HOUSE & SENATE · RESEARCH UPDATE · OCTOBER 20, 2026

What the model review found

Total receipts are now included provisionally. The wider review found several useful research leads, but no additional variable earned automatic inclusion.

Polling cutoff: October 19 · Historical comparisons use the same 15-day forecast horizon where comparable inputs exist.

What changed

I added a capped Senate adjustment based on total receipts through June 30. Comparable June reports cover 22 of the 35 races. The remaining races keep their existing estimates rather than receiving guessed financial values. House race estimates are unchanged; joint House–Senate outcomes include the revised Senate estimates.

This is a provisional judgment while the financial model is refined. Its historical margin improvement is small and statistically uncertain. The existing ranges are retained and are conditional on that adjustment. Financial formula, source amounts and race effects.

The most useful comparisons

Each row compares two models on identical races and elections. Average error is measured in margin points; lower is better. Levels from different rows should not be compared because their coverage differs.

Comparison	Elections	Existing → trial error	Decision
June total receipts beyond existing controls	2014–2024 · 6	4.87 → 4.84	Included provisionally; p = 0.47
Best full-cohort Senate combination: poll depth/age + lean + common correction	2016–2024 · 5	4.60 → 4.24	Research only; p = 0.25
Revised House incumbent prior-election context	2018–2024 · 4	4.77 → 4.67	Research only; p = 0.12
Pooled ranked-choice House performance in Senate model	2016–2024 · 5	4.60 → 4.70	No addition; larger error

The best combination is selected retrospectively from many alternatives, so its apparent gain is likely optimistic. Its adjustment for multiple comparisons does not pass. Probability, close-race and worst-election checks also matter; a smaller margin error alone is insufficient.

Alaska and Iowa

Race	Receipt adjustment toward D	Current projected margin	D win chance
AK	+0.85 points	R+1.72	37.9%
IA	-0.08 points	R+3.11	27.2%

Alaska: I checked Mary Peltola’s certified 2022 and 2024 House results, including the final ranked-choice rounds, and repaired missing historical candidate matches. Her adjusted House record is a strong personal-performance signal, but it does not translate point for point into a Senate bonus. The pooled version would add about 0.12 margin points to the current Alaska estimate; that formula worsened overall historical Senate accuracy, so it remains a comparison.

Iowa: Josh Turek outperformed the Democratic presidential margin by 13.32 points in the same 2024 precincts of his state House district. That is relevant evidence about the candidate. I do not yet have enough comparable state-legislature-to-Senate cases to estimate how much carries statewide. Ashley Hinson’s federal House history is included in the candidate comparisons. I have not forced an upward Iowa adjustment from one district result.

All variable families reviewed

Family	What I tested	Result
Polling-firm and sponsor weights	Partisan 0.5; grade bonus 1.25; combined; grades available at the time versus later hindsight grades. Generic-ballot weighting tested separately.	Small Senate gains; House results mixed. Keep as comparisons. Results
Fundamentals versus polls	Six strengths from zero to 1.25 times the existing contribution, plus choices based only on earlier elections.	Removing Senate fundamentals worsened margin and probability scores. No new weight selected. Results
2020 influence	Full, half and quarter fitting weights; all elections still scored equally.	Lower 2020 weights worsened average margin error in both chambers. Results
Bias-only alternatives	Common or firm bias correction, with partisan weights 0.5, 0.75 and 1.	Worse than the current model on the same historical races. Results
Presidential approval and midterms	Gallup level/net approval, dated-reading sensitivities, generic-ballot controls and fixed/learned opposition-party effects.	No reliable additional improvement. A valid current approval reading is not available in the saved series; no value is invented. Results
Candidate record and experience	32 structural comparisons of prior House, Senate and governor results; latest/pooled ranked-choice results; revised polling blends.	Identity matches repaired. No additional candidate formula qualified. Results
Incumbent prior-election context	Eight House definitions: national environment, prior-map lean, previous challenger/open-seat/incumbent role and uncontested-vote assumptions.	Context adjustment modestly improved margins; insufficient election-level confirmation. Results
Fundraising	29 Q3 receipt/donor/cash/spending comparisons; 8 expanded June-series comparisons, plus candidate/national combinations.	June total receipts included provisionally by choice. Other money measures remain research. Results
Lean, candidate factors and national conditions together	100 fixed combinations, including small-state incumbency, congressional approval interactions, poll depth/age and nonlinear lean.	No combination met all evidence and forecast safeguards. Results
Expert ratings	Cook, Sabato, Inside Elections and consensus, with historical final ratings.	Retrospective benchmarks may contain information published after the forecast cutoff; not sufficient for deployment. Results
Geography and demographics	Map-matched education and urbanicity profiles; eight shared-error structures, two simulation seeds and 100,000 draws per case/election.	Refreshed on the corrected baseline. No reliable improvement; regional default retained as an explicit scenario choice. Results
Additional state urbanization	Dated Census 2010 metro share and urban share, separately and together, after lean/incumbency controls.	Senate margin accuracy worsened; no addition. Results

Regional and demographic simulations

I refreshed the existing map-matched education and urbanization study rather than treating those inputs as unavailable. Profiles cover all states and districts for the modeled cycles. Models are fitted using earlier elections and keep each race’s total error variance consistent, so they cannot appear better simply by shrinking the uncertainty.

Eight error structures were compared in 2022 and 2024, using two seeds with 50,000 simulations each. Small score changes differed by chamber and election. Two elections cannot establish that an extra demographic factor improves joint control probabilities. The regional forecast remains the default scenario, with its experimental status stated explicitly.

Structure	House seat score, 2022 / 2024	Senate seat score, 2022 / 2024
baseline	3.40 / 2.88	0.66 / 0.51
education factors	3.40 / 2.89	0.66 / 0.51
urbanicity factors	3.40 / 2.88	0.66 / 0.52
demographic partial	3.40 / 2.88	0.66 / 0.51
geography demographics	3.42 / 2.91	0.66 / 0.52

The seat score rewards a distribution that puts more probability near the actual seat total; lower is better. Respondent overlap across polling firms remains a separate data gap.

How I checked the results and what remains uncertain

I used earlier elections to fit each historical prediction, kept repeated ballots from counting as separate races, and compared alternatives on matched race cohorts. National effects were tested at the election level. Hundreds of races and thousands of simulations do not supply hundreds of independent national environments.

An independent calculation reconciled all 100 combined-variable comparisons: 34,577 historical predictions and 17,308 current scenarios, including the provisional receipt adjustment. Separate checks covered financial coefficients, candidate matching, official precinct totals, prior-election incumbency formulas and demographic simulation draws. Page verification checked the saved data and structure; it did not establish a live browser interaction test.

Historical final expert ratings can contain later campaign information and remain retrospective benchmarks. Some early financial publication dates are less precise than recent ones. June and Q3 financial studies use different matched cohorts and cannot be ranked by their overall error levels. Missing donor geography, respondent overlap and comparable state-legislative transfer histories remain gaps rather than measured zero effects.

These are development elections already inspected during model building. Five independent elections cannot produce a two-sided exact sign-flip p-value below 0.0625; two elections cannot do better than 0.5. Failure to meet a 0.05 threshold is not proof that a factor has no effect. The most valuable next evidence is a precommitted test on new elections and better matched donor/candidate histories, rather than further searching these same results.

Evidence and downloads