Data & methodology
Readable by people. Reusable by software. Trace every number to its inputs.
Explore the Election Lab ↗ · Browse poll questions
Download the evidence
| Resource | Contents |
|---|---|
| Poll catalog JSON v1.0 · CSV | Eligible survey questions and candidate toplines, including questions outside the selected model matchup. Poll ID identifies a survey; question ID identifies a question. CSV has one row per candidate per question. Included means a question contains the selected model pair, not that every displayed answer enters the calculation; model_pair identifies the exact comparison in JSON. |
| Election analytics JSON v1.0 | Senate scenario grid, race uncertainty, polling weights, coverage, control leverage and reproducible simulation metadata. |
| Senate forecast · House snapshot | Original model outputs, assumptions, seeds, source hashes and input polls. House control remains null. |
| Historical baseline tests | Walk-forward historical results-only test. This does not validate the current polling blend or chamber odds. |
| Catalog schema · AI discovery guide | Stable fields, units, interpretation and links for citation. |
| Senate update health · House health | A failed source refresh retains the last successful snapshot and its original timestamp. |
Read the numbers correctly
Win probability is a share of simulated outcomes, not vote share. Margins are percentage points, positive toward the named opposition candidate. Senate margins use reported candidate percentages; House model margins normalize D/R support to two parties. Catalog candidate percentages are the original feed toplines and are not normalized. The catalog omits undecided and noncandidate answers, so displayed candidate percentages may sum below 100.
Scenario shifts move every Senate race margin toward the opposition (Democratic or independent) or toward the Republican, with the same random errors. A pinned race is an assumed outcome imposed on each simulation. It does not update the national error after observing a result and is not the conditional probability of other races given that result. Independent wins do not imply a caucus commitment. Democrats need 51 caucus seats; Republicans need 50 under the model’s tie rule.
Poll weights describe this model’s weighting choices, not pollster quality ratings. No probability is a certified result. Candidate matchups are inferred from polling, not a complete official nominee roster. The full forecast has not been calibrated through a polling-era backtest. Existing Senate methodology · House limitations.
Provenance, freshness and licensing
Polling comes from The New York Times Senate feed and House feed, reused under CC BY 4.0 and independently filtered and modified by DutyStation. The Times does not endorse this work. Poll links and feed labels are upstream observations; not every original release is independently verified, and linked pages can change or cover several contests. Report discrepancies rather than treating an aggregator label as an official fact.
Historical inputs are MEDSL Senate results and presidential results. Source hashes, source retrieval time, generated time, model version and as-of date appear in JSON. Generated time is not a new poll date. Updates run hourly through the existing process; failures preserve the last successful forecast. Polls older than 90 days, future field dates, hypothetical questions and subpopulation-only samples are excluded. Multiple questions from the same survey are not independent polls.
Schema version 1.0 is additive within this version; missing values are null, never fabricated zeros. Dates are ISO 8601; probabilities are 0–1, margins are percentage points, intervals are model outcome intervals rather than poll sampling margins of error. Cite the canonical race or Lab page alongside generated_at, model_version, source hash and the original poll release where available. CSV text that could execute as a spreadsheet formula is prefixed with an apostrophe.
Control paths and dependence on pollsters
Control pathways count joint outcomes in the exact baseline simulations across six races selected by model control leverage. The headline percentage is the fraction of all simulations with both that pattern and the selected chamber-control outcome. The secondary percentage uses only simulations ending with that control outcome. Other races vary. These are not products of independent race probabilities or guarantees that the displayed pattern is enough for control. Low counts are identified; zero sampled outcomes do not prove impossibility.
Race comparisons show the published mean and 80% model interval alongside a separate evidence sensitivity test. Each omission removes all selected observations from one pollster and recomputes the same existing weighting and historical blend, holding the chosen matchup fixed. The omission range spans those recomputed means, not a confidence interval or validated alternative forecast. If a Democratic matchup loses its only pollster, the output identifies its fallback to historical-only evidence. An independent with no remaining polls is unavailable. Pollster weight concentration is not an accuracy rating, and different organizations can share correlated errors.
The additive analytics.insights object has its own version, matching forecast/sample hashes and source timestamps. Read the insights JSON schema. Automated clients should use a descriptive User-Agent; existing abuse protections still apply.
Reviewed corrections & model changes
October 23, 2026 · Senate baseline 1.2.0. Dan Osborn is modeled as an independent. The polling feed labels his rows DEM; the FEC identifies Independent and the Nebraska certified ballot lists By Petition. An exact state/candidate-ID/name correction preserves the raw source party. No caucus is assumed. The existing model’s independent uncertainty treatment now applies. Older saved snapshots remain as originally published; differences across this revision include a model change.
House 0.1.1. Alaska’s projection is withheld because reducing a multi-candidate ranked-choice contest to D/R omits a leading independent. Its historical record remains intact. The catalog retains the actual independent-candidate and final-round polling questions. Read a source release illustrating the first- and final-round distinction.
Compare approaches
Explore Race to the WH’s forecasts and polling tools, FiveThirtyEight’s archived 2022 forecast codebook and the NYT polling collection. These are independent resources. Their forecasts and ratings are not imported or presented as consensus in this model. DutyStation’s contribution is the linked poll evidence, reproducible scenario experiment and visible distinction between weak data, assumed outcomes and modeled uncertainty.