misleading. Major independents remain modeled as independents rather than being folded into a party.

Current limits and weighting rules

Candidate history can help, but it is easy to give it too much credit. The share of a past result already explained by the state, national environment, incumbency and any independent allowance is subtracted first. The normal candidate cap is three points in either direction. The wider six-point cap requires at least two consistent results before this forecast. After the poll blend, a candidate at that wider cap can retain 0.93 points. A candidate with three consistent statewide results can retain one point if the modeled strength is at least three. The post-blend adjustment never exceeds one point.

An independent is not a generic third-party candidate. When the same person has run statewide as an independent, half of that weighted overperformance is used and any candidate-strength credit already counted is subtracted. Without that history, the temporary three-point allowance applies only when the Senate race has polling. An unpolled independent gets no bonus, and neither does a routine third-party candidate.

Sample size is scaled by its square root, with a floor of 0.75 and a cap of 1.25. Likely-voter polls receive full electorate weight; registered-voter polls get 0.90, and adult polls get 0.75. Related releases from the same firm are also discounted. Pollster-quality multipliers run from 0.85 to 1.15.

The map colors show margins. Control probabilities come from a separate simulation and are still beta output. The model does not yet adjust for pollster house effects, and some candidate judgments still need a documented manual decision.

Sources for the Senate model: Silver Bulletin’s generic-ballot average and pollster predictive-plus-minus table; presidential results from 2016, 2020 and 2024, measured against the national result.

GOVERNOR MODEL · v0.0.1 EXPERIMENTAL

Governor methodology

This model is experimental, and its backtest is not finished. Estimates may change as additional election cycles are tested.

Governors can run ahead of—or behind—their party in a way senators often cannot. This experimental model therefore leans more on polls, candidate details and outside race ratings.

  1. Build a simpler baseline

    The model combines 65% of a state’s partisan lean with 55% of the national generic-ballot margin. Those weights are smaller than a full partisan baseline because governor races often diverge from federal voting.

    Simple = 0.65 × lean + 0.55 × GCB

  2. Account for candidates and incumbents

    Candidate experience can add up to three points, and incumbency is worth 2.5. Verified elected incumbents also get an adjustment for their electoral history; challenger quality is capped at two points.

  3. Translate expert ratings

    Cook and Sabato ratings are translated into margins: zero for Toss-up, five points for Lean, 10 for Likely and 18 for Safe or Solid. The two estimates are then averaged. It is a deliberately simple translation of expert categories into numbers.

  4. Blend polling and estimate uncertainty

    Governor polls have a 30-day half-life and can account for as much as 90% of a forecast. Win chances use 100,000 correlated t(7) simulations; the heavier tails allow more extreme outcomes than a normal curve.

Validation and limitations

The incumbent-history rule was set using elections before 2014, then tested on 254 usable governor races from 2004 through 2022. Complete historical polling and expert-rating series could not be reconstructed, so the poll blend and uncertainty settings remain provisional.

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

BETA · 2026 U.S. HOUSE

House map methodology

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

Most district boundaries come from the Census Bureau’s 2026 TIGER/Line files for the 120th Congress. Missouri uses the 119th-Congress map listed in the boundary manifest. Census labels at-large districts “00”; the public map displays them as 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 map spreads that estimate across counties—or Alaska House districts—by preserving how each place voted relative to its state in 2024.

  1. Start with 2024 presidential results

    For each county or district, the calculation begins with the Democratic share of the 2024 Democratic-plus-Republican vote. Third-party votes stay out of the calculation and are not assigned to either major party.

  2. Apply one statewide log-odds shift

    One shift is applied to every place, preserving its 2024 relationship to the state. Log-odds keep the result between 0% and 100%, including in places where one party won nearly every vote.

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

  3. Match the statewide forecast

    The shift is solved so vote-weighted local estimates add up to the statewide COYNTOSS margin. Every geographic ID must match the corresponding election result before publication.

What the local map does not claim

The 2024 vote supplies the map’s weights; it does not predict 2026 turnout. A Senate or governor candidate is not assumed to recreate the presidential coalition county by county. The colors distribute a statewide estimate; they are not 50 miniature county forecasts in a trench coat.

Alaska appears as 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. Source files are left intact rather than inventing votes to make the map tidier.

Connecticut and Massachusetts also have town maps, built from official 2024 results and current Census municipal boundaries. Other town views will be added only when election and boundary files line up cleanly.

Sources: 2024 cartographic files from the U.S. Census Bureau; county and Alaska House returns from Tony McGovern’s documented 2024 compilation; town returns from the Connecticut Secretary of the State and Massachusetts Secretary of the Commonwealth. Every geographic ID is checked before publication.