2026 U.S. midterm forecast

A Monte Carlo forecast of the 2026 U.S. House, Senate and governor elections, with an experimental forecast of 22 state legislative chambers. It is published at scorvec.com/midterms; the reader-facing explanation is How the forecast works (web/about.html).

The election was held on 3 November 2026. The forecast was frozen at 17:04 UTC that day, before the first polls closed (data/final/, with its SHA-256 and commit; see The daily run), and the site now opens on the results and how the frozen forecast did. results.yml reads the calls and counts, scores the forecast, writes the results page's data and projects the races still out (live/remaining.js: the frozen forecast's own simulations, from live/engine.js and data/final/whatif_2026.json at the frozen commit, reweighted by every call and count); the daily forecast no longer runs.

Each day the model re-reads the polls, refits a national generic-ballot trend from individual polls, blends every race's polls with a fundamentals prior and simulates the elections 20,000 times with correlated errors: a national miss, a statewide miss, regional and demographic swings, and each race's own error. The whole run happens in GitHub Actions and commits its outputs to web/; the pages are static, and every position of the "national mood" slider is precomputed.

page	what it shows
web/index.html	the races still out (the front page): each chamber's median outcome and seat range, and every uncalled race's chance from the calls and counts so far
web/results.html	the results: each chamber as called, the national House vote against every election since 1946, every race next to its frozen forecast, and the scorecard
web/forecast.html	the final forecast, as frozen: chamber and governor odds, seat distributions, every race and how each moved
web/about.html	how the forecast works, the track record and every source
web/legislatures.html	state legislative chambers (experimental), with district maps

How the model works

National environment (generic.py, national_mood.py). The generic-ballot trend is fitted to individual polls, not to other people's averages: a penalised piecewise-linear trend (knots every 14 days), one house effect per pollster (ridge-shrunk, centred), likely / registered / adult offsets and a partisan-sponsor offset. Each poll sits at the midpoint of its field period, and the versions of one survey (for example likely and registered voters) count as one poll. A poll's weight is the square root of its sample size (capped at 1,500), lowered for a pollster whose polls scatter about the trend more than sampling explains, times the pollster's quality rating squared (score out of 3; a C+ pollster keeps 0.36 of its weight); its precision halves every 60 days of age, and no pollster's polls may carry more than 15 % of today's value. The approval trend is fitted the same way. The trend is taken as it stands: there is no correction for the direction of a polling miss. The size of past misses, in either direction, sets the national error s(L), which grows with the days L to the election (fitted on the 15 House elections 1996-2024; data/static/national_mood_fit.json). The part of that error that is late opinion movement is scaled state by state by how much of the vote is already banked (early_vote.py), and so is the part of each race's own error that is late movement in that race (0.25 pts² a day, measured on 538's raw_polls from how polls of the same race by different pollsters drift apart over the final two months; data/static/race_move_fit.json). A state banking at the historical pace keeps its total error; one banking faster gets tighter, one banking slower wider.

House (model.py, run2026.py). Each seat's expected Democratic two-party margin starts from a fundamentals prior: the district's partisan lean (our own, from precinct returns: 0.75 x its 2024 presidential margin against the nation's + 0.25 x the same for 2020, on the 2026 maps; cd_lean_build.py), incumbency and the nominees' fundraising through June 30 (money.py), fitted on the contested seats of 2018 and 2022 with the same lean on those elections' maps, plus the national environment. Demographic terms carry the Hispanic and Asian swing since 2024, with Texas, Florida and Western Hispanic factors, white non-college and urbanization factors; the district polls update those factors (model.factor_update). Race polls are corrected before they are blended, with the corrections measured on 538's archive of rated polls (race_poll_calibration.py): sponsored polls are shifted and down-weighted, each pollster's measured lean is removed, polls with many undecided or third-party respondents and polls from inexperienced pollsters get more variance, older polls fade (63-day half-life, 150-day window), and several versions of one survey count once. The blend is a Student-t (robust) update against the fit's own prior sd, so a poll that disagrees with everything else counts for less until other polls confirm it; the measured poll errors enter as standard deviations (the t scale is adjusted for the degrees of freedom). A poll by a pollster rated A or better with no sponsor is taken at face value instead: its error is normal with the t's scale (model.TOP_RATED_SD; ratings as known before each election, neutral in the backtest).

Senate (senate2026.py). The same pipeline with a prior refitted on every Democrat-v-Republican Senate race 1982-2024, with linear time trends in the lean and incumbency slopes (the Senate has nationalized). Independent challengers (a wider prior; 12 points more for one with no Democrat on the ballot and at least a quarter of the Republican's money, see Independent challengers), appointed incumbents and split anti-Republican fields have their own rules. A common trend estimated across all of this year's Senate polls brings older polls up to date. Control is counted the way it is settled: 50-50 goes to the Republicans while the vice president is a Republican, and independent winners are assumed to abstain. The page gives the chance that each party controls the Senate, the share of simulations in which the independents' choice decides it, and the answer if they side with either party.

Governors (gov2026.py). A prior fitted on every governor race 1998-2024 - the lean slope with a time trend, elected and successor incumbents, and the nominees' prior office (gov_quality.py) - with the Senate's poll corrections, poll errors and blend (polls from VoteHub and the Wikipedia race pages).

Ranked-choice races (rcv.py). In Maine (Senate, both House seats) and Alaska (Senate, House, governor) a poll's own final round is used where it reports one; a first-round-only poll is converted to an expected final-round margin with transfer rates measured on the official round-by-round tabulations, and the spread of those rates adds to the race's polling uncertainty.

Simulation. 20,000 draws with Student-t (5 df) errors: a national draw for the House, a statewide draw shared by the Senate and governor races (correlated 0.3 with the national draw), a state draw, urbanization and demographic-group factors, and each race's own error. A slider position is a shift of the same draws (House margins one for one, Senate and governor races by their fitted national slopes), so moving the slider changes the answer only through the assumption. A forward heating-oil adjustment (below) is added after the poll blend.

State legislatures (stateleg.py, experimental). Seat-by-seat simulations of 22 chambers on the same national draw: the district's 2024 presidential lean, incumbency and the national environment, a state error shared by both chambers and a seat error, with multi-member districts, held-over senate seats and tie rules handled per chamber.

Diagnostics that do not feed the forecast. Early and mail ballots by race, party or composition in North Carolina, Florida, Georgia, Pennsylvania and Nevada's Clark and Washoe counties; Texas early-vote turnout (off by default); voter registration in North Carolina and Pennsylvania; same-pollster poll-to-poll changes. Details below.

Validation

Changes to the model are scored on leak-free, walk-forward harnesses that run the live code as of past dates with only the information available then. Model notes records the scores, including what was tested and left off.

• House and Senate (the track record on the about page, build_web.BACKTEST_NOTE; computed before the Senate error re-split below, with a shared statewide miss of 4.2, and with the 538 / Cook district leans the House prior used before its own lean - on the live-path harness that switch moved the September 15 calls by about a seat, see District lean). House, each year forecast from the other years' data: on September 15 the model gave Democrats 227 seats (10-90 % range 202-257) in 2018 against 235 won, and 213 (191-234) in 2022 against 213; over 12 dates from July to November the result fell inside the 10-90 % range every time. Senate, 2018-2024 at 6 dates each: the seat count fell inside the 10-90 % range at all 24; in all four cycles the Senate polls missed toward the Republicans - a direction the model deliberately does not assume.
• Governors (gov_backtest.py): every race 2006-2024 at up to six dates (234 races, 1,086 race-dates, 10 cycles); see Governor model review for the scores.
• State legislatures (stateleg_backtest.py): 2018 and 2022 on October 1; see State legislatures.
• Ranked-choice transfers (python -m midterms.rcv --backtest): each count predicted from its official first round with rates fitted on the other counts (web/about.html#rcv).
• Reconstructed history (backfill.py): the history lines on the forecast page are the current model run as of every past date with only the polls released by then, the fundraising reports received, the pollster experience, the heating-oil price and the early-vote snapshot of that date. python -m midterms.backfill --check reruns today's reconstruction and compares it with web/data/model.json.

