Machine Learning Winds Forecasts

How the forecasts work

Each forecast is a simulation: we draw thousands of possible elections from a model of how far polls and fundamentals usually miss, then count how often each candidate or party wins.

US Senate

Each race starts from fundamentals: the state's lean in the last presidential election, whether an incumbent is running, and the national environment measured by generic-ballot polls. That gives a starting estimate and a range around it.

The polls then update every race at once. The model treats each poll as a noisy reading of where the race will finish, and it accounts for four things a simple average misses:

  • Pollster leans. Some firms consistently lean toward one party. The model estimates each pollster's lean from all the races it polls in a year and corrects for it. Campaign and party polls are allowed a larger lean.
  • Time. Opinion moves before Election Day, so a poll from six weeks ago says less than one from last week. How fast opinion drifts is learned from past elections.
  • Shared errors. Polls can miss in the same direction everywhere, as they did in 2020. The model includes that possibility, so it never treats fifty polls as fifty independent pieces of evidence.
  • Similar states. States share a national swing and a regional one, so a poll in one state also nudges its neighbors.

Because every piece is a normal distribution, the result is calculated exactly rather than approximated. We then draw 10,000 simulated elections from it to get each candidate's chances and the range of Senate seats. A poll counts as public two days after its last day of fieldwork.

How accurate it has been

We test the model on every Senate election from 2006 to 2022, forecasting each year using only earlier years, including the settings for drift and pollster leans. The Brier score measures probability accuracy: 0 is perfect and 0.25 is a coin flip. The table also shows our earlier, simpler model, which blended a weighted poll average with fundamentals.

Days before electionBrier scoreEarlier modelRaces calledVote-share error (points)Seat count inside 80% range
10.0410.04293%2.67/9
70.0450.04893%2.77/9
140.0500.05292%2.97/9
210.0460.05293%3.18/9
280.0490.05593%3.38/9
350.0540.05792%3.57/9
420.0570.06191%3.88/9
490.0720.07190%4.36/9
560.0960.09686%5.25/9

Spain

The Spanish model works in three steps.

National vote shares. Every published poll is treated as a noisy reading of each party's vote on election day. As in the US model, the polls are combined with pollster leans, drift over time and an error shared by all pollsters. Shares of different parties move together the way a fixed electorate splits between them: when one party gains, the others lose in proportion to their size. A joint figure, such as Broad Front and Compromís–FV reported together, is used as a reading of their sum. The state pollster CIS is allowed a much larger lean than private firms. The starting point is the previous election's result, with a wide range around it.

Constituencies. Each simulated national result is spread across the 52 constituencies by proportional swing: a party that is up 10% nationally is assumed up about 10% everywhere it stands, not up the same number of points. Each simulation adds random shifts for every autonomous community and every province. Their sizes come from how far real results strayed from proportional swing between 2015 and 2023. The provinces are then adjusted so they add back up to the national result.

Seats. Seats are allocated with the D'Hondt method in each constituency, among lists with at least 3% of the valid vote there (blank ballots count as valid). Ceuta and Melilla elect one deputy each. We draw 10,000 simulations.

Each day's forecast uses only polls published by that day. A poll counts as public two days after its fieldwork ends. Spanish law bans publishing polls in the five days before the vote, so the forecast freezes on November 23.

How accurate it has been

We tested the model on the elections of 2016, April 2019, November 2019 and 2023. Each test used only the polls available at the time, plus constituency shifts estimated from the other elections. Two simpler methods get the same polls for comparison: a 14-day poll average spread to constituencies by uniform swing (the same number of points everywhere) and by proportional swing.

Days before electionSeats misallocated: our modelPoll average, proportional swingPoll average, uniform swingVote-share error (points)Seat counts inside 80% range
11818200.740 of 41
72120200.838 of 41
142122230.937 of 41
282421241.037 of 41
563336361.734 of 41

Seats misallocated is half the sum, over all parties, of the gap between forecast and actual seats, averaged over the four elections. Our model's figure uses the most likely seat count given its central vote estimate.

With only four past elections, these numbers are rough. The model's seat error is roughly in line with the simple methods at every horizon: a little better at some, a little worse at others. Most of the remaining error comes from the polls themselves: in 2023 they underestimated the PSOE by more than three points, and in 2016 they overestimated Unidos Podemos by about as much. The model's main advantage is honest uncertainty. Across all four elections, about 9 in 10 parties' seat totals landed inside its 80% range, so if anything the ranges are slightly too wide.

We set the model's drift, pollster-lean and shared-error settings in advance rather than fitting them to these four elections. Trying a range of alternatives changed the backtest very little.

Sources

  • Election results: MIT Election Data and Science Lab, U.S. Senate and President statewide returns (CC0).
  • 2026 polls: VoteHub Polling API (CC BY 4.0), with links to every original release.
  • Historical polls: FiveThirtyEight pollster-ratings archive (CC BY 4.0).
  • 2026 generic-ballot polls: Wikipedia, "2026 United States elections polling" (CC BY-SA 4.0).
  • Candidates: FEC candidate data and Ballotpedia.
  • Spain: constituency results for 2015 to 2023, voting-intention polls and the 2026 seat allocation, from English Wikipedia (CC BY-SA 4.0): the "Results of the ... Spanish Congress election" and "Opinion polling for the ... Spanish general election" pages. The underlying figures come from Spain's Ministry of the Interior and the pollsters.

Corrections

When we find a mistake in our data or code, we fix it, rerun the forecast, and note the change on this page.

Changes

  • November 14, 2026: the Spain forecast launched.
  • November 14, 2026: the US Senate model now estimates pollster leans, opinion drift and shared polling errors jointly. It replaced a model that blended a weighted poll average with fundamentals.