The forecast

The final prediction, Oct 27, 2026

The Netanyahu bloc reaches 61 in 1% of runs, the anti-Netanyahu Zionist bloc in 0%, neither in 99%. The anti-Netanyahu bloc with the Arab lists and the lists between the blocs reaches 61 in 99%. The Netanyahu bloc is the larger of the two Zionist blocs in 66% of runs. Largest list: Yashar! in 57% of runs, Likud in 43% of runs.

Forecast seats: whole-number party predictions reconciled to 120; each bloc is the sum of its parties. The simulated bloc median is a separate summary and can differ from that sum. Vote share includes the model's adjustments; prior is the party-class adjustment in percentage points. Means, ranges and probabilities summarize the simulations.

What goes in

The field: the newest poll from each publisher within 14 days of the latest, 7 polls to Oct 23, 2026, less the two firms below. The weight is the square root of the sample, shared between outlets polled by one firm.

Left out of the field and used as checks: Channel 14 and i24NEWS, Shlomo Filber's firms. Their historical lean over the field is what step 4 below supplies from results; counting both would count it twice.

Surplus-vote agreements in the count (5, all published by the Central Elections Committee on October 19): Yashar! with The Democrats; B'Yachad with Yisrael Beiteinu; Joint List with Ra'am; Likud with Religious Zionism–Zehut; Shas with UTJ. Withdrawn after the ballot was set and dropped from the model: Blue and White.

How it is made

Version 2, October 24: eleven steps, each a rule from the record wherever the record allows one, each tested on the six elections since 2015 before it was kept. The two marked as judgments are the size of the width and of the consolidation draw.

1. Polls. The newest poll from each publisher within 14 days of the latest. The field is every publisher except Channel 14 and i24NEWS, which the model uses as checks rather than inputs. Two outlets polled by one firm share that firm's weight (Maariv and Walla by Lazar; Kan 11 and Israel Hayom by Kantar), weights by the square root of the sample. Today: 7 field polls.
2. Seats to shares. A poll's seats are the firm's own Bader-Ofer allocation, so a list with k seats holds a share between k and k+1 times the poll's divisor, and at least the threshold since it passed: the model takes the midpoint. The old seats-out-of-120 rule put every 4-seat list at 3.1 to 3.2%, under the line, and called lists failing that every poll had passing. A list the firm put below keeps the percentage it printed; with none printed, the firm's last printed figure, else the other polls' median.
3. Average. The weighted mean share per list, scaled to 98.2% of the vote: 1.8% is reserved for the lists the polls do not report (Noam, Haredi Public, micro-lists) and for Blue and White, which withdrew on October 19 and is dropped, its ballot slips still counted as wasted votes.
4. Priors by list. Vote-share points added where the final-week polls have missed one kind of list in one direction in both halves of the record since 2013, shrunk 20 to 40% toward zero: Shas +0.8, Ra'am +0.6, B'Yachad −0.8, Reservists–Economic −0.8, People of Israel −0.8. Shas is under-polled in six of seven elections; Ra'am beats its polls (its base votes at a steady rate), while the Joint List's correction is the turnout term of the next step; every new or first-time-solo list polled at 2.4 to 4.6% has failed, five of five; Bennett's lists have fallen short six times in six.
5. Arab turnout. The polls behave as if anchored on the previous election's Arab turnout: the change in Arab list votes per Arab elector, relative to general turnout, explains in hindsight, with the actual turnout, about nine tenths of the Arab lists' final-poll miss since 2013; forecast from the record alone it explains less, cutting the Arab lists' error by about half. That index reverts to its long-run mean; from 2022, which sat on the mean, the expected change is +0.8% of the Joint List's share, and each run draws a further swing of 19.6% around it, every other list diluted in proportion. This replaces the old fixed Joint List prior and the Arab turnout swing. Pre-blackout turnout expectations of about 60% were not added on top: the corrected polls already imply them.
6. The lean. A common swing of 1.7 seats toward the Netanyahu bloc: the field's miss on that bloc in the final week of the six elections held under the 3.25% threshold, 2015 to 2022, weighted toward the recent ones with a half-life of two elections (the table below). The spread left after the lean is 1.5 seats. Taking the lean net of Shas's own miss was tested and rejected: it under-predicts the bloc by three seats on the record.
7. Why the two firms are checks. In 2020, 2021 and 2022 Direct Polls sat about two seats above the field, and so did the result. That house lean is the signal the lean already encodes, estimated from results rather than from the firm. This year the two firms sit 8 to 12 seats above the field, four times their historical lean; nothing in the record explains that, so the model neither follows it nor rules it out: it widens.
8. Width. Each simulated election moves the whole Netanyahu bloc by a random swing drawn from a Student-t distribution with 5 degrees of freedom and a scale of 1.7 seats: the predictive distribution of the next miss given the six on the record (the residual, widened for a seventh draw; never under 1.5 seats). Version 2 used a normal swing of 3.1 seats, twice the residual; on the record that put all six past results between the 31st and 66th percentiles, a width the data did not support.
9. List noise. Every list is moved on its own, 1.08 points at an 8% share scaling with the square root of the share, never under 1 point for a small list (the misses of lists near the threshold run about a point). Inside each bloc the moves sum to zero: history shows lists trading votes within a bloc while the bloc total holds, so the noise is substitution, not a second bloc movement. Ra'am, outside the turnout term, keeps its share of a turnout swing of 1.7 seats, the size of the Arab lists' misses in 2020 to 2022.
10. Consolidation (judgment on the size). One draw per bloc, 1.5 seats standard deviation, moved between the bloc's largest list and its partners: in 2015, April 2019 and 2020 Likud beat its polls by 3 to 9 seats while its bloc moved under one.
11. Count. The 3.25% threshold on the whole vote, the reserve included, then 120 seats by Bader-Ofer with the 5 surplus-vote agreements the Central Elections Committee published on October 19, in 50,000 runs.
12. Central seats. Each party's median simulated seat count is reconciled to a total of 120. Headline bloc totals sum those party predictions across the homepage, forecast, polls and Coalition Builder. Simulated bloc medians, uncertainty ranges and probabilities are separate summaries.