Where cycles are compared, significance is judged per election cycle, since a cycle shares one national miss: an exact sign-flip permutation of the cycle means and a race-cluster bootstrap stratified by cycle.

The daily run

.github/workflows/daily.yml is scheduled for 02:30 UTC (22:30 New York time, after the day's poll releases; GitHub can start scheduled runs late). It runs python -m midterms.bootstrap, then python -m midterms.weekly --daily:

1. Downloads the fixed public inputs into data/raw/ if they are missing (bootstrap.py, checksums verified).
2. National polls: VoteHub's open API (generic ballot, approval and race-poll catalogues), Polling USA and Political Poll Bot posts on Bluesky, pollresults.org and hand-entered releases; generic.py refits the generic-ballot trend.
3. UF Election Lab early-vote counts (early_vote.py) and EIA's weekly residential heating-oil price (heating_oil.py).
4. Wikipedia: the House seat table (2026 maps, incumbents, nominees), race ratings, every Senate and governor race page and the 50 state House-election pages; VoteHub and pollresults.org race polls; the Bluesky feeds (refresh.py).
5. build_web.py: the House, Senate, governor and state-legislature simulations -> web/data/model.json, polls.json, whatif.json and stateleg_seats.json.
6. The daily snapshot web/data/history/model_<date>.json; the reconstructed history (backfill.py, rebuilt only when the methodology files change), web/data/race_history.json and web/data/summary.json (the site's homepage panel).
7. Diagnostics that do not feed the forecast (Florida and Pennsylvania early / mail vote, North Carolina and Pennsylvania registration on Mondays, same-pollster deltas); they are kept in the Actions cache or uploaded as a workflow artifact, never committed.
8. On Mondays, reports/weekly_<date>.md: which races moved in the past week and why.

build_web.py also writes data/cache/forecast_detail.json, the scorable forecast (final.py): every race's chance and the percentiles 1-99 of its simulated margin, the chambers' seat distributions and the national vote.

Freeze. Once, before the polls close, dispatch daily.yml with freeze: python -m midterms.final freeze copies that run's detail and model.json into data/final/ with a SHA-256 and the commit, and python -m midterms.scorecard template writes a row per race into data/results/results_2026.csv. From then on scheduled and dispatcher runs stop (after_freeze runs one anyway).

Publish gate (weekly.publish_gate). The outputs are held when a race that had polls suddenly has none, the race-poll count falls by more than 10 %, or the generic-ballot poll count falls - usually a Wikipedia table that stopped parsing. The workflow then commits only the poll records in data/state/ and fails; force_publish on a manual run overrides it.

Commit. The outputs are snapshotted, then on each push attempt the job resyncs to the newest commit of the branch, restores the snapshot and commits (no rebase, no merge of stale trees). data/raw/ and data/cache/ persist between runs through actions/cache. A run takes a few minutes on a normal day and about 20 minutes when the reconstructed history is rebuilt or the download cache is cold.

Publishing. The website copies the pages and web/data/ from this repository at the commit SHA of main (site_integration/midterms-sync.yml is a reference copy of that workflow) and refuses a forecast that is broken or older than the one it has.

Other workflows:

workflow	what it does
results.yml	hourly, on demand and on every hand edit of data/results/results_2026.csv or outside_calls.csv (calls a news organisation made first, entered by hand with their source): the calls, counts and national House vote from Wikipedia (results_wiki.py; rows from official returns are kept, a count entered from a news tally keeps its numbers and takes Wikipedia's call), the projection of the races still out (live_obs.py, then node live/remaining.js -> web/data/live.json, live_history.json), the scorecard (scorecard.py, rules in its docstring, -> data/results/scorecard_2026.json) and the results page's data (results_page.py -> web/data/results.json, and the homepage's results_summary.json with the median outcome)
stateleg-geo.yml	builds the district maps web/data/stateleg_geo/ from Census TIGER/Line when the chamber list or the district leans change
stateleg-data.yml	one-time builds of the state-legislature tables in data/static/stateleg/ (triggered from a feature branch)
cd-lean.yml	manual: the congressional-district leans (cd_lean_build.py) on the 2026 maps or on the 2018 and 2022 maps
district-inputs.yml	manual: the urbanization index, the House map shapes and (for states back on their old lines) the demographic tables on the maps in force for 2026 (demog.cd_blocks); dispatch after a court ruling changes a state's map

Run it yourself

Python 3.11 and network access:

pip install -r requirements.txt
python -m midterms.bootstrap            # fixed public inputs into data/raw/ (once)
python -m midterms.weekly --daily       # the daily run: fetch, fit, simulate; writes web/data/
python -m http.server -d web            # serve the pages, then open the printed address

The first run downloads every source page and takes longer than a daily run. Individual pieces:

python -m midterms.generic                      # generic-ballot and approval trends -> data/cache/generic_model.json
python -m midterms.build_web                    # simulate from the cached inputs -> web/data/
python -m midterms.backtest 2022 2022-09-15     # the live House path as of a date in a past cycle, scored
python -m midterms.gov_backtest all             # governor harness: every variant, then the report
python -m midterms.stateleg_backtest            # state-legislature backtest
python -m midterms.backfill --check             # today's reconstruction against web/data/model.json

Tests (tests/) need only the committed files and no network, and take a few seconds; CI runs them with the ruff checks on every code change (.github/workflows/ci.yml):

pip install pytest
pytest

Switches read from the environment: MIDTERMS_STATELEG (state legislatures, default on), MIDTERMS_STATELEG_KAPPA (the tested-and-rejected state-poll signal, default off), MIDTERMS_RCV (ranked-choice handling, default on), TX_EARLY (Texas tracker, default off) and MIDTERMS_FORCE_PUBLISH. Model switches that were tested and left off are module constants, each documented where it is defined.

