> 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/2026-forecast-methodology.md).
# 2026 Forecast Methodology
A Quantitative Framework for Forecasting 2026 U.S. Congressional Elections
Methodology Overview
2026 Election Cycle | U.S. House and Senate
## 1. Introduction
We describe the statistical methodology underlying a probabilistic forecasting system designed to estimate the likelihood of Republican and Democratic victory for each US House and Senate race in 2026. The model integrates three distinct classes of information: structural fundamentals derived from historical electoral and financial data, aggregate public polling, and aggregated prediction market information. The outputs are race-level win probabilities, qualitative competitive ratings, and seat-distribution simulations for each chamber.
The forecasting framework is motivated by a long-standing literature in political science demonstrating that election outcomes are meaningfully predictable from both structural conditions and contemporaneous public opinion measures. The present approach draws on this tradition while incorporating advances in ensemble machine learning and market-based inference to produce estimates that are robust to the limitations of any single data source.
## 2. Data Sources and Variable Construction
The model draws on four primary data sources: a historical training dataset spanning multiple election cycles, Federal Election Commission (FEC) campaign finance records, the DDHQ live polling database, and prediction market price feeds.
### 2.1 Historical Training Data
The core training dataset is assembled at the race level for contested general elections from 2016 through 2024. Each observation encodes the two-party vote margin, alongside a set of structural predictors.
### 2.2 Partisan Baseline
Public opinion today is sharply polarized along partisan lines, and political scientists widely recognize baseline partisanship as among the most reliable predictors of future electoral outcomes \[1]\[2]. Our model leverages this insight: the prior electoral lean of the district or state. For House races, this is operationalized as the most recent presidential margin adjusted for national environmental swing. For Senate races, the analogous measure is the adjusted prior presidential margin within the state. These variables capture the underlying partisan composition of the electorate and serve as a strong prior on competitiveness, particularly in races where polling is sparse or absent.
### 2.3 National Political Environment
The political science literature widely recognizes the generic ballot as a strong predictor for House election outcomes \[3]\[4]. A single continuous variable summarizes the broader electoral environment in a given cycle, encoded by the DDHQ generic 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 allows the model to adjust all race-level predictions in a consistent direction when national forces favor one party, reflecting the well-documented tendency for election outcomes across individual districts to co-move with national partisan tides.
### 2.4 Campaign Finance
Fundraising data are drawn from the Federal Election Commission and supplemented with outside expenditure records. The model constructs a normalized receipts measure for each candidate that incorporates direct campaign receipts, coordinated outside support, and offsetting outside opposition spending:
Fnorm = RReceipt + SupportOutside - Oppositionoutside
The resulting variable captures the net financial advantage accruing to each party's candidate in a given race. In the absence of available finance data for a given race, the variable is imputed to the sample mean, preventing data gaps from distorting predictions for unpolled or under-reported contests.
### 2.5 Incumbency and Structural Covariates
For Senate races, incumbency status is included as a categorical predictor. Incumbency is well-established in the political science literature as having historically conferred a meaningful electoral advantage, particularly in lower-information environments \[5]\[6]. An additional interaction variable encodes the combination of the incumbent president's party and cycle type (midterm versus presidential year), capturing the consistent pattern by which the party holding the White House tends to underperform in midterm elections relative to presidential election years.
## 3. Model Specification and Estimation
### 3.1 Training Architecture
The model is trained separately for House and Senate races, and separately for each target cycle using a strict temporal holdout: all observations from prior cycles are used for training, while the target cycle constitutes the test set. This design prevents data leakage and produces out-of-sample forecasts that are directly analogous to real-world conditions in which the future is unobserved at the time of prediction.
Continuous predictors are standardized prior to model fitting using z-score normalization estimated on the training set.
### 3.2 Ensemble Approach
Point predictions of the two-party margin are generated by an ensemble of complementary algorithms. The three algorithms differ substantially in their inductive biases: Ridge regression imposes linear structure with L2 regularization to prevent overfitting to small historical samples; Random Forest aggregates many shallow decision trees to capture nonlinear interactions without explicit feature engineering; XGBoost employs sequential boosting of residual errors to achieve low bias under controlled regularization.
The fundamentals-driven ensemble prediction MF for race i is computed as the average margin m predicted by each model n :
MF(i) = 1N đșnNmn
Averaging across models with diverse error structures reduces variance relative to any single estimator, a property that is especially valuable given the limited size of the historical training sample. The ensemble's aggregate prediction serves as the fundamentals-based margin estimate for each race.
