Method

20 sources are catalogued and 7 are read on a schedule — poll compilations, the Central Elections Committee, and the official gazette for certified results. Every fetch is stored as a dated snapshot and none is ever overwritten.

That last part matters more than it sounds. The compilations revise history: a poll's figures can change under them after publication. Keeping the original bytes is what lets a forecast published in August still be rebuilt, months later, from exactly what it was built from.

Pollsters do not agree on what to call a party, and parties do not stay still. One list arrives under three different spellings; two merge mid-campaign and the field takes weeks to catch up. Every list is folded to one canonical name, mergers and splits are tracked as they happen, and a poll published by two outlets is counted once.

Get this wrong and the failure is silent: a party quietly becomes two, each with half its support, and nothing about the output looks broken.

The 101 polls behind the current forecast come from 10 polling houses. They are combined into one reading of where the parties stand, with recent polls counting for more than older ones. Polls commissioned by Channel 14 carry 50% of the weight of any other poll the same age: they have put the Netanyahu bloc well above the same week's polls from other outlets. The weight is a judgment, not a measured correction, and polls from other outlets that also run above the field count in full.

No single poll moves the result much, which is the point. A striking number from one house is the kind of thing that leads a news bulletin and should not move a forecast; what moves this one is several houses agreeing.

An average is not a forecast. It says where the parties are, not how wrong it might be. So the election is simulated 20,000 times, each run drawing a different plausible result given how far polls have missed in the past.

The seat apportionment runs inside every draw, not once at the end: each simulated election divides the 120 seats by the same d'Hondt rule the law uses, surplus-vote agreements included. Because the model works in seats rather than votes, the threshold becomes a floor: a list that a draw leaves between nothing and 4 seats goes to 4, with a chance that grows the closer it came, and otherwise to nothing, and the other lists absorb the difference. That is what makes the arithmetic behave like the real thing: a list near the line comes out of a simulated election as it would come out of the count, with nothing or with about 4 seats, not left at 2, where no election can leave a list.

A model that has never been wrong in public is a model nobody has checked. This one is refit against 8 past elections, each time seeing only what was known before that election was held.

That is the difference between a forecast and a story. Fitting a model on an election you already know the result of will always look impressive; it tells you nothing about the next one.

Individual decisions get written up as they are made.

• Analysis
• What this particular forecast is made of
• What this forecast does not claim