Repository layout

path	what
midterms/	the model (Python 3.11; requirements.txt); paths.py defines the data layout
web/	the results page, the final forecast, how it works and the state legislatures; web/data/ holds the published outputs
tests/	the test suite: every module imports, the early-vote profile and exposures, release dates, simulation pieces
data/static/	small inputs derived once from public sources (below); committed
data/state/	poll records the daily run accumulates (the feeds only reach back a few weeks); committed
data/manual/	hand-entered polls (releases not yet in any feed) and checked corrections to scraped polls, with source URLs
data/early/	early-vote diagnostics, written nightly until election night; not used by the forecast or the website
data/final/, data/results/	the frozen forecast (final.py), the results as they come in and the scorecard (scorecard.py)
data/raw/, data/cache/	downloads and intermediate files; git-ignored, kept between runs by actions/cache
reports/	weekly change reports
site_integration/	reference copy of the website's sync workflow

Data sources and licences

Committed inputs (data/static/):

file	contents	source / licence
district_groups_2026.csv, state_groups.csv, district_wnc_2026.csv	citizen voting-age population by Hispanic / Asian origin group, white non-college share, per 2026 district and state	U.S. Census Bureau ACS 2020-2024 5-year tables + PUMS, the block equivalency file of the map in force (demog.cd_blocks: the 120th Congress, Missouri on its 119th-Congress lines after the Supreme Court blocked its 2025 map); built by demog.py, the Missouri rows from the Census API's district tables (demog.patch_cd119)
district_heat_oil_2026.csv, state_heat_oil.csv	share of households heating with fuel oil / kerosene	ACS B25040 (public domain); demog.heat_oil_shares, Missouri as above
urbanization_2026.csv	urbanization index on the 2026 maps	538 urbanization index 2022 (CC BY 4.0) carried to the maps in force with Census block-group populations; urban.py
early_share_2022.json	each state's share of 2022 ballots cast before Election Day	U.S. EAC 2022 EAVS (public domain); early_vote.py
early_profile.json	the timing profile of pre-Election-Day ballots (W, P of h(tau) = (1 - tau/W)^P) with the derived shares it was fitted to	UF Election Lab national early-vote tables (2024 via Wayback Machine captures; shares only, the tables are not redistributed); python -m midterms.early_vote profile
fec_june30_{2018,2022,2026}.csv, fec_senate_june30_2026.csv, fec_reports_2026.csv, fec_nominees_2026.csv	candidate fundraising through June 30 (House nominees; Senate nominees for the funded-independent rule)	FEC / OpenFEC (public domain); the House files are built by fec_money.py
house_national_vote_1946.csv	national House vote by party since 1946	Wikipedia election pages (CC BY-SA 4.0)
national_mood_fit.json	fitted coefficients only: the national-vote error s(L) (and the unused directional and approval variants), the cycles used and the late-movement slope	fit of the generic-ballot average to the House vote, 15 cycles 1996-2024 (538 / ABC News averages and poll archives, HuffPost Pollster archives, Gallup approval). The underlying table is not distributed; rebuilds the fit from it

web/data/districts_2026.topo.json (district shapes) is built by district_shapes.py from Census 2020 cartographic block groups and the block equivalency file of the map in force (demog.cd_blocks; public domain).

data/state/ holds poll toplines as posted by the feeds (Polling USA, pollresults.org - poll data from The New York Times, CC BY 4.0), race_poll_seen.csv (the date each race poll first appeared on the forecast page), the Center Square generic ballot series used to confirm held polls (source_watch.py) and pollster_scores.csv (a numeric score per pollster from pollresults.org's letter grades; it weights the national generic-ballot and approval polls, not race polls; the grades themselves are not committed).

Downloaded at run time (never committed): the 538 poll archives and data repository files (CC BY 4.0), FEC results workbooks, MIT Election Lab presidential returns (CC0), VoteHub, Wikipedia (CC BY-SA 4.0), Bluesky posts, pollresults.org, UF Election Lab early-vote counts (CC BY-NC-ND 4.0 - used only to compute each state's scaling, never republished), EIA heating-oil prices (public domain), Florida Division of Elections and Pennsylvania Department of State files (and, when switched on, the Texas Secretary of State's 2026 early-voting county files). Race ratings from Cook, Sabato and Inside Elections are read from Wikipedia's ratings tables and shown for comparison only. Every poll belongs to its pollster.

Fetching politely

fetch.py sends one identifying User-Agent, scorvec.com midterm model (+https://scorvec.com/midterms/about.html), which the bootstrap, registration and early-vote downloads also use. Pages are fetched one at a time with pauses, nothing already held is downloaded again, and fetch.py prints its request and byte totals per host at the end of the run.

source	per daily run
Wikipedia (~130 race and table pages)	the MediaWiki API's current revision ids, 50 pages per request; only pages edited since the cached copy are downloaded; historical results pages are fetched once
pollresults.org race pages (~220)	the sitemap (1 request); a page is downloaded only when its lastmod is on or after the cached copy
Bluesky (Polling USA, Political Poll Bot)	posts are cached; each feed is read only back to the first post already held
VoteHub API (5 poll types)	5 requests, a few MB: the API has no date filter or ETag
EIA heating oil, UF early vote, Florida statistics files	conditional requests (ETag / Last-Modified): unchanged files come back as 304
Texas SOS early-voting turnout	off unless TX_EARLY=on; then the election index and one county file per early-voting day, each at most once a day (conditional requests); per-voter rosters are never requested
Pennsylvania mail ballots (data.pa.gov)	2026 aggregates only; 2022 and 2024 are queried once and kept
NC / PA registration (Mondays)	new weekly NC snapshots only; past years' files are kept
538, FEC, MIT inputs (bootstrap)	once, then kept
OpenFEC	never in the daily run; a manual rebuild only (cached per candidate)

The one-time builders (demog.py, urban.py, district_shapes.py, cd_lean_build.py, stateleg_build.py, fec_money.py) need large downloads (ACS summary files and PUMS, Census geography, precinct returns; several GB) or an API key and are not part of the daily run; their docstrings list the exact inputs. The geographic builders also need GeoPandas (the state-legislature and district-inputs.yml workflows install it), and district_shapes.py runs mapshaper through npx. Every 2026 district input reads the map through demog.cd_blocks, which places the states whose adopted map a court blocked (demog.MAP_FROM_CD119: Missouri) on their 119th-Congress lines; cd_lean_build.MAP_STATUS holds the dated sources and a test keeps the two in step.

Secrets

secret	used by	needed for
OPENFEC_API_KEY	fec_money.py, only on a manual run with fec_refresh	re-pulling June 30 fundraising (an api.data.gov key)
CENSUS_API_KEY	demog.patch_cd119, only in district-inputs.yml	the demographic tables of a state back on its old lines, from the Census API (a free api.census.gov key)

The daily run needs no secret; it commits with the workflow's own GITHUB_TOKEN.

Model notes

The sections below record what was measured behind the main choices, including what was tested and left off. The switches named here are module constants or the environment variables listed above; the comment at each definition carries the same evidence.

National environment and error

• No directional correction (national_mood.DIRECTION = False). Generic-ballot averages overstated Democrats in 13 of the 15 House elections 1996-2024, but most of that record comes from older averages built differently. Rebuilt with this estimator for the seven elections with poll-level data (2012-2024), the average overstated Democrats in 5 of 7 a month out, by anything from nothing to five points, so no direction is assumed. An approval-adjusted variant (national_mood.APPROVAL) is built and off.

• House effects shrunk lightly (generic.LAMBDA_HOUSE = 1 rather than 4): predicting the next 24 days' polls on 538's generic archive 2018-2024 at 7 dates, weighted RMSE 2.623 -> 2.551 (23 of 28 cases, 4 of 4 years; not significant on its own, year-block t 1.95).

