Modelling the 2026 US Midterm Elections

Model prediction

The model gives Democrats 246 House seats and Republicans 189. Across the 35 Senate contests, it gives Democrats 18 and Republicans 17.

I built this after Election Day using inputs dated November 2. These are retrospective estimates, not a forecast published before voting.

Model description

I compared 13 model families. The best performer was a simple blend: 75% race polling and 25% partisan baseline. Without a matched poll, it uses the baseline alone.

The model predicts the Democratic share of the two-party vote, excluding third-party votes:

The baseline adds local presidential lean, ℓ, to the national Democratic generic-ballot share, q. Lean measures how far a district or state sits from the country. House figures use the relevant district boundaries:

A matched race poll, p, receives three times the weight of the baseline:

I tested polling weights of 25%, 50%, 75% and 100%. Earlier-year tests selected 75% in every evaluated election and for the final model.

Polls cover the preceding 90 days, with a 14-day half-life and square-root sample-size weighting. A pollster's total weight is capped at its strongest survey. Candidate questions must match the saved roster.

The training data contain 3,063 D/R races: 2,656 House races from 2012–2024 and 407 Senate races from 2000–2024. Regression and tree models also used incumbency, election cycle and presidential party.

Comparing the models

The benchmark covers 13 families and 47 settings. I fixed the comparison rules before running it and ranked models by mean absolute error (MAE), in percentage points.

For each test election, 2018, 2020, 2022 and 2024, models fit and tune on earlier years only. Tests on 2014 and 2016 provide the initial tuning history. Each election and chamber receives equal weight.

The blend has an average error of 2.57 points, compared with 3.11 for the baseline, a reduction of 17.3%. Most of that improvement comes from Senate races: 2.50 points versus 3.49. The House difference is much smaller, 2.63 versus 2.72.

The blend loses to the baseline in both chambers in 2020. They tie in the 2024 House, where the saved inputs contain no race polls. The blend still ranks first when any one test election is removed, but four elections are a small sample.

I chose the winning family after seeing these scores. Choosing both the family and its settings from earlier years only gives 2.93 points of error, rather than 2.57. These elections also informed earlier versions of the project, so neither score is an untouched final test.

Among close races, where the baseline predicts 45–55% Democratic support, the blend orders the candidates correctly in 83.7% of 343 House races and 76.3% of 38 Senate races. The baseline scores 81.3% and 63.2%. Every model uses the same subset; the Senate sample is small.

With polling hidden, the blend becomes the baseline. It also uses the baseline for 327 House races and 8 Senate races without matched polls in 2026.

Race estimates and uncertainty

The ranges use earlier prediction errors, grouped by chamber and polling availability, with equal weight per election. Groups with fewer than two earlier cycles use the full chamber error history. Errors from 2014 and 2016 allow range checks to begin in 2018.

The 95% ranges contain more than 98% of House results in each test year, so they are wide rather than precisely calibrated. Brier scores assess the probability of a Democratic majority of the two-party vote. Neither check gives the chance of controlling Congress.

Ballot assumptions

In 14 House contests, all listed candidates belong to one party. Those seats are assigned from the roster, without estimating vote shares or choosing between same-party candidates.

For independent or minor-party opponents, I borrow the missing major party's baseline and use matched polls where available. This assumption is not validated by D/R history, so candidate-win probabilities are left blank.

Ranked-choice estimates describe first preferences, without simulating transfers. Louisiana House estimates combine candidates within each party and cannot identify a runoff winner.

Limitations

No 2026 outcomes enter fitting, model selection or calibration. The live comparison scores the fixed estimates separately. The November 2 cutoff relies on dates in files retrieved later; I cannot verify that every input was available in the same form before voting.

Historical rosters come from election returns, and older polling archives selected the eventual top candidates. Some presidential percentages are rounded. The data do not recreate exactly what a forecaster knew at the time.

The seat bars count which party leads each estimate. They do not model shared national errors or the probability of controlling Congress. The benchmark supports this blend on the saved data, not a claim that it is best for every election.

Error patterns