> For the complete documentation index, see [llms.txt](https://votes-docs.decisiondeskhq.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://votes-docs.decisiondeskhq.com/forecast-methodology/2026-governor-forecast-methodology.md).

# 2026 Governor Forecast Methodology

A Quantitative Framework for Forecasting 2026 U.S. Gubernatorial Elections

## 1. Introduction

We describe the statistical methodology underlying a probabilistic forecasting system designed to estimate the likelihood of a Republican or Democratic victory in each of the 36 gubernatorial races on the 2026 ballot. The model integrates three distinct classes of information: structural fundamentals derived from historical electoral data, aggregate public polling, and prediction market prices. The outputs are race-level win probabilities, qualitative competitive ratings, and a seat-distribution simulation across all gubernatorial contests.

The forecasting framework applies the same three-layer blending architecture used in the companion House and Senate models, adapted to the particular data constraints of gubernatorial elections: a smaller number of contests per cycle, the absence of a campaign-finance signal with coverage comparable to that available for federal races, and a state executive office whose electoral behavior does not always track federal partisan trends as closely as House and Senate races do.

## 2. Data Sources and Variable Construction

The model draws on three primary data sources: a historical training dataset spanning multiple gubernatorial election cycles, the DDHQ live polling database, and prediction market price feeds. Unlike the companion House and Senate models, this model does not currently incorporate a campaign-finance covariate, reflecting the comparatively limited and less standardized financial disclosure data available for state executive races (see Section 8).

### 2.1 Historical Training Data

The core training dataset is assembled at the race level for contested gubernatorial general elections from 2016 through 2024. Each observation encodes the two-party vote margin alongside a set of structural predictors. Gubernatorial elections occur on a staggered schedule across states, and roughly three dozen contests are held in any given cycle. This is a substantially smaller sample than is available for House races, and is comparable in scale to the sample available for the Senate. That smaller sample constrains both the complexity of the model that can be reliably estimated and the granularity of the historical back-testing described in Section 5.

### 2.2 Partisan Baseline

As with the companion House and Senate models, the model leverages the well-documented reliability of baseline partisanship as a predictor of electoral outcomes. For gubernatorial races, this is operationalized as the state's most recent presidential margin, adjusted for the national environment described in Section 2.3. Unlike the House and Senate models, the prior gubernatorial result itself is not used as an independent predictor. Outgoing governors' personal margins are frequently idiosyncratic, reflecting name recognition, tenure-specific circumstances, or a personal brand that is not transferable to a successor, and once incumbency is accounted for separately (Section 2.4), the prior result does not improve on presidential lean as a measure of a state's underlying partisan composition.

### 2.3 National Political Environment

In all of our models, the generic ballot is a single continuous variable that summarizes the broader electoral environment in a given cycle; it is encoded by the DDHQ generic congressional ballot polling average. The average is smoothed via a Kalman filter to prevent transient polling fluctuations from unduly influencing the forecast outlook. The resulting smoothed average is used to adjust each race's partisan baseline (Section 2.2) to reflect the current national mood, allowing the model to shift all race-level predictions in a consistent direction when national forces favor one party.

### 2.4 Incumbency

Rather than including incumbency as a fitted categorical predictor, the model estimates a direct, non-parametric incumbency adjustment: the median historical gap between a race's actual margin and its partisan baseline, computed separately for Republican-held, Democratic-held, and open seats, using only cycles prior to the target year. The median is used in place of the mean because a small number of outlier gubernatorial races — most notably several instances of unusually popular incumbent governors substantially outperforming their state's presidential lean — would otherwise exert disproportionate influence on the estimated effect, given how few gubernatorial contests are available for estimation within any single incumbency category.

### 2.5 Candidate Quality (WAR)

A supplementary candidate-quality adjustment, constructed using the same methodology applied to the companion Senate and House models, is added to the fundamentals estimate for each race. This adjustment captures candidate-specific factors, primarily previous electoral over- or underperformance.