• Poll weights: quality, age and a cap per pollster (generic.QUALITY_GAMMA = 2, AGE_HALF_LIFE = 60, SHARE_CAP = 0.15). Same harness (538's generic archive 2018-2024, 7 dates a cycle, 3-day release lag): weighted RMSE of the next 24 days' polls and the trend's miss of the national House margin, 2.576 / 3.523 without the three. None of the differences below is significant (year-block t on four elections, 3 df).

  change	next 24 days' polls (better in: dates, years; p)	margin vs House vote (better in: dates, years; p)
  quality weight (score/3)^2	2.574 (12/28, 2/4; 0.78)	3.388 (14/28, 2/4; 0.52)
  age weight, 60-day precision half-life	2.568 (14/28, 1/4; 0.77)	3.492 (18/28, 3/4; 0.12)
  15 % share cap	2.600 (12/28, 0/4; 0.07)	3.437 (16/28, 3/4; 0.64)
  all three	2.565 (15/28, 2/4; 0.72)	3.436 (14/28, 2/4; 0.78)

  The quality weight scales each pollster's house-effect ridge by the same factor squared; without that, a down-weighted pollster's house effect is also pulled toward zero and the next polls get worse at every exponent (2.70 at 2, worse in 4/4 years). Exponents 1, 3 and 4 score 2.572 / 2.580 / 2.596 and 3.452 / 3.337 / 3.286; 2 takes most of the margin gain before the next polls start to slip, and scores best on the next polls with the other two on. The age weight discounts the trend's smoothness penalty by the same factor, so the trend through past months keeps its shape (with a fixed penalty, year-old polls at 1 % of their precision flattened the 2026 trend by up to 2 points, approval by 6). Half-lives of 30, 45 and 90 days score 2.583 / 2.568 / 2.568 with the other two on; a looser trend (LAMBDA_TREND 10 instead of 30: 2.564 / 3.466 alone, 2.600 / 3.373 with the other two) or weekly knots (2.566 / 3.417, 2.611 / 3.368) lean on recent polls by extrapolating their slope instead. The margin gain of the three together comes from 2018 (3.09 -> 1.94) and 2020 (4.00 -> 3.80); 2024 is worse on both counts (next polls 2.19 -> 2.50, margin 4.10 -> 4.51). The cap is a judgement rule, not a backtest result. A pollster's share is the summed positive influence of its polls on today's value (the poll page's "influence on today"), and over the cap all its polls' precision is scaled down until it sits at the cap. Morning Consult's weekly tracker carried 0.36 of the generic trend on 17 October 2026, and with the quality and age weights on, the leading pollster of a backtest date carries 0.19-0.83. The cap binds at all 28 backtest dates, on 2-7 pollsters at a time, and moves the trend by -0.9 to +1.2 points. On 17 October 2026 the three moved the generic trend from D+8.6 to about D+9.6 and approval from -21.1 to about -22.9, and Morning Consult's share from 0.36 to 0.15 (generic) and from 0.25 to 0.10 (approval).

Governor model review

Governor races are the least nationalized of the three offices, so the fundamentals prior matters most there. The harness (python -m midterms.gov_backtest all) scores every even-year cycle 2006-2024 at September 1, September 15, October 1, October 15, November 1 and the day before the election (2006-2016 from October 1 only: 538's raw_polls, the poll source for those years, holds the last ~60 days; 2018-2024 use 538's full governor poll archive). It runs the live path - gov2026.Prior, prepare (sponsor shifts and pollster leans fitted on earlier cycles only), collapse_versions, the robust blend and gov2026.simulate on the shared national draw - with only what was known on the date: polls three days after their last field day, the prior fitted on the governor races of earlier cycles (1998 onward), and the generic-ballot average known at that lead (538's average to 2016, this estimator from 2018; no directional correction). The 2026-only Hispanic / Asian loadings and the heating-oil term are left out. n = 234 races (every D-v-R race of each cycle, polled or not), 1,086 race-dates, 10 cycles; 984 race-dates had polls. Scores: Brier and log loss of P(D wins), error of the race mean, the CRPS of the margin, 80 % / 50 % interval coverage. Each entry reads difference (cycles better; permutation p / bootstrap p), with an exact sign-flip permutation of the cycle means (2^10 assignments) and a race-cluster bootstrap stratified by cycle (4,000 draws).

Baseline (the prior before these terms): log loss 0.1944, Brier 0.0573, margin MAE 6.11 and RMSE 8.59 points, CRPS 4.39, 80 % intervals cover 76.1 % and 50 % intervals 45.1 % (too narrow), PIT KS 0.121.

variant	log loss	Brier	CRPS	margin MAE	80 % cov.
baseline	0.1944	0.0573	4.390	6.11	0.761
successor incumbents coded as incumbents (parser fix)	-0.0044 (7/10; 0.129 / 0.059)	-0.0017 (7/10; 0.152 / 0.062)	+0.035 (3/10; 0.248 / 0.183)	+0.03 (0.469 / 0.372)	0.758
lean slope with a time trend	-0.0050 (6/10; 0.061 / 0.018)	-0.0015 (7/10; 0.070 / 0.070)	-0.230 (7/10; 0.025 / 0.000)	-0.32 (0.023 / 0.001)	0.789
recency weights, half-life 12 years	-0.0006 (6/10; 0.484 / 0.266)	-0.0001 (5/10; 0.676 / 0.578)	-0.039 (9/10; 0.064 / 0.017)	-0.06 (0.041 / 0.004)	0.762
recency weights, half-life 8 years	-0.0007 (5/10; 0.553 / 0.377)	-0.0001 (5/10; 0.871 / 0.757)	-0.053 (9/10; 0.072 / 0.028)	-0.08 (0.037 / 0.006)	0.762
fit without races with a 15 %+ third candidate	-0.0007 (3/10; 0.666 / 0.354)	-0.0003 (3/10; 0.484 / 0.249)	-0.021 (6/10; 0.184 / 0.103)	-0.02 (0.328 / 0.228)	0.766
previous governor race residual	+0.0007 (7/10; 0.965 / 0.569)	+0.0002 (7/10; 0.938 / 0.729)	-0.049 (8/10; 0.160 / 0.086)	-0.07 (0.096 / 0.092)	0.770
previous race residual, own v inherited	-0.0009 (7/10; 0.383 / 0.468)	-0.0004 (7/10; 0.312 / 0.467)	-0.033 (7/10; 0.441 / 0.338)	-0.04 (0.459 / 0.406)	0.761
incumbent's own previous residual	-0.0006 (7/10; 0.518 / 0.286)	-0.0003 (8/10; 0.314 / 0.196)	-0.023 (7/10; 0.562 / 0.469)	-0.03 (0.555 / 0.482)	0.761
prior sd: open v incumbent	+0.0007 (5/10; 0.268 / 0.059)	+0.0002 (6/10; 0.471 / 0.254)	+0.015 (1/10; 0.018 / 0.106)	+0.02 (0.021 / 0.152)	0.748
prior sd grows with distance from 0	+0.0021 (3/10; 0.105 / 0.033)	+0.0005 (4/10; 0.439 / 0.221)	+0.055 (2/10; 0.029 / 0.005)	+0.07 (0.035 / 0.009)

Adopted (gov2026.PRIOR_SPEC = {"lean_t": True, "succ": True, "qual": "all"}; {} restores the baseline prior):

• The lean slope's time trend (lean x (year - 2010) / 10). Governor races have nationalized: the per-cycle lean slope rose from 0.1-0.2 (1998-2006) to 0.43-0.54 (2018-2022); a pooled slope (0.36) is too flat for a current race and compresses every prior toward 50-50. Fitted 1998-2024: 0.277 + 0.110 per decade -> 0.45 in 2026 (one cycle beyond the data, the same reach the walk-forward test made every cycle).
• Successor incumbents as their own term. Governors who succeeded mid-term and then ran (Ivey 2018, Reynolds 2018, Hochul 2022, McKee 2022, Parson 2020, Quinn 2010 ...) had been read as open seats by the history parser; they are neither open seats nor full incumbents. Fitted: 10.8 points for an elected incumbent, 6.9 for a successor. Simply recoding them as incumbents (INC_FIX) scored the same within noise and is not used.
• Nominee experience, D minus R (+1 / 0 / -1): a nominee is experienced if, before the race, he or she held a statewide elected office (governor, lieutenant governor, attorney general, secretary of state, treasurer, comptroller, auditor, superintendent, an elected state commission) or a seat in Congress, read from each nominee's Wikipedia infobox (gov_quality.py); a sitting governor counts as experienced, and a nominee without an article as not. Symmetric in party by construction. Fitted: 6.1 points.
• Together against the baseline: margin MAE 6.11 -> 5.76; 80 % coverage 76.1 -> 79.6 %, 50 % coverage 45.1 -> 47.2 %, PIT KS 0.121 -> 0.109; better at every date; unpolled race-dates' MAE 14.2 -> 11.5.
• 2024 is in the history (gov2026.USE_2024, data/static/gov_results_2024.csv) and is the harness's tenth test cycle.

