Nine polls, not sixty-nine

For most of today the top of this page read Netanyahu bloc leads with about 55 seats, Opposition bloc with 54, and Arab parties with 11. The polls did not say that. Of the 69 polls in the window, 52 had the opposition bloc ahead and 17 had the Netanyahu bloc ahead, and all 17 were commissioned by Channel 14 or i24 News. No other outlet’s poll has put the Netanyahu bloc ahead since July. Weighted the way the model weights them, the polls’ own bloc sums came to 53 for the Netanyahu bloc, 56 for the opposition and 11 for the Arab parties.

The page now reads Opposition bloc leads with about 57 seats, Netanyahu bloc with 51, and Arab parties with 12, from 9 polls. This piece is about how the first sentence got written, and why the second one rests on 9 polls rather than 69.

A list that did not exist yet

The model averages each list over the polls that measured it. That is the right rule for a list one pollster left out of one table, and it is written into the method for a reason: reading a blank as a zero would drag every small party toward nothing in proportion to how often it is omitted.

It is the wrong rule for a list that did not exist when the poll was taken. Amcha Israel entered on 25 August, and its level came from the polls that had a column for it. Likud and Religious Zionism were averaged over all 69, and the 60 polls from before Winter’s entry, which still carried most of the weight, had his voters inside them. So the model added Amcha Israel’s seats on top of lists that still contained them. The vector summed to about 122 seats, and the step that scales it back to 120 took the excess from every list in every bloc. The opposition bloc paid about a seat for a shift that happened entirely inside the Netanyahu bloc.

Then the printed table did the rest. Amcha Israel and Religious Zionism both sat at a median of exactly 4, each with roughly an even chance of entering, and the table seats a median of 4 in full. The simulations still had the opposition bloc as the larger one in 52.6% of draws; the table said 55 to 54; the headline was generated from the table.

The two outside aggregators we compare against printed the same thing, and for the same reason. The 2026 compilation’s two-week average read 55, 54 and 11, and its own methods note says a counted party “uses only the polls that actually seat it”. TheMadad’s per-party average adds up to about 125 seats. Agreeing with them was agreeing on the arithmetic, not on the polls.

What changed

A new list is a break. From today, a list’s first seated poll starts the evidence. Once a list has been seated in at least half the polls since it appeared, and in at least 5 of them, polls fielded before it appeared are dropped for every list, not just for it. A list the tables carried all along below the line and which then crosses does not count: that is a threshold crossing, not an arrival. For Amcha Israel that leaves the 9 polls fielded since 26 August. The count at the top of the page says 9 because 9 is what the forecast is built from. Scored against the 5 past elections the model can be scored on, the rule fires 14 times, every time for a list that had genuinely just formed. It lowers the average seat error and leaves the final-week horizons where they were.

“Leads” needs the simulations’ agreement. The headline now reads how often each bloc is the largest across the draws before it says a bloc leads. It takes 3 draws in 5. Short of that, the two blocs are reported as close, in the table’s order with the table’s numbers and no verb between them. Today the opposition bloc is the larger in 69% of simulations, so the word stands.

Channel 14’s polls carry half weight. Every poll Channel 14 has commissioned this campaign has put the Netanyahu bloc 10 to 12 seats above the same week’s other firms, and at equal weight those polls alone moved the bloc by nearly a seat and a half. This is a judgment, not a measurement: the model’s house-effect estimator centres every firm against the field it is in, so it can say that Channel 14’s polls differ from the field but not that the field is right. Halving their weight is the mildest form of that judgment. Nothing else is down-weighted.

Where that leaves the forecast

The Netanyahu bloc is on 51 seats, the opposition bloc on 57 and the Arab parties on 12. In the simulations the Netanyahu bloc reaches 61 in 2.9% of draws and the opposition bloc in 14.6%. The opposition and the Arab parties together are 69 on the printed table; that is arithmetic, and those 12 seats are two lists that do not act as one.

Yashar is the most likely largest list at 63%, with Likud at 37%. This morning the two were level at 23. The 9 polls since Winter’s entry have Yashar at 21 to 26 and Likud at 19 to 31, and only Channel 14 and i24 News have Likud ahead; Maagar Mochot has the two level.

Amcha Israel is on 5 seats and clears 4 in 77% of simulations. Religious Zionism prints at nothing, and here the model is still wrong in a way worth saying plainly: 4 of the 9 polls seat it on 4 or 5, and 5 put it under the line, and the model reads that as a list averaging 2 seats with a 13% chance of 4 or more. The truth is closer to a coin toss between 0 and 4. That is the threshold problem the method already describes, and it is the next thing to fix.

The piece published earlier today, “Two coin tosses, three seats each way”, described the 55 to 54 table accurately and drew the wrong conclusion from it. The table had not moved because of two coin tosses. It had moved because one list was being counted twice.

Addendum, later the same day. The Religious Zionism problem above is fixed, and the fix is measured. The model now simulates a list the polls disagree about as the two outcomes they describe rather than around their average: seated at its polled level in the share of simulations the polls seat it, and at nothing in the rest, with the chamber renormalised. Religious Zionism now wins nothing in 58% of simulations and 4 or more seats in 33%, against 16% and 13% this morning. Its mean is unchanged, so no bloc figure moved. Scored on the 5 past elections, the change improves the forecast’s spread scores at every horizon and leaves the seat and bloc errors where they were.