Polling evidence
The model