About thirty variants were scored, and with ten cycles the smallest attainable permutation p is 0.002, so no single result survives a strict multiple-comparison correction. The adopted terms are the ones that pass on both tests, help in most cycles and have a reason (a measured trend, a coding error, a long-established predictor); everything else stays off.

Measured, not used:

• Governor poll errors, measured as the Senate's were (538 raw_polls, last 60 days, races with 3+ polls): within-race sd 4.89 v 4.63 for the Senate, race-average systematic 5.73 v 5.32. Scaling the Senate's 4.2 / 5.0 by those ratios made every score worse (CRPS +0.06, p 0.016 / 0.001), so the governor blend keeps the Senate's poll errors.
• The shared miss: the cycle-mean governor polling miss (non-partisan polls, last 21 days, 1998-2022) has RMS 3.5 points against the Senate's 2.8, and the two offices' cycle means correlate 0.90. The governors' shared statewide shock (senate2026.GOV_SHARED_MISS 4.2, scaled by the governor national slope) is left as it is: x0.5 and x1.5 each lose on log loss or coverage.
• Undecideds do not split in proportion to the candidates' shares: on 538's governor polls with cycle fixed effects they lean to the trailing candidate (coefficient -0.19, se 0.04, in the last 21 days). Shifting polls accordingly is tested below and left off; undecideds stay a symmetric variance term.
• Results run more lopsided than the race means: after removing each cycle's shared miss the winner beats the mean by about 2 points (both parties' favourites). The lean trend removes part of it; a symmetric pull of the polls toward the state's fundamentals was not significant.
• Not testable here: Alaska's ranked-choice count (the harness scores Alaska 2022 on its first round) and strong independents (too few governor cases since 2006).

Senate prior and errors

• Time-varying prior (senate2026.SEN_CONST, B_LEAN, C_INC, NAT_SLOPE, PRIOR_SD). Per-cycle fits show the incumbency coefficient falling about 3.8 points per decade, the lean slope rising about 0.11 and the residual sd falling about 1.35. The prior has linear time trends in the lean and incumbency slopes, fitted on every D-v-R race 1982-2024, the national slope of the last 10 cycles and a trend-extrapolated residual sd; at 2026: constant 0.66, lean 0.837, incumbency 6.7, national 0.795, sd 8.8. Against the pooled 10-cycle fit on the walk-forward harness (Senate 2006-2024, 10 cycles, 1,615 race-dates): log loss -0.0079 (7/10 cycles; sign-flip p 0.061, race bootstrap p < 0.001), Brier -0.0025, CRPS -0.23; 2014-2024 better in 6/6 cycles, worse in 2006, 2008 and 2012. The gain is largest in cycles whose polls overstated Democrats; unpolled races show the same time pattern.
• Shared and race errors (STATE_SHARED_MISS 3.1, RACE_EXTRA 2.83, SEN_S_NAT). Two settings tuned separately on the harness, which do not reconcile arithmetically and are documented as such in senate2026.py: each race's blend nets a lead-dependent shared part (SEN_S_NAT x the national slope, about 2.3 points three weeks out) out of the measured midterm race-average miss of 4.0, keeping about 3.3 as its own systematic error; the simulation then draws a fixed 3.1-point shared shock (the RMS cycle-wide statewide miss 1998-2022, about zero) and adds 2.83 points of race-level spread, so a race's systematic spread is about 5.3 points. Against a single 4.2 shared shock (which implied a correlation between race errors of 0.34 against 0.13 measured) race scores were unchanged and the seat-count PIT sd moved 0.265 -> 0.299 (uniform 0.289); the deep review had found race-level errors larger than modelled. Governors keep GOV_SHARED_MISS 4.2 with no extra.
• Midterm poll error (MIDTERM_POLL_SYS_K 0.8): the shared per-race Senate polling miss is smaller in midterms (4.1 v 4.7 in presidential years); scaling it by the measured ratio improved log loss in all five midterms 2006-2022 and is applied in midterm years only.
• Common trend (SEN_COMMON_TREND, on): how all of the year's Senate polls have moved together, estimated with each race's own level and each pollster's lean, shifts every older poll by the shared movement since it was taken. Walk-forward 2006-2024: log loss -0.0035 (8/10 cycles). The same adjustment taken from the generic-ballot average, or applied to governor races, made the forecasts worse, so it is used for the Senate only.

Independent challengers

Four 2026 Senate races have an independent as the main challenger: Nebraska, Idaho and South Dakota with no Democrat on the ballot, Montana with one (senate2026.IND_SHIFT, IND_SHIFT_SD, IND_MIN_RATIO, SPLIT_PRIOR_MULT). Measured on every Senate race since 1988 and every governor race since 1998 with no Democrat on the ballot (MIT and Wikipedia results; Sanders and Mississippi's 2008 special, Democrats in all but label, aside), against the D-v-R prior fitted without the race's cycle (the Senate form on the cycles before it; the governor prior on the other cycles), at the actual national vote:

race	June 30 money (x the Republican's)	independent minus Republican	prior	above the prior
Orman, Kansas 2014	$0.63M (0.45x)	-11.1	-38.5	+27.4
McMullin, Utah 2022	$3.25M (0.85x)	-10.9	-31.6	+20.8
Osborn, Nebraska 2024	$1.58M (1.00x)	-6.7	-31.0	+24.3
Walker, Alaska governor 2014	(state filings)	+2.3	-30.7	+33.0
Harrington, Arkansas 2020	$0	-33.1	-31.7	-1.4
O'Hara, Mississippi 2002	none	-69.2	-32.3	-36.8
Spannaus, Virginia 1990	none	-63.3	-21.9	-41.5
Steve Osborn, Indiana 2006	none	-74.8	-25.7	-49.1

