Which parties count as being “in the race”
Israel’s electoral threshold is 3.25% of the vote, worth roughly four of the 120 seats, and a party below it wins no seats. Pollsters publish seat projections rather than vote shares, so the threshold has already been applied before we see a number: a party that a pollster leaves out (or shows at zero) is one that pollster has judged to fall short. We do not apply the threshold ourselves. Our rule decides what to do when pollsters disagree about whether a party clears it.
We count a party in the two-week average if it passes either test: it is seated in at least half the window’s polls, or in at least half of the six most recent. A counted party’s average uses only the polls that seat it, so a firm that leaves the party out lowers the party’s chance of being counted without pulling its seats toward zero.
We added the second test on 30 July 2026, after the case of Yoaz Hendel’s list, which has run since 6 September with Yaron Zelekha’s as the Reservists and the Economic Party. The list crossed the threshold partway through that window: it appeared in 7 of the 17 polls then in range, under half, so it was dropped from every average, although three of the last four pollsters seated it at four or five. The first test counts how many polls seat a party and ignores which, so on its own it leaves a rising party out of the averages for up to two weeks. We keep the older test alongside because the second alone would admit a single fresh outlier: on that date Blue & White failed both, seated in 1 of those 17 polls and by no pollster since 16 July, and under both tests a single old poll does not add a party to the averages.
Adding a party changes every other party’s seats, because the projection is rescaled to total exactly 120: counting Hendel’s four seats shifted the blocs from 54/55/11 to 53/58/9, and neither bloc reaches 61 seats under either count. A party near the line can also enter or leave the averages between refreshes, and the change that follows in other parties’ seats is arithmetic, not a change in support.
We now apply the same rule in the nowcast simulation. Until then, membership in the simulation was judged after house effects (each firm’s persistent lean against the other firms) were evened out, so max(0, 0 − a negative lean) could produce a seat share for a party a pollster had placed below the threshold, which put Balad (seated by no pollster at all) and Blue & White into the simulation. We now judge membership on the raw published seats in every panel: a house effect can change a party’s seat number but not whether the party is counted as running.
One survey, one entry - guarding against double-counting
Because we pool published results, each real survey must be counted once. The same poll can enter the data under different labels, and a poll counted twice shifts every average without any visible sign. Our audit of the full data in July 2026 found three ways a poll is duplicated:
- Re-attribution - one Channel 13 poll was logged first under one polling firm (Maagar Mochot) and later another (Midgam) as its attribution was corrected, leaving two entries for one survey.
- Re-dating - a poll appearing under two nearby dates, such as a fieldwork-end date beside a publication date.
- Mis-dating - one Lazar/Maariv poll was listed under both a May and an April date a month apart. The April copy reported the merged “Together” (Bennett + Lapid) party, which did not exist in early April and merged about three weeks later; the same firm was polling those two parties separately a week on either side of that date. A party cannot appear merged three weeks before it merges, so we inferred that the April date was an error and that the two entries were one poll.
We treat two entries as the same survey when they have an identical seat breakdown from the same commissioning outlet. The outlet that ordered a poll, together with its numbers, identifies it; the polling firm’s name and the exact date vary across sources and do not. We keep the better-sourced copy (a sample size, an independent check) and drop the duplicate; where identical numbers are far apart in time and the case is ambiguous, we keep both and flag them rather than guess. Before this audit, a poll strongly favouring the Change bloc was counted twice, which made the opposition’s lead look steadier than the once-counted data support.
The nowcast simulation
The nowcast figures are seat arithmetic: we count which combinations could reach 61 seats, not who would agree to govern together. We report results with and without Joint List (Hadash-Ta’al-Balad), Ra’am, the parties commonly called Arab-majority, because the main Change-bloc parties (Bennett, Lapid, Eisenkot) have said they will not rely on them, a position held more firmly since 7 October 2023. For that reason the opposition can lead in seats and still fall short of a government. The position is a political choice and can change: in 2021 Ra’am joined a governing coalition. In the simulation a whole bloc can miss in the same direction at once, because real polling errors are shared across allied parties rather than independent, which widens the spread. That spread still comes only from how far recent polls disagree, which is usually narrower than how far polls miss on the day, so the figures are broad signals, not exact odds.
A nowcast is an estimate of current opinion from current polls, as opposed to a forecast of the eventual result. We treat the recent polls as a snapshot of opinion and turn their disagreement into probabilities:
- Recent window. For each party we take the polls from the last 2 weeks (all pollsters pooled), rescale each poll to the 120-seat Knesset, and average the party over the polls that list it, keeping the parties that at least half the pollsters list.
- Centre & spread. A party’s average seats in that window is its central estimate; the standard deviation of those polls (at least 1 seat) is its uncertainty, covering both real change and disagreement between pollsters.
- Simulate 10,000 elections. Each run is one simulated election, in which every party’s seats are drawn at random from a normal distribution centred on its average, with that spread; allied parties shift partly together, since a real polling miss tends to affect a whole bloc.
- Threshold. A draw below the 3.25% electoral threshold (~4 seats) becomes zero, and the party fails to enter. Parties near the line can therefore swing the whole result.
- Re-balance. The surviving parties are rescaled so the seats sum to 120, then rounded.
- Count. We total each bloc’s seats and record which combinations reach 61 seats.
Each percentage is the share of the 10,000 runs in which an outcome occurred. We chose 10,000 runs because the figures are then stable to within about half a percentage point (the Monte-Carlo error is roughly √(p(1−p)/N) ≈ 0.3–0.5 pp); more runs would not change the headline, since the model’s assumptions contribute more of the uncertainty. A fixed random seed makes the numbers reproducible.
Why allied parties rise together - the “correlated-error” term. In each simulated election we shift every party a little away from its polled seats, because polls are never exact. Real misses come in groups: if ultra-Orthodox turnout is higher than the pollsters expected, Shas and United Torah Judaism both finish above their polls, since both depend on the same voters. Had we shifted each party independently, a gain for Shas would often be offset by a chance fall for United Torah Judaism, and the Netanyahu bloc’s total would vary too little between runs and look more certain than it is. We therefore let the parties within a bloc vary partly together, with a “within-bloc correlation” of 0.35. We set 0.35 by judgment; it was not measured from past Israeli elections.
Our simulation is a deliberately simple version of the Monte-Carlo approach that professional election forecasters use (FiveThirtyEight, for example, which popularised the distinction between a nowcast and a forecast). It omits the parts of those models that make them elaborate: weighting pollsters by track record, fitting the full party-by-party error-correlation matrix from past elections, calibrating the spread to how far polls have historically missed at this distance from election day, and blending in non-poll “fundamentals.” (The toggle at the top of the tab evens out each pollster’s house effect; it is on by default, so the headline figures are already adjusted, though only relative to the other firms.)
Assumptions and limits. The uncertainty reflects disagreement between published polls; we do not hold the pollsters’ raw data, so we cannot model each survey’s own sampling and weighting. The simulation is seat arithmetic: we count who could reach 61 seats, not who would agree to govern together (hence the separate columns with and without the parties commonly called Arab-majority). It is also a nowcast rather than an October forecast, and the result could differ from it as opinion changed before election day. The spread is built from how far apart this cycle’s polls are, which differs from how far polls miss on the day, so the figures are broad signals rather than exact odds.
The width of the bands against past polling misses. The simulated range for a bloc is much wider than the error in the final polls at recent elections. At the last five elections the final polls put the right bloc within 0.2, 1.1, 0.4, 1.3 and 3.4 seats of its result, a spread of about 1.7 seats, whereas the simulation gives the Netanyahu bloc a spread of about 7 seats, four times as much. Part of the difference is timing, because the simulation uses polls taken weeks before the vote rather than the last ones published. A larger part comes from the firms’ disagreement with each other, from which the spread is computed; that disagreement measures real uncertainty about current opinion rather than sampling noise, and we do not narrow it for the number of polls, because one firm’s house effect is not cancelled by another’s. The rest is the shared bloc shock, set at a correlation of 0.35 between any two parties in the same bloc, which widens the bloc totals by design. The bands are therefore cautious by construction and are not an estimate of how far the polls will turn out to have missed.
The forecast to 27 October, and how far apart the pollsters are
The nowcast estimates current opinion. The forecast estimates the result on election day, and its figures appear in the gap between the pollsters panel in the Nowcast. We start from the same seat totals as the scoreboard, so the two agree by construction, and add three sources of error: the uncertainty in the current level, how far opinion has drifted over a period of this length in this campaign, and how far the final polls missed at the last five Israeli elections (between 0 and 3.4 seats on the Netanyahu bloc, and in the same direction four times out of five).
