CalendarAll →

The Capivote Forecast

The Capivote Forecast is the proprietary methodology of Capivote, an independent election analysis project, for estimating election results across countries. It aggregates voting-intention polls before the vote, projects the result during the count and checks every estimate against the official result.

Models by type of election

Two-round presidential elections, proportional parliamentary elections and referendums follow different rules and require different calculations. The methodology comprises eight polling models and two vote-count models, applied according to each country's electoral system.

Poll selection criteria

The estimates draw on polls from independent pollsters that disclose their methodology, with priority given to those with the best record in previous elections. Polls commissioned by parties or campaigns are excluded.

Probabilities and margins of uncertainty

Every poll is subject to error, and the electoral record makes it possible to measure it. The methodology simulates the election thousands of times within that margin and records how often each outcome occurs. Estimates are therefore presented as ranges and probabilities.

Projection during the count

The first votes counted rarely represent the electorate as a whole. The methodology compares each locality with its behaviour in the previous election and projects the final result while the count is under way. In Brazil, with 59% counted, the projection was within 0.1 point of the result.

Validation on past elections

No rule is adopted untested. Every methodological change is assessed on elections already held and is incorporated only when it reduces the error relative to the rule in force.

Forecast record and official sources

Every forecast is recorded before the election and later compared with the figures released by the country's electoral authority. The sources used are cited on each page.

Capivote is independent and does not work for parties or campaigns.

Numbers without AI

Polling averages, forecasts and live counts come from code with fixed rules. The simulations use fixed seeds, so the same data always give the same numbers. No published number is estimated, adjusted or rounded by an artificial intelligence model.

The method is described on this page. A new rule is adopted only after it has been tested on past elections, and every change is recorded in a decision log with its date, test and result. Every forecast is recorded before the vote and later compared with the official result.

Where we use AI: Capivote uses Claude, by Anthropic, to help write the code and the texts and to translate them between Portuguese, Spanish and English. Headlines from other newspapers on the home page are machine-translated by Claude Haiku and carry the label "machine translation".

Sources and licenses

TSE (Brazil); Interior Ministry (Spain); Israel's Central Elections Committee; Wikipedia (CC BY-SA) for polling and rating tables; University of Florida Election Lab (CC BY-NC-ND 4.0) for early voting; MEDSL (CC0) for results by vote mode; Silver Bulletin (pollster ratings).

Enjoyed it? Follow along and share it with friends.