• Prior, no Democrat on the ballot. Being an independent explains nothing (all eight: mean -2.9, sd 34); money does: the four with a campaign finished +26.4 above the prior (sd 5.2, 95 % 18 to 35), the four without -1 to -49. "Funded" is at least a quarter of the Republican's June 30 money, with no dollar floor: the earlier $1M floor would have filed Orman with the unfunded. The record places the cut only between no money and Orman's ($0.63M, over $0.8M in 2026 dollars; 0.45x): Nebraska ($5.5M, 0.74x) is well clear, South Dakota ($0.78M, 0.58x) and Idaho ($0.60M, 0.29x) fall in that untested range and now qualify. Early polling strength separates the groups too, but the polls already enter the blend, so the prior does not condition on them. On the live-path backtest (next point) dropping the floor touches one race, Kansas 2014, and scores worse there (log loss +0.11 two to four weeks out): Orman led his final polls by 3 and lost by 11, with Kansas's governor polls 6 points off the same way.
• Size of the shift. The prior moves by 12, not the measured +26. Live path (that prior at the generic-ballot average, each race's own polls, leave-one-race-out, the four funded races): shifts 0 / 12 / 26 score log loss 0.507 / 0.551 / 0.617 two to four weeks out and 0.636 / 0.549 / 0.503 five to nine weeks out. Late in 2014 and 2024 these races' polls ran ahead of the results, as those cycles' D-v-R polls did, so a prior further below the polls scored better late; four races in three cycles cannot settle it. The shift sd 9 (prior sd 12.6 in all) is wider than the group's spread (5.2, 95 % 2.9 to 19).
• Polls. In the four funded races the non-partisan average of the final three weeks missed by +2.8 for the independent (Kansas +11.4, Alaska -1.2, Utah -2.0, Nebraska +2.9; se 3.1): +0.1 (se 3.0) net of the same cycle's D-v-R statewide polling miss and -1.2 (se 3.1, three races) net of the same state's D-v-R race that year (Kansas governor 2014 +6.4, Nebraska's other Senate race 2024 +6.7). Their spread (RMS 5.1 net of the cycle) is that of D-v-R races in those cycles (3.8-6.4). The independent's margin fell from the 22-60-day average to the final three weeks in all four (mean -4.2, se 1.7), -1.2 (se 0.7, three races) net of the cycle's D-v-R movement. Undecided and minor-party voters went 67 % to the Republican between the final polls and the result, against 71 % in D-v-R races in red states (lean below -15; 168 races). So there is no independent poll bias, extra variance or late fade to add: these misses follow the cycle and the state, which the shared statewide shock already carries.
• Sponsors. An independent's own polls ran +5.5 for the independent (se 2.1; 13 polls, 4 races) against the same race's non-partisan polls within 21 days, like the Democratic sponsor effect (+4.3) the model removes from them. Democratic-sponsored polls in independent races ran -3.0 (se 2.8; 12 polls, 6 races; race-level se 4.0) and keep the independent's-side tag: too few races to drop it.

Tested and left off

Same-state Senate and governor errors (senate2026.JOINT_ON = False). Within a cycle, a state's Senate and governor polling errors correlate 0.58 (95 % 0.46-0.68; 149 state-years, 13 cycles). With the switch on, each race's own error is split into c x a draw shared by the state's statewide races plus sqrt(sd^2 - c^2) x its own draw (c = 3.68 points on 1998-2022, capped at 0.9 sd), so only the covariance moves. Joint test: the governor harness paired with the Senate harness, state by state, one joint simulation per pair on the shared national draw; n = 682 pair-dates, 142 state-years, 10 cycles:

	off	on	difference (cycles better; permutation p / bootstrap p)
log loss, 4 outcomes	0.3325	0.3347	+0.0023 (5/10; 0.73 / 0.76)
Brier, 4 outcomes	0.1793	0.1813	+0.0020 (4/10; 0.47 / 0.40)
Brier of P(split ticket)	0.0775	0.0792	+0.0016 (5/10; 0.53 / 0.50)
mean P(split) (actual 24.9 %)	27.4 %	26.1 %
simulated correlation of the pair	0.31	0.56	(backtest residuals: 0.32)

No gain: the shared national draw already gives a state's two races a correlation of 0.31, and the backtest residuals show 0.32. The simulated national shock is larger than the realized cycle-wide miss, so it already carries the state part. Each office's seat distribution is a sum over one office's races, so a cross-office covariance cannot move the headlines.

Undecideds toward the trailing candidate (model.UND_TRAIL_K = None). On, each race poll's margin moves by k x margin x (100 - D - R) / (D + R) before the sponsor lean is removed, with k from model.und_trail_fit(before=cycle) (538 raw_polls, Senate and governor, non-partisan, last 60 days, cycle fixed effects; -0.22 when fitted on every cycle). Against the model as it is:

harness	n	log loss	Brier	CRPS	margin MAE	80 % cov.
governors	234 races, 1,086 race-dates	-0.0003 (2/10; 0.99 / 1.00)	+0.0001 (3/10; 0.51 / 0.72)	+0.18 (2/10; 0.008 / <0.001)	+0.24 (0.014 / <0.001)	79.6 -> 76.0 %
Senate (walk-forward prior)	326 races, 1,528 race-dates	+0.0031 (1/10; 0.006 / <0.001)	+0.0006 (2/10; 0.031 / 0.002)	+0.26 (0/10; 0.002 / <0.001)	+0.34 (0/10; 0.002 / <0.001)	79.6 -> 77.1 %

It makes the race means worse in both offices: they are already too close to 50-50, and the switch pulls them further in. It is not a party lean (mean shift of the race means +0.03 points for governors, +0.05 for the Senate).

Pollster shared error (model.POLLSTER_SHARED_ERROR = None). A pollster's house error is shared by all of its polls in a race, so four polls from one firm are not four independent reads. With tau set, each poll's error in the robust blend splits into its own part and a pollster-race part common to that firm's polls, integrated out exactly; each poll's marginal variance is unchanged. tau is fitted on 538's raw_polls (race_poll_calibration.pollster_shared_sd: same-pollster v different-pollster pairs at the same time gap, lean-corrected, net of sampling): tau = 2.8 points. On the leak-free backtests (Senate 2018-2024, governors 2018-2022, House 2018 and 2022; six dates from September 1 to November 1; tau from earlier cycles) it changed nothing measurable: log loss Senate +0.0002 (95 % interval -0.0014 to +0.0016), governors -0.0006 (-0.0024 to +0.0008), House +0.0001 (-0.0001 to +0.0003).

Directional state corrections (model.STATE_DIRECTION = False). State polling misses persist net of the national miss (next-cycle slope 0.24, 538 raw_polls 1998-2022) and statewide polls ran too Democratic in states with more white non-college voters (1.17 points per state sd). Both are measured and implemented; like the national rule, the live model assumes no direction and widens the state-level error instead.

Heating-oil adjustment

A judgment term, not backtested (no past cycle had a comparable price shock): the change in retail heating oil over the past year, times each district's (state's) share of oil-heated homes, times a typical household's 650 gallons, at 0.78 generic-ballot points per dollar of gasoline spread over 750 gallons, times the part of the change not yet in the race's polls, against the president's party (model.heating_oil_shift; a fall enters with the opposite sign). The unpriced part is race-specific (model.unpriced_change): the change since the race's newest poll was fielded counts in full, the change before it at half (HEAT_UNPRICED); a race without polls takes half of the whole change. The price is EIA's weekly Maine No. 2 heating oil residential price; EIA's survey runs from October to March, and outside the season the last published week is held. The reconstructed history evaluates it at each past date.

State legislatures (experimental)

stateleg.py forecasts control of 22 chambers - both chambers in Michigan, Minnesota, Wisconsin, Arizona, Pennsylvania, New Hampshire, North Carolina, California and Kentucky; the Georgia, Iowa and Texas Houses; and the Texas Senate - inside the daily run (MIDTERMS_STATELEG, default on; workflow input stateleg). It writes state_legislatures into web/data/model.json and the per-seat file web/data/stateleg_seats.json; web/legislatures.html shows them.

