Senate and House: a broader research forecast
Separate experimental model · conditional coverage boundsPolls for competitive Senate races, fundamentals for sparse races, and district baselines on the maps used in 2026. Stan estimates uncertainty shared across states and both chambers.
The recorded Senate polling pilot remains unchanged. This extension is a research shadow. The ranges below assign uncovered contests to each extreme; they are coverage bounds, not confidence intervals or unconditional control odds.
What if both chambers move together?
Apply the same additional Democratic-minus-Republican margin shift to every modelled race, using the same simulated cycles. This isolates a shared miss. It leaves reported same-party fields fixed and uncovered contests unresolved.
Negative shifts favour Republicans; positive shifts favour Democrats. This is a manual sensitivity overlay on the model’s existing uncertainty, not a new calibrated forecast.
Scenario data table
Every Senate contest and House district
Race estimates below stay at the recorded forecast. “Same-party field” is conditional on the reported candidate list being complete; it is not an independently certified ballot capture. Independents and unresolved ballot rules are retained explicitly.
| Contest | Model D win | D−R margin | 80% margin range | Coverage / latest poll | Reported candidates |
|---|
Historical checks before interpreting the map
Expanding-cycle fits use four or more previous election cycles. Outcomes from the test year never enter its fit. These are retrospective, selected D/R contests: publication lags are proxies, some redrawn historical baselines are missing, and unopposed or multiparty outcomes are excluded from margin fitting. The 2024 Senate polling supplement uses Electoral-Vote, a source change from the older 538 archive.
These tables score the accepted population only. The chamber-wide historical population and exclusions are shown below. A good sampler does not establish predictive skill, and three correlated held-out elections do not establish an edge.
Stan versus a fundamentals-only independent-error baseline
Senate polling comparison on exactly matched polled races
Original polling-v1 parameters are refitted only to earlier cycles under its unchanged code. This comparison preserves v1’s probability clipping at 0.000001. The new main report uses raw Brier scores and clips log-loss probabilities at 0.01. No market prices enter either fit.
Full historical population and accepted subsets
Download all historical contests, including excluded rows and their reasons. November multiparty/runoff rows are population records, not a complete final-winner chamber score.
Excluded historical contests
Stan sampling diagnostics
Four chains; stricter sampling replaced an earlier run with two divergent transitions. Priors, data and the Stan program were unchanged. Published fits must pass zero divergences and treedepth hits, R-hat at most 1.01, bulk and tail ESS at least 400, and the energy check.
Sources and assumptions
House fundamentals use earlier presidential votes re-tabulated onto each target election’s district lines, not a district-number join across redistricting. State and district lean is measured against the national presidential margin of the same earlier year. The current generic ballot uses the latest observation per firm within 45 days; same-date variants are averaged. Historical 2024 archive cutoffs also respect insertion-date proxies; a later archive capture is not claimed as an observed pre-election vintage.
The model has chamber-specific lean and generic-ballot effects, a fitted Senate poll weight, correlated national cycle errors, state-cycle errors shared across chambers, and separate race noise for polled Senate, unpolled Senate and House. Modelled races use a D/R outcome approximation; minor-party wins are not simulated. House district polls, candidate-specific effects and incumbency adjustments are not included in this version.
Reported candidate lists are an evidence layer, not official nomination certification. New Senate independents are not assigned to a caucus. Senate arithmetic starts with 34 Democratic-aligned holdovers and 31 Republican holdovers. Democratic majorities require 51 Senate seats and 218 House seats; a 50–50 Senate is not counted as a Democratic majority. Source gaps and the accepted historical population limit interpretation.