## 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 DDHQâs live database and aggregated to a race-level average as of the most recent available date. The polling weight assigned to a given race is a linear function of the number of available polls n, scaled to a maximum weight at a ceiling of nâ polls:
wpoll = wmax min( 1, nn0)
Races with no polling receive a weight of zero, defaulting to the pure fundamentals prediction. The combined final model margin m for race i is then computed as:
m(i) = \[1-wpoll(i)] MF(i) + wpoll(i) ppoll(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 PR(i) for race i based on polling and fundamentals is:
PR(i) = (m(i))
where Ί(·) denotes the standard normal CDF and Ï is the residual standard deviation of the model, estimated empirically from historical out-of-sample forecast errors. This formulation is consistent with standard probit-type frameworks in the political science literature and produces well-calibrated probability estimates across a wide range of competitive conditions.
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. Prediction markets have demonstrated competitive accuracy relative to polling-based models in a number of forecasting studies \[7]\[8], and their inclusion 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 and matched to corresponding race. The effective trading volume V associated with each market, aggregated over multiple prediction exchanges, determines the weight assigned to the market signal. A race must clear a minimum effective volume threshold Vmin, with weight scaling up to a maximum of wmax as effective volume V approaches a ceiling Vâ :
wmkt (V) = wmax VV0 if V Vmin , else 0
The final win probability for race i is computed as a weighted average of the model-plus-polling baseline probability PR and the market-implied probability Pmkt:
PF(i) = (1 - wmkt) PR(i) + wmkt Pmkt
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 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 the chamber following the election, the model generates Monte Carlo simulations for each individual race outcome. 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:
yiN =(PF(i)) + + i , i \~ N(0, 2)
where λ governs the strength of cross-race correlation and η is a shared draw from the standard normal distribution. The Republican candidate wins race i in simulation N if yiN > 0. Uncontested racesâdefined as those in which one party has no viable opponentâare excluded from the probabilistic model and assigned deterministically to the appropriate party. The distribution of simulated seat outcomes yields chamber-level majority probabilities, expected seat margins, and full distributional uncertainty intervals.
## 8. Limitations and Caveats
Several limitations of the present framework merit explicit acknowledgment. First, the historical training sample available to any election forecast is relatively small by machine learning standards: congressional elections only occur every 2 years, and each cycle provides fewer than 500 additional observations, introducing a degree of irreducible estimation uncertainty. The regularized ensemble approach partially mitigates this concern by reducing overfitting, but it does not eliminate it.
Second, the model assumes that the structural relationships observed in past cycles are stable over time. To the extent that the 2026 electoral environment reflects genuinely novel conditionsâwhether in candidate quality, issue salience, or voter behaviorâthe model may systematically under- or overestimate the magnitude of partisan shifts. Third, polling averages and prediction market prices are themselves imperfect signals subject to systematic biases, late-breaking information, and coverage gaps, particularly in lower-profile House races where public polling is sparse or nonexistent. Prediction markets in particular are very recent innovations, and are rapidly evolving. Their forecast integration will surely continue to develop and evolve 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.
## References
\[1] Bartels, L. M. (2000). Partisanship and voting behavior, 1952â1996. American Journal of Political Science, 44(1), 35â50.
\[2] Abramowitz, A. I., & Webster, S. W. (2016). The rise of negative partisanship and the nationalization of U.S. elections in the 21st century. Electoral Studies, 41, 12â22.
\[3] Bafumi, J., Erikson, R. S., & Wlezien, C. (2010). Balancing, generic polls and midterm congressional elections. Journal of Politics, 72(3), 705â719.
\[4] Bafumi, J., Erikson, R. S., & Wlezien, C. (2010). Forecasting house seats from generic congressional polls: The 2010 midterm election. PS: Political Science & Politics, 43(4), 633â636.
\[5] Gelman, A., & King, G. (1990). Estimating incumbency advantage without bias. American Journal of Political Science, 34(4), 1142â1164.
\[6] Ansolabehere, S., & Snyder, J. M. (2002). The incumbency advantage in U.S. elections: An analysis of state and federal offices, 1942â2000. Election Law Journal, 1(3), 315â338.
\[7] Atanasov, P. D., Witkowski, J., Mellers, B., & Tetlock, P. E. (2025). Crowd prediction systems: Markets, polls, and elite forecasters. International Journal of Forecasting, 41(2), 580â595.
\[8] Berg, J. E., Nelson, F. D., & Rietz, T. A. (2008). Prediction market accuracy in the long run. International Journal of Forecasting, 24(2), 285â300.