## 3. Model Specification and Estimation

### 3.1 Training Architecture

The model is trained using a strict temporal holdout: for any given target cycle, the incumbency and candidate-quality adjustments described above are estimated only from cycles strictly prior to that year. This design prevents data leakage and produces out-of-sample estimates that mirror real-world conditions, in which the future is unobserved at the time of prediction.

### 3.2 Fundamentals Estimation

Unlike the ensemble machine-learning approach used in the companion House and Senate models, the gubernatorial fundamentals estimate is constructed as a direct sum of its component parts, rather than as the output of a fitted statistical or machine-learning model. The fundamentals-driven estimate $$M\_F$$ for race i is:

$$
M\_F(i) = \text{Baseline}(i) + \text{Incumbency}(i) + \text{Quality}(i)
$$

This specification reflects the more limited historical sample available for gubernatorial races. Flexible, tree-based ensemble methods, of the kind used in the House and Senate models, tend to compress predictions toward the center of the training distribution, a form of shrinkage that is especially costly for races with few or no close historical analogues. An explicit additive specification instead preserves the full range of the underlying baseline and incumbency signals, at the cost of forgoing any interaction effects a fitted model might otherwise capture. Given the comparatively small number of available observations, this tradeoff is judged to favor the additive approach for state executive races.

## 4. Polling Integration

Where public polling is available for a given race, the fundamentals-based margin estimate is blended with a polling average to produce a combined model margin. Polling data are sourced from DDHQ's live database and aggregated to a race-level average as of the most recent available date; for races without an official DDHQ polling average, that average is computed directly from individual published polls. The polling weight assigned to a given race is a linear function of the number of available polls n, scaled to a maximum weight $$w\_{\text{max}}$$ at a ceiling of $$n\_0$$ polls:

$$
w\_{\text{poll}} = w\_{\text{max}} \cdot \min\left(1, \frac{n}{n\_0}\right)
$$

Races with no polling receive a weight of zero, defaulting to the pure fundamentals prediction. The combined model margin m for race i is then computed as:

$$
m(i) = \left\[1 - w\_{\text{poll}}(i)\right] \cdot M\_F(i) + w\_{\text{poll}}(i) \cdot p\_{\text{poll}}(i)
$$

This framework ensures that heavily polled races are more responsive to public opinion data, while thinly polled contests remain anchored to the structural prior.

## 5. Probability Estimation

Win probabilities are derived from the combined model margin using the normal cumulative distribution function. Under this specification, the Republican win probability $$P\_R(i)$$ for race i based on polling and fundamentals is:

$$
P\_R(i) = \Phi\left(\frac{m(i)}{\sigma}\right)
$$

where Φ(·) denotes the standard normal CDF and σ is the residual standard deviation of the fundamentals estimate, derived empirically from out-of-sample forecast errors across the three most recently completed gubernatorial cycles (2020, 2022, and 2024) and scaled by a fixed multiplier to avoid overconfidence in the resulting probabilities. Cycles prior to 2018 are excluded from this back-test, as too few gubernatorial races preceded them to reliably estimate the incumbency adjustment described in Section 2.4.

Races are subsequently assigned to a seven-category competitive rating scale ranging from Safe Democratic through Safe Republican, with intermediate categories of Likely, Lean, and Toss-Up. Rating thresholds are applied to the win probability output and are designed to match the conventional qualitative classifications used in political analysis.

## 6. Prediction Market Integration

A third input layer incorporates prices from prediction markets, which aggregate the probabilistic assessments of financially motivated participants. The inclusion of prediction markets provides an independent signal that may incorporate information not yet reflected in public polling or structural variables.

Market-implied win probabilities are sourced from a market price feed, aggregated across the Polymarket and Kalshi exchanges, and matched to the corresponding race. Prices are clipped to a fixed range near the boundaries prior to use, to prevent a single extreme or thinly traded quote from dominating the blend. The effective trading volume V associated with each market, aggregated across exchanges, determines the weight assigned to the market signal. A race must clear a minimum effective volume threshold $$V\_{\text{min}}$$, with weight scaling up to a maximum $$w\_{\text{max}}$$ as effective volume approaches a ceiling $$V\_0$$:

$$
w\_{\text{mkt}}(V) = w\_{\text{max}} \cdot \sqrt{\frac{V}{V\_0}} \quad \text{if } V \geq V\_{\text{min}}, \text{ else } 0
$$

