# Oregon Governor Forecast Charter

Published by **Consequential Races**  
*Measured forecasts for elections that could change how a state governs.*

Status: **Research preview**  
Election: **Oregon governor, November 3, 2026**  
Initial model series: **0.x**

## 1. The forecast's job

The tracker answers two different questions:

1. **Nowcast:** If voting ended today, what is the probability that each candidate would win?
2. **Election-day forecast:** What is the probability that each candidate will win the certified statewide popular vote on November 3, 2026?

The forecast reports probabilities, median vote shares, a median major-candidate margin, and 50% and 80% uncertainty intervals. A probability is not a race rating or a declaration of certainty. A candidate given a 30% chance should win roughly three times in ten across comparable forecasts.

## 2. Forecast target and conventions

- The target is the winner of the certified Oregon gubernatorial popular vote.
- The model preserves all ballot-qualified candidates. It also publishes a normalized Kotek–Drazan two-candidate margin for comparison across time.
- Positive margins mean Kotek leads; negative margins mean Drazan leads.
- Displayed probabilities are rounded to whole percentage points. Internally, full precision is retained.
- The public page shows the data cutoff, run time, model version, data version, and a data-quality assessment.
- The forecast is frozen at 8:00 p.m. Pacific on election night. Live returns belong to a separate results product and never rewrite the final forecast.

## 3. Model views

The tracker exposes three estimates rather than hiding their disagreement:

- **Polls:** a quality- and recency-adjusted estimate of current voter preference.
- **Fundamentals:** a structural estimate based on the 2022 rematch bridge, subsequent election results, incumbency, political environment, registration, campaign strength, and measurable state conditions.
- **Combined:** the official forecast, which updates the structural prior with polling evidence and simulates remaining uncertainty.

An expert-rating or prediction-market comparison may be displayed later, but it does not enter the initial official forecast.

## 4. The 2022 rematch bridge

The 2022 result is a candidate-specific anchor, not a deterministic vote-transfer rule.

The bridge retains uncertainty across five stages:

1. Certified 2022 all-candidate result.
2. Estimated 2022 Kotek–Drazan counterfactual without Betsy Johnson.
3. A 2026 structural prior combining electorate change, incumbency, national environment, campaign strength, and state conditions; the artifact retains those subcomponents.
4. Current quality- and recency-adjusted polling evidence.
5. The November forecast after combining the structural and polling estimates with remaining movement and election-day uncertainty.

For a comparable horizontal axis, the bridge positions use the normalized Kotek–Drazan two-candidate margin. The first station also reports the literal full-ballot margin (Kotek +3.42 in 2022), so normalization is never mistaken for the certified result.

Former Johnson voters may choose either major candidate, another candidate, or not vote. Major-candidate voters may also switch or abstain. Precinct-level patterns describe places, not individual voters, so ecological inference uncertainty must remain in every simulation.

The exact Johnson-voter split is not identified by certified results or any available direct second-choice question. The production model therefore represents Kotek's share **conditional on a Johnson-origin voter choosing one of the two major candidates** as a truncated-normal prior with mean `0.48`, standard deviation `0.16`, and bounds `0.05`–`0.85`. A separate Johnson-to-major rate leaves room for residual voting; its `0.90` center is an unvalidated judgmental participation assumption and is not inferred from VoteCast or Clout. The production bridge holds the 2022 Kotek and Drazan source blocs at 100% and does not separately estimate their retention or crossover; public, non-probabilistic stress tests show sensitivity to those omitted behaviors.

Two imperfect 2022 measurements inform the center and width of that prior:

- AP VoteCast permits an individual-level relative-favorability comparison among respondents who reported or intended a Johnson vote. That comparison is closer to the individual-voter question, but relative favorability is not a second-choice ballot test.
- Clout measured an initial three-candidate ballot and a forced Kotek–Drazan ballot in the same survey sample. The aggregate change constrains plausible transfers, but without a respondent-level cross-tab it cannot isolate Johnson voters from initially undecided, other-candidate, or switching major-candidate respondents.

