# Model research

## What the forecast uses

The forecast combines recent polling with adjustments learned from earlier polling misses. When polling is limited, state or district lean, national conditions, incumbency and prior candidate results receive more influence. The default simulations share national, state and regional changes across the two chambers while allowing individual races to differ.

Historical candidate performance and Senate fundraising are provisional additions. Selected Senate scandal penalties are judgments about possible electoral effects; they are adjustable and are not findings of wrongdoing. The current uncertainty ranges condition on these choices.

## Candidate performance

Earlier results are compared with the political environment, with attention to office, incumbency, recency and comparable boundaries. A paired refit replaces overlapping history terms rather than adding another incumbent bonus. Changes are capped at three margin points; unsupported histories retain their existing treatment.

Mary Peltola's House overperformance is present in the data. Its learned Senate transfer is small, so the entire replacement changes Alaska's projected margin by only about 0.12 points. Josh Turek has no numerical state-legislative overperformance record in the current series. That missing information is not evidence of no candidate appeal. The October 21 review documents both the blend and the candidate-transfer tests.

## Fundraising

The forecast uses comparable September 30 two-year receipts for eligible Democratic/Republican Senate pairs, with earlier-election-only fits. Twenty-four current pairs are covered. Receipts include more than individual donations and do not include outside spending.

In the matched 2014–2024 review, September receipts lower average margin error from 4.891 to 4.857 points. Probability error is almost unchanged from baseline and worse than the June comparison. The margin improvement is small and not statistically reliable. A fresher financial figure is not, by itself, evidence of a better forecasting formula. The Sullivan reporting-period difference is reconciled exactly; historical amendment timing remains a limitation.

## Polling, national conditions and expert ratings

Poll age, sample size and limits on a firm's influence control the polling average. Historical signed misses inform corrections; uncertainty also allows for errors shared across pollsters. Generic ballot, presidential approval, partisan lean, pollster weighting and expert ratings have been tested in several combinations. A factor's correlation with results does not establish that it improves a poll-based forecast.

Final historical expert ratings were accepted for exploratory comparisons. They can contain information unavailable at an earlier forecast date, so they must not be treated as a date-matched prospective test. Missing poll methods also limit attempts to separate methodological changes from firm or election effects.

## Simulations and election rules

Each default simulation links the House and Senate through shared national, state and regional changes. Shared geographic variation replaces part of local variation rather than simply increasing every interval. The regional scenario is a modeling choice with limited historical validation. Demographic error patterns require consistent district boundaries and historical demographic data before a reliable additional effect can be claimed.

Georgia allows for a runoff and further movement. Ranked-choice final-round polling is preferred where available, but complete voter-transfer models are not fitted. Independent winners remain separate from party control unless an explicit caucus assumption is supplied. Fixed-party House outcomes do not invent a separately modeled winning candidate.

## Reading the historical checks

Margin error measures the typical distance from the result. Probability scores assess confidence in winners; interval scores assess both width and missed outcomes. Comparisons use earlier information to fit each election. Only a few national cycles support the tests, and those elections have also informed model development.

The October21 review's weaker lean/fundamentals variants raise Democratic chances in Alaska and Iowa but increase historical error. That does not prove the current 2026 estimates are correct. It means those particular changes have not earned adoption from the available tests.

## Evidence

- [Current review and scenario tables](<../../audit/review_2026-10-21/REVIEW.md>)
- [Ranked research priorities and primary academic sources](<../../audit/review_2026-10-21/research/Research priorities.md>)
- [Organized source index](<../../audit/review_2026-10-21/organization/Sources and Research.md>)
- [Exact-duplicate inventory](<../../audit/review_2026-10-21/organization/duplicate_inventory.json>)