Click a bloc to show only its parties. Hover to highlight it on the map.
Each dot is a seat. Central allocation: mean of forecast vote shares, converted by Bader-Ofer with a 3.25% threshold. Full ranges below.
For each bloc - the seat distribution across election-day simulations. The dashed line is 61, the majority mark: one line on the distribution, not its target. Hover over the bars.
Bloc composition is a political definition, not a statistical one - based on the parties’ current declarations.
A vote for a list that misses the threshold is discarded, and the 120 seats are redistributed among all who cleared it - from all blocs, in proportion to their size. So a list falling at the threshold does not "return" its seats to its bloc - it scatters them. Here: the lists the model is unsure about, and what happens to each bloc in each scenario. Drag the slider - results are computed live from simulations.
Seats in the central scenario (parties always sum to 120), the range holding 90% of simulations, vote share and the chance of clearing the electoral threshold. Click a party for the full distribution and its path over time.
The line is the model’s estimate of true support each week, after netting out each pollster’s bias; the band is the 90% range. The dots are the polls. Move your mouse over the chart for values on any date, or pick a source to compare one pollster with the model.
How far each pollster measures each party above or below the industry average, in percentage points, after accounting for poll date. Pollster-to-pollster comparison only - an error they all share will not show here.
Blue: the pollster measures the party above average; orange: below it. Color intensity - size of the gap.
Every election poll published to the public must, under Section 16E of the Election Propaganda Law, be filed with the Central Elections Committee and is published on its website. Project 120 reads all polls from there - as filed, including sample size, fieldwork dates and method - not from news sites. Sub-population polls, polls without a vote question and duplicate filings are excluded.
Hierarchical Bayesian model: each party’s support is a time series that drifts at a rate learned from the data, separately per party, with the rates shrunk toward a shared industry-wide level (a list with a short history borrows strength from the others instead of being clipped by hand). A list that entered mid-series is measured only from the polls after its entry: polls that did not ask about it, and the first week after entry, are not treated as data about it. Each pollster gets a learned house effect - its systematic lean relative to the rest - which is netted out. Every poll is weighted by sample size, and extra noise is learned per pollster. The output is not a single number but thousands of possible election-day scenarios.
An error shared by every pollster cannot be detected from the polls - they agree with each other, and agreement looks like accuracy. So the model is given a stated "industry bias", from two sources detailed in the tables below: Multi-election calibration - the systematic gap between the final polls and the result in the five elections of 2019-2022, by political family, weighted toward the most recent elections; and The midpoint - for parties on which pollsters are currently split into two schools. Both sources enter as a distribution, not a number: each party gets a standard deviation measured from the spread across elections, and the midpoint gets uncertainty of its own size.
Every vote-share scenario is converted to 120 seats by the Bader-Ofer method: 3.25% electoral threshold, quota, surplus allocation by largest average, and the surplus agreements actually signed: Likud-HaTzionut HaDatit, Shas-Yahadut HaTorah, Yashar-HaDemokratim, Yisrael Beitenu-Beyachad, Joint List-Ra’am. The method was validated against the 25th Knesset election results and reproduces them exactly. The model knows nothing about turnout, events that have not yet happened, or who will form a government.