The record the rules come from

Final-week polls against the result for the Netanyahu bloc, every election under the 3.25% threshold. The field is the mean of each firm's final poll other than Direct Polls; the miss is the result less the field.

Lists shown at 4 or 5 seats by most final-week polls since 2015 passed 7 times in 12.

The daily record

Every run that changed the figures, oldest first. Bloc totals here sum the recorded party predictions; the original simulated bloc summaries remain in the data record. The final prediction is preserved through exit polls and the count.

Changes to the rules

Each change to the algorithm, with its date and its effect on the backtest. The reasoning behind each is in the research notes linked below.

• Oct 23 First run: the rules described on this page. Netanyahu bloc 52 seats; 61 or more in 6% of runs.
• Oct 24 Version 2, after a six-memo research round backtested on 2015–2022 (docs/research/polls-method/round1). Poll seats are read back to vote shares by the D'Hondt midpoint, never below the threshold (the old seats/120 rule put every 4-seat list under it). Priors by list class from the record since 2013: Shas +0.8 points, Ra'am +0.6, Joint List +0.4, B'Yachad −0.8, Reservists −0.8, Amcha Yisrael −0.8. Blue and White, withdrawn Oct 19, dropped; 1.5% reserved for lists the polls do not report. Per-list noise sums to zero inside each bloc, with a turnout swing of the Arab lists' own (1.7 seats) and a 1.0-point floor for small lists. Two outlets polled by one firm share its weight; 7-day window. Central seats are each list's median run, reconciled to 120. Backtest, six elections: per-list error 1.77 → 1.36 seats, wrong threshold calls 9 → 4, mean bloc error 2.00 → 1.33; Netanyahu-bloc error 1.00 → 1.83. Netanyahu bloc 54, opposition 51, Arab lists 15; 61 or more in 3% of runs.
• Oct 24 Version 3, after a ten-memo research round backtested on 2015–2022 with a red-team review (docs/research/polls-method/round2). Every list's central seats are the same as in version 2; what changes is the spread and the bloc figure. The Netanyahu-bloc swing is a Student-t with 5 degrees of freedom and a scale of 1.7 seats, the predictive distribution of the next miss given the six on the record, instead of a normal of 3.1 seats (which had put all six past results between the 31st and 66th percentiles). A bloc's central is now the median of its simulated total and is the scored bloc figure; the sum of its lists' centrals is reported beside it, and the bloc centrals can add to 119 or 121 (today 119). An Arab-turnout term on the Joint List, mean-reverting from the 2022 turnout index, replaces its +0.4 prior and the Arab swing; Ra'am keeps its +0.6 prior and its share of the swing. 1.8% reserved for unpolled lists (1.5%); a list printed below the threshold with no share counts 2.5% (2%). Rejected on the evidence: pollster house effects, trend extrapolation and blackout widening, a state-space rebuild, ensembles, a Haredi turnout term, strategic-desertion and skewed threshold terms. Backtest, six elections, three seeds, Filber-firm polls dropped poll by poll: per-list error 1.385 → 1.306 (about three seats of it on the Arab lists; the rest seat rounding that reverses under leave-future-out priors, so the honest size is 0.04–0.07 seat per list, inside the noise), Netanyahu-bloc error 2.11 → 0.83, wrong threshold calls 4 → 4, majority Brier 0.091 → 0.071. Netanyahu bloc 53 (list sum 54), opposition 51, Arab lists 15; 61 or more in 1% of runs. Version 2 is rerun on the same polls each day (data/forecast-v2.json) and scored beside it under its own rule.
• Oct 26 No model change. Since the display change of October 25 every page shows a bloc as the sum of its parties' seats, with the simulated bloc median in a separate column. The scoring rule now scores the bloc total both ways, the headline sum and the median, so that the figure scored is the figure shown; the median stays scored because version 3's claim that it is the better bloc estimator was pre-registered. Figures unchanged: Netanyahu bloc 53, opposition 52 (median 51), Arab lists 15; 61 or more in 1% of runs.

How it will be scored

Fixed before the result. On the official count, for the forecast and for both site averages:

1. Mean absolute seat error per list, over the lists above (the averages: their scaled figures).
2. Secondary diagnostic: absolute error of each simulated bloc median, retained under version 3's recorded scoring rule.
3. Threshold calls: a list is called passing at 50% of runs or more (the averages: by their passes-in-k-of-n rule).
4. Whether the outcome given the highest probability above occurred.
5. Every current poll's final figures, scored the same way, for the pollster record.

What would have to be true for this to be badly wrong

Channel 14 at 62 is right. That requires the field to miss the Netanyahu bloc by 8 to 10 seats, more than twice its worst miss since 2015, or a late swing. The model gives 61 or more a 1% chance; if it happens, the lesson is that right-wing non-response grew after the war faster than six elections of history could show, and the next model's lean should be estimated from 2026 alone. The other way to be wrong is a first-time list passing: both are given about one chance in four or five, and every such list since 2013 has failed.

Reproducible: the script, the data file, the run history, the backtest and the research notes, all in the site's public repository. The same figures are a choice on the home page, the polls page and the Coalition Builder. Home · Polls · Coalition Builder