Seat model. Each seat's expected Democratic two-party margin is a + beta * lean + c * inc + g * E: the district's 2024 presidential margin minus the national one (lean), incumbency (+1 / -1 / 0) and the model's national environment E. The simulation adds g x the House's national draw, a state error shared by both chambers (partly chamber-specific) and a t5 seat error. Multi-member districts (AZ House, NH House) elect their top-k candidates (party share +- the candidate's own deviation from its slate, sd fitted, plus an incumbency bonus); a party with fewer nominees than seats can win at most that many. Seats with only one major party on the ballot are fixed. Staggered senates (PA and WI: odd districts up; CA, KY and TX: the seats elected in 2024 are held over) add the held-over seats by their current holders. Ties: the MI and PA senates go to the party of the 2026 governor winner (the lieutenant governor), the NC Senate to the Democrats (Lt Gov Hunt), elsewhere counted as shared.

Fitted parameters (data/static/stateleg/params.json; python -m midterms.stateleg fit): incumbency c = 6.0 points, national slope g = 0.77, 2-year state swing error 4.8 (state error 3.4), between-chamber correlation 0.62 - Klarner returns 1972-2022, 42,851 consecutive same-map district pairs; lean slope beta = 0.91, intercept a = 2.9, seat error 5.8 - 2018 (2016 lean) and 2022 (2020 lean) results; slate sd 8.6 % of the party mean (11,276 multi-member slates).

Backtest (python -m midterms.stateleg_backtest, data/static/stateleg/backtest.json): 2018 and 2022 on October 1, Klarner parameters from earlier elections only, the lean slope from the other cycle, E = the 538 generic-ballot average with no correction, who is on the ballot from the candidate records. The original 14 chambers (24 chamber-elections with full district data): control Brier 0.115 against 0.250 for "the current majority holds", called right in 20 of 24 (the current-majority rule: 18), seat count inside the 80 % range 23 of 24. The Georgia, Iowa and Texas Houses (5 chamber-elections with complete district data): all called right, as the current-majority rule also was; Brier 0.007; seat count inside the 80 % range 5/5. The California and Kentucky chambers and the Texas Senate (10 chamber-elections): all called right, as the current-majority rule; seat count inside the 80 % range 8 of 10. The daily run publishes these per-group scores from backtest.json (stateleg_backtest.published, state_legislatures.backtest in model.json), and about.html and legislatures.html quote them from there. No published forecaster's chamber calls were scored (none available under a usable licence).

State-poll signal (tested, not adopted; MIDTERMS_STATELEG_KAPPA=on restores it). The legislative state swing residual tracks the state's statewide polling miss (slope -0.31, se 0.05; 360 state-cycles 1998-2020), and a fitted link kappa x the state's Senate / governor surprise (kappa 0.074, se 0.024) changed seat log loss by +0.0005 (2018) and +0.0003 (2022) and the original 14 chambers' control Brier 0.115 -> 0.119 (arm "state" of backtest.json): no measurable gain.

Coverage. stateleg_scan.py reads every 2026 chamber page (data/static/stateleg/coverage_scan.csv) and checks the candidates (stateleg_scan.check -> pending_check.csv): a chamber is added only when the page has a general-election slate for every seat up, the seats up match the 2024 rule, every district has a lean on the current map, and the one-party seat count is plausible against 2024 (MEDSL) and 2022 (Klarner). Held back: CT Senate and House (2024 MEDSL party coding under fusion voting gives no usable comparator), UT and WV Senates (one-party seats far below 2022 / 2024), IA Senate (a district missing from the page), WA and OR Senates (lean check failed: rms 4.0 and 8.9).

Approximations. NH House: candidates are not listed on Wikipedia, so every seat is contested by full slates, floterial districts are simulated as ordinary multi-member districts, and the 2018 backtest has no NH House. MI Senate: no 2024 Senate election, and the 2026 court-ordered redraw of the Detroit-area districts is not in the Census TIGER files yet; leans are VEST 2020 precincts moved to 2024 by each House district's measured swing on the 2022 map (the same method reproduces known 2024 Senate leans with rms 0.3-1.8 in PA, GA, IA, TX, NC). PA odd Senate seats use the same method on the current map. MN Senate candidates are not on Wikipedia: incumbents are assumed to run unless listed as retiring. Georgia House nominees come from the Georgia Secretary of State's official primary and runoff results (ga_sos.py), cross-checked against Wikipedia and Open States (ga_house_crosscheck.csv).

Data (stateleg_build.py, run in GitHub Actions by stateleg-data.yml; downloads cached, never committed):

file	contents	source
stateleg/lean_2026.csv	2024 president by 2026 district (MEDSL precinct rows joined to the same precinct's legislative labels; MN / WI senates from nested House districts; AZ House = LDs; MI Senate / PA odd seats as above) + a spatial cross-check	MEDSL 2024 precinct returns (CC0), VEST 2020 (CC BY 4.0), TIGER (public domain)
stateleg/lean_hist.csv	2016 president on the 2018 maps, 2020 president on the 2022 maps	VEST 2016 / 2020 x TIGER 2018 / 2022; NH House by town names x MEDSL 2022 labels
stateleg/medsl_results.csv.gz	candidate-level legislative results 2022, 2024	MEDSL (CC0)
stateleg/klarner.csv.gz	Klarner returns aggregated to district-elections, all states 1972-2022	Klarner, doi:10.7910/DVN/FJOGJB (CC0)
stateleg/openstates_current.csv	sitting legislators	Open States people (CC0)
stateleg/qc.json, stateleg/geo_audit.json	coverage, cross-checks, map-change audit; the district-map join audit (the map build fails on a gap)	-
web/data/stateleg_geo/{ST}.topo.json, index.json	district maps, one TopoJSON per state (stateleg_geo.py, stateleg-geo.yml)	U.S. Census Bureau TIGER/Line 2025 state legislative districts (public domain); large water cut with the Census 2024 cartographic state outline; NH floterials from MEDSL 2024 precinct labels

Candidates and retirements are read each day from the Wikipedia 2026 chamber pages (stateleg_wiki.py, cached like every page). On the maps, districts are shaded by the chance a Democrat wins the seat (multi-member: the expected Democratic share), hatched where only one party is on the ballot and dotted where the seat is not up in 2026. Boundaries are the TIGER/Line 2025 districts (the maps in force for 2026), except the Michigan Senate's court-ordered Detroit-area redraw, which is flagged on its map; each New Hampshire floterial is drawn as the outline of the base districts whose towns vote in it.

District lean (the House prior)

cd_lean_build.py builds the House prior's district lean from public precinct returns: 0.75 x the most recent presidential two-party margin minus the nation's + 0.25 x the same for the election before, using only elections held before the cycle

• the 2026 maps from 2024 and 2020 (cd_lean_2026.csv), the maps of the 2022 election from 2020 and 2016 and of the 2018 election from 2016 and 2012 (cd_lean_hist.csv, on which the prior is fitted). The national margins are the 50-state sums of the same tables. Every state is calibrated to its official totals, each nominee's votes on every ballot line (MIT's party code leaves out New York's fusion lines through 2016).