Rather than blending market and model probabilities directly, the market price is first converted to an implied margin using the same residual scale σ described in Section 5, and blended with the fundamentals-and-polling margin in that common unit:

$$
M\_{\text{mkt}}(i) = \sigma \cdot \Phi^{-1}\left(\tilde{p}\_{\text{mkt}}(i)\right)
$$

$$
m\_{\text{final}}(i) = \left\[1 - w\_{\text{mkt}}\right] \cdot m(i) + w\_{\text{mkt}} \cdot M\_{\text{mkt}}(i)
$$

The final win probability for race i is then computed once, from this fully blended margin:

$$
P\_F(i) = \Phi\left(\frac{m\_{\text{final}}(i)}{\sigma}\right)
$$

This three-layer blending architecture — fundamentals, polling, and markets — ensures that each input class contributes in proportion to the strength of the available signal, rather than treating all data sources as equally informative regardless of coverage or depth.

Market probabilities are sourced from the Polymarket and Kalshi exchanges. DDHQ is not paid or sponsored by Polymarket or Kalshi to use prediction market data in its forecast.

## 7. Seat Distribution Simulation

To estimate the probability that each party controls a majority of the 36 governorships on the 2026 ballot, the model generates one million Monte Carlo simulations of the full set of race outcomes. Each simulation draws a correlated vector of race outcomes from a single-factor probit model, in which a shared national factor η induces positive correlation across all races. For race i in simulation N, the latent outcome is:

$$
y\_i^N = c \cdot \left\[\Phi^{-1}\left(P\_F(i)\right) + \lambda \cdot \eta\right] + \epsilon\_i, \quad \epsilon\_i \sim N(0,1)
$$

where λ governs the strength of cross-race correlation, η is a shared draw from the standard normal distribution common to every race within a given simulation, and c is a scaling constant chosen so that each race's marginal win probability, averaged across simulations, continues to match $$P\_F(i)$$. The Republican candidate wins race i in simulation N if $$y\_i^N > 0$$.

The distribution of simulated outcomes yields the probability of a Republican or Democratic majority, the expected and median number of governorships held by each party, and full distributional uncertainty intervals. The simulation output additionally supports race-by-race correlation analysis and identification of the tipping-point race most likely to decide the majority in any given simulation, alongside sample full-map draws for scenario exploration.

## 8. Limitations and Caveats

First, the historical training sample available for gubernatorial elections is smaller than for House races and comparable to the Senate: roughly three dozen contests occur in any given cycle, and gubernatorial elections prior to 2018 are excluded from back-testing entirely due to an insufficient number of preceding observations. This introduces a meaningfully greater degree of irreducible estimation uncertainty than in the companion House model, and is the principal reason the fundamentals estimate is specified additively rather than through a fitted ensemble (Section 3.2).

Second, and unlike the companion House and Senate models, the gubernatorial fundamentals estimate does not currently incorporate a campaign-finance covariate. State-level campaign-finance disclosure is less standardized and less consistently available than federal FEC filings, and a comparable normalized receipts measure has not yet been constructed for gubernatorial candidates.

Third, the model assumes that the structural relationships observed in past cycles are stable over time. To the extent that the 2026 electoral environment reflects novel conditions, whether in candidate quality, issue salience, or voter behavior, the model may systematically under- or overestimate the magnitude of partisan shifts.

Fourth, polling averages and prediction market prices are themselves imperfect signals subject to systematic biases, late-breaking information, and coverage gaps, particularly in lower-profile gubernatorial races where public polling and market liquidity are sparse. Prediction markets in particular are a very recent innovation and are evolving rapidly; their integration into election forecasts will continue to develop over the coming years.

Finally, all probabilistic forecasts should be interpreted as estimates under uncertainty rather than definitive predictions. The competitive ratings and win probabilities reported by this system represent the model's best assessment of the likely range of outcomes given available information at a point in time; they will be updated as new data are incorporated across the election cycle.