Methodology

The model estimates the probability of each of the 350 seats from quantitative and qualitative data: the actual vote at every polling station, CIS survey microdata, the changing electorate, the overseas vote, published polls and an analysis of each party's line-up in each constituency. The output is not a single number but 50,000 possible scenarios.

The building blocks

1. Starting point. Polling-station results from the Ministry of the Interior and constituency results from the BOE, 2008 to 2023; each station's 2023 vote is the basis of the local split.
2. Changing electorate. Deaths since 2023, first-time young voters and naturalised citizens, with priors from the academic literature.
3. Polls. An aggregator separating sampling error, each pollster's house effect and the error common to the whole industry, calibrated on how polls missed in 2015–2023.
4. CIS. Its individual responses, not its published estimate: vote intention, recalled 2023 vote and stated likelihood to vote, used to build the vote-transfer matrix and the provincial split. The sample is weighted to INE population figures (sex, age, education and municipality size), and respondents who hide their past vote or are still undecided are imputed with a model (profile, ideology and closest party), never pro rata. In the poll aggregator it enters as one more pollster through its direct vote intention weighted by likelihood to vote, adjusted by its average deviation from the results of the nine 2023–2026 elections with a CIS pre-election survey.
5. Profile of each section. The same model is evaluated on the roughly 36,000 census sections (age, sex, education, employment, municipality size, foreign-born, estimated religiosity and INE 2024 income) and adjusts the provincial split according to who lives in each one.
6. Voters abroad (CERA). A separate model: they weigh heavily in Galicia and have grown with the 2022 nationality law.
7. Ballot and context. Coalitions and candidate lists in each constituency, agreements between parties and a correction for how polls missed at recent regional elections.
8. Seats. In each of the 50,000 scenarios, D'Hondt is applied with the 3% threshold in every constituency and the 2026 seat counts (Madrid 38, Cádiz 8).

Seat ratings

Once the model is live, each seat is rated by the probability that the favourite wins it: safe (95% or more), likely (80–95%), lean (60–80%) and toss-up (under 60%). Until then, the front page uses a historical measure: each seat's cushion with the 2023 vote.

Validation

Before publication, the model is tested by forecasting 2016, 2019 (April and November) and 2023 using only what was known before each election, and must beat simple baselines (uniform and proportional swing). All input data are reconciled with the official BOE totals.

Limitations

• If all polls miss in the same direction, the model will miss too; that is why its intervals have wide tails.
• Final candidate lists are known around 3 November: until then, some parties' line-ups are assumptions.
• No new estimate is published from 24 to 28 November (art. 69.7 of the Spanish electoral law).

The model and its specification are altnalysis's own work and are not published. Sources, what the model combines and downloads are under Data.