2026 maps. Redrawn states are found from the Census 119th v 120th Congress block equivalency files. Unchanged states use MEDSL 2024 precinct rows by their U.S. House label, where the file matches the official state totals and the presidential vote sits on labelled precinct rows. Redrawn states (and files that fail those checks): VEST 2020 precincts -> 2020 blocks -> 120th-Congress districts, carried to 2024 by the county-by-district swing on the labels (qc_cd.json lists every check and shift). The Census 120th-Congress block file does not follow court orders, so cd_lean_build.MAP_STATUS records the map in force per state with a dated source; Missouri uses its 2022 lines (the U.S. Supreme Court blocked the 2025 map on 2026-09-25). For comparison only, against 2 x Cook PVI: rms 1.09 points over all 435 (0.93 on unchanged maps, 1.29 on redrawn ones), mean +0.25, r 0.9994.

2018 and 2022 maps. VEST 2016 and 2020 precincts -> 2020 blocks -> the 118th-Congress block equivalency file (2022) or the TIGER/Line 116th-Congress districts by each block's internal point (2018). VEST has no 2012 files and MEDSL's county returns sit behind a Dataverse guestbook, so 2012 is the 2016 block votes moved by each county's 2012-16 swing and turnout change, with the county results from Wikipedia's state pages (every state's table matches its official totals within 1 %). Checked against The Downballot's published presidential results by district (a check only; their figures are not used or redistributed): 2016 on the 2018 maps rms 0.16 points (largest 1.7, MO-5), 2020 on the 2022 maps rms 0.23 (3.4, LA-6), 2012 on the 2018 maps rms 2.5 (10, TX-7: the county step cannot see swings inside the large metro counties), so the 2018 lean is within rms 0.63 (2.5) of the same lean from their figures. The 2016 vote alone would be off by rms 2.45 (7.6). The county step measured where both elections are known (2016 rebuilt from the 2020 blocks): rms 2.7 by district. 538's published leans for the same maps differ by rms 3.3 (2018 partisan lean, which also weights state legislative results; largest in West Virginia and Kentucky) and 2.4 (the PVI column of its 2022 urbanization index).

Fit (model.fit_prior, contested seats 2018 and 2022, n 795; the 538 / Cook leans it replaced in brackets): lean slope 0.918 (0.951), incumbency 4.47 (4.58), residual sd 7.35 (7.41); with fundraising (money.coefficients) 0.855 (0.889), 3.34 (3.51) and 1.31 (1.25) per unit of the log money ratio. Nothing about the prior is a stored constant: every run refits it, and the intercept k0 (the contested-margin intercept over the national House vote, data_prep.national_house_vote) and the prior sd come out of the same fit - 0.42 (2018 -0.79, 2022 +1.63) and 7.35 today. Until October 2026 the blend used a hand-copied prior sd of 6.7 and k0 of 0.4 from the earlier 538 / Cook fit, so the switch of lean had left the polls weighed against a prior 20 % tighter than the fit said. Fundraising still helps out of sample on the new lean (all-seat RMSE fit on one cycle and scored on the other: -5.6 % in 2018, -2.6 % in 2022).

Backtest. The live House path (model.house_blend, factor_update, simulate) on 2018 and 2022, each year forecast from the other's fit (slopes, k0, fundraising), 538's House polls as released, the generic-ballot average at the lead, 7 dates from August 1 to the day before the election (5,565 seat-dates):

lean	log loss	Brier	CRPS	margin MAE	80 % cov.	Sep 15, 2018 (235 won)	Sep 15, 2022 (213 won)
538 / Cook	0.1153	0.0353	3.642	4.98	0.854	221 (197-247)	212 (190-233)
ours	0.1113	0.0338	3.593	4.93	0.866	222 (198-249)	212 (189-233)

Ours is better in both cycles on all three scores (log loss 2018 -0.0070, 2022 -0.0009; race bootstrap p 0.13, Brier 0.055, CRPS 0.20): with two cycles not a significant gain, and not a loss. The seat count is inside the 10-90 % range at all 14 dates with either lean. On the inputs of 16 October 2026 the switch moved the House from 241.7 to 241.2 Democratic seats (P(majority) 0.894 -> 0.888): +0.7 seats from the lean with k0 refitted on both leans, -1.1 from replacing the stale k0.

Early vote

These ran until election night and were retired with the early-vote page on 4 November 2026. The trackers behind the tests of the early vote's skill (North Carolina, Florida, Pennsylvania, Clark County) keep their code and committed aggregates; the Georgia, Texas and Washoe County trackers, the turnout projection and North Carolina's election-night tracker were removed with their data. None of these fed the forecast. Votes cast before Election Day enter the model only through the early-vote movement split described above (national and race-level).

North Carolina, the tracker and the gate (nc_tracker.py): accepted mail and one-stop ballots of 2026 against the 2018 and 2022 midterms and the 2020 and 2024 presidential years on the same number of days out, by party, age, race, county (against each county's 2024 presidential margin) and whether the voters voted in the previous two generals; rates are per registered voter of the same group on that year's rolls (2018-2024 committed in data/static/nc_tracker/; 2026 from the morning's file, this week's registration file reduced again only when the board replaces it, and nc_early's cached voter history). The gate: the party mix of the early vote (D - R share of ballots banked by L days out, minus the same on the rolls, change from the previous cycle of the same kind) against North Carolina's statewide poll miss, 2012-2022: 5 scoreable cycles, r -0.72 at 17 days (p 0.17), -0.27 at 10, -0.22 at 5, the sign opposite to the usual reading. No skill, so it stays out of the forecast (data/static/early_gate.json); rerun after 2026.

North Carolina by race and ethnicity (nc_early.py; race and ethnicity are self-reported on the NC voter record): accepted mail and one-stop ballots by days before Election Day, 2026 against 2022 and 2024, for white, Black, other, undesignated and Hispanic voters - each group's share of ballots, that share against its share of registered voters on the same days out, ballots per 1,000 registered, the share who skipped at least two of the three previous generals, and what the same voters did in the previous general of the same kind. Inputs: the NCSBE absentee file of 2026 (one conditional request per run), the committed 2022 / 2024 baseline, NCSBE weekly registration statistics, and the 2016-2024 generals reduced from NCSBE's voter-history file (voter-level, so it lives only in the Actions cache, never committed). Output data/early/nc_early.json and the day's table data/early/nc/2026_<date>.csv.

Florida (fl_early.py; party only): the Division of Elections' daily statistics files are snapshotted into data/early/fl_stats/ (Florida keeps no history, so a missed day is lost) and compared with the Wayback Machine's 2024 captures at the same days out, as the returns' R-D margin, that margin minus the registration R-D of the latest month end, and minus the R-D of all mail ballots sent. Output data/early/fl_early.json. Hurricanes Helene (Sep 26, 2024) and Milton (Oct 9, 2024) fall on the 2024 baseline dates.

Pennsylvania (pa_early.py, daily run): mail-ballot applications and returned ballots by party at matched days out, 2026 against 2022 and 2024, from the Department of State's datasets on data.pa.gov; uploaded as a workflow artifact.

Credits

Model and code: Shawn Corvec. Poll data belong to their pollsters. Governor results and nominees' offices, and the 2026 calls and counts: Wikipedia (CC BY-SA 4.0). Ranked-choice transfer rates are derived from the Maine Secretary of State's and the State of Alaska Division of Elections' official tabulations. See web/about.html ("Sources and credits") for the full list. Not affiliated with or endorsed by any of the sources.

License

Code: MIT. Data keep the terms of their sources, listed above. Race polls taken from Wikipedia's race pages are used under CC BY-SA 4.0; the race poll averages in web/data/model.json are derived from them and published under the same licence. The 2026 calls and counts in data/results/results_2026.csv and web/data/results.json are taken from Wikipedia's election pages under the same licence, until official returns replace them.