Neither proxy is presented as an observed Johnson transfer rate. The wide distribution is a judgmental synthesis of conflicting, indirect evidence, not a fitted posterior. The repository publishes only derived AP VoteCast statistics and audit metadata; it does not redistribute the raw public-use microdata.

## 5. Poll eligibility and review

Every poll is reviewed before it affects the model. The review record preserves:

- pollster and sponsor;
- partisan or campaign affiliation;
- field dates and release date;
- sample size and population (adult, registered voter, or likely voter);
- mode, sample frame, weighting, and likely-voter method when disclosed;
- exact ballot question, candidate order, preceding questions, and all response options;
- original source URL and an immutable source snapshot identifier;
- whether another release reuses or overlaps the sample;
- decision and reason: `accepted`, `accepted_with_penalty`, `excluded`, or `superseded`.

Partisan polls are not automatically excluded. They receive an explicit sponsorship adjustment and wider uncertainty. Poll aggregators are discovery tools; the pollster's original release is the source of record whenever it is available.

Poll weights account for recency, effective sample size, population, mode, pollster quality, house effects, sponsorship, and correlated observations. The model never treats the published sampling margin of error as total forecast uncertainty.

## 6. Fundamentals

Candidate or party effects enter only through measurable, historically testable inputs. Candidate-specific fundamentals may include:

- recent state partisan lean and elasticity;
- incumbency and approval;
- prior performance in a competitive election;
- candidate elected experience;
- fundraising and cash on hand, with reporting-lag and endogeneity cautions;
- national political environment;
- state labor-market, income, and fiscal indicators;
- changes in registration and turnout composition.

The live model must not add undocumented analyst points for campaign narratives.

## 7. News and evidence policy

News is collected into a sourced evidence ledger. The default numerical effect of a story is zero.

An event may change the model when it changes a defined input, such as a certified ballot change, candidate withdrawal, qualifying poll, finance filing, approval measure, or economic release. A rare direct model intervention requires:

- a written rationale;
- a probability distribution rather than a certain point adjustment;
- a model-version change;
- a public changelog entry;
- a sensitivity result showing the forecast without the intervention.

AI-assisted discovery, classification, summarization, and deduplication are permitted. Automated news sentiment is not a forecast input.

Approval, registration, turnout, economic, finance, and advertising observations follow the same restraint. They may be published in the context ledger after source verification, but they have exactly zero mathematical effect until a feature definition and coefficient pass a documented out-of-sample gubernatorial backtest. Observations drawn from the same survey sample carry a shared sample identifier and may not be counted as independent evidence.

## 8. Update policy

- One scheduled public run each morning using a fixed data cutoff.
- An additional run after an approved qualifying poll or material certified candidate-field change.
- News may refresh more frequently without forcing a forecast change.
- Each published run is immutable. Corrections append a new run and correction note.
- Inputs that are revised later, including economic data, retain their original release vintage for historical backtests.

## 9. Validation and launch standard

The official probability does not leave research-preview status until it passes rolling-origin backtests on gubernatorial elections outside Oregon as well as Oregon-specific reasonableness checks.

Backtests use only data that would have been available at each historical cutoff and compare:

- repeat-prior-result baseline;
- simple poll average;
- polls-only model;
- fundamentals-only model;
- combined model.

Primary metrics are margin mean absolute error, Brier score, log loss, and 50%/80%/95% interval coverage. Validation snapshots should include approximately 120, 90, 60, 30, 14, 7, and 1 day before Election Day. The model specification is frozen before inspecting its 2026 output; later changes are versioned and documented.

## 10. Reproducibility

Every run records:

- data cutoff and generation time;
- model and methodology versions;
- source-manifest and input hashes;
- code revision when available;
- configuration and random seed;
- simulation count;
- output artifact hash;
- validation status and warnings.

The normalized inputs, generated forecast JSON, methodology, and archived run summaries are downloadable. No calculation required to reproduce a published forecast occurs only in the browser.