Our treatment of the first source of error is a judgement. Eight pollsters have published in the last fortnight, and they are about 4.9 seats apart on the Netanyahu bloc, from the high fifties to the low sixties at one end and the high forties at the other. Until 9 October 2026 we divided that spread by the square root of the number of pollsters, the usual standard error of a mean, which reduced it to 1.7 seats. That division measures how precisely the average of these eight is known. A forecast needs how far the election will fall from that average, and the two are equal only if each pollster’s difference from the average is an independent error that cancels in the average. Past elections suggest otherwise: the final polls missed the same bloc in the same direction at four of the last five elections, an error the pollsters share rather than eight unrelated ones. We now use the full spread, which widens the range and raises the chance that the Netanyahu bloc reaches 61 seats from about 2% to about 22%.
Using the full spread errs towards caution. The five-election error was measured on poll averages, so it already contains some of the same disagreement between pollsters, and using both counts a little of it twice. Eight pollsters in one campaign are too few to estimate how strongly their leans are correlated, and without that estimate the choice is a judgement.
How to cite
The data change every day, so a citation needs to name the day; this citation names the last poll included in the build:
Levi, Ron. Israel 2026 Election Forecast and Polls. University of Toronto. Includes polls to 23 October 2026. https://israel-polls-2026.pages.dev/ (accessed [your date]).
Released under CC BY 4.0, so it may be reused, redistributed and built on, including commercially, as long as it is credited. The downloadable CSV carries the same credit on every row, in its compiled_by, source_dataset, source_url and license columns, so a copy of the file stays attributed wherever it is copied.
Surplus-vote agreements
When the votes are counted, the 120 seats are shared out in proportion to each list’s share, and the division never comes out even: every list is left with votes too few for another seat. In a surplus-vote agreement, two lists agree before polling day to have those leftovers added together. If the combined total is enough for one more seat, that seat goes to whichever of the two is closer to earning it, so a pair can take a seat that neither would have taken alone. The agreements are worth a seat or two across the whole Knesset, and nothing to a pair in which either partner finishes below the threshold. Each party may pair with one other.
Who has signed
Five pairs signed, none of them across the two blocs; every deal, its date and its source are on the surplus-vote agreements page, which also lists who went unpaired. Filing closed at the Central Elections Committee at noon on 16 October, so these five are the set that will count on the night.
Why they are not modelled
Our numbers are the pollsters’ own seat figures, averaged and rounded to 120; that rounding is arithmetic, not the election’s allocation. A bloc shown a seat short could reach 61 seats on the night without any poll having been wrong. A partner who finishes below the threshold contributes nothing.
What we leave out, and why
Two rules govern which polls enter an average. First, we list individual published polls only: other aggregators’ and models’ averages are excluded, because pooling an average of polls with the polls it already contains counts the same surveys twice. Second, a poll has to be sourced well enough to check. Where it is not, we record the figures and the reason and leave them out of the averages rather than dropping them without notice, so that the data list what they omit instead of implying they are complete.
Holding an item off is not a judgment that its numbers are wrong. Each is recorded in data/held_off.csv with its reason, so it can be added to the counted data once the missing source appears. The DRI poll of 7 September was held off until 24 September and now counts, with its sourcing described beside its figures.
The trend line - pooling, the uncertainty band, and house effects
For the seats-over-time line we pool every pollster’s published polls into one curve with a Gaussian kernel (nearby polls count more); a model-based (Kalman) option smooths the same series as election forecasters do. The shaded 95% band is computed from how much the polls near each date disagree, so it is wider where polls are fewer. A house effect is a firm’s persistent difference from the consensus of all pollsters; “evened out” means we subtract that difference before pooling, and we compute the band from the raw spread so that the adjustment never overstates precision. We measure the difference, not its cause. Each firm’s current house effect per party appears beside the chart on The parties.
Robustness of the findings
The deeper analyses
For Co-movement & factor structure (Dive into the data), the matrix gives the correlation between two parties’ seat series across polls; we apply principal-component analysis to the standardized (and compositionally transformed) series and report a range for the share of shared movement, because that share changes with the choice of method. For Challenger parties we use fuzzy-set Qualitative Comparative Analysis (Ragin): each party’s membership in “large” is calibrated from its recent seats, and we report which combinations of traits (newly formed, centrist) occur together with size, with consistency and coverage in the footnote. For Bargaining power (Coalition builder) we use the Banzhaf index, which counts, across every possible coalition and with no regard to political compatibility, how often a party’s seats take a group that is short of 61 seats up to 61 seats; a second column, If pledges over whom to partner with are kept, repeats the count with only the coalitions consistent with the publicly ruled-out pairings. The live figures and fuller notes for each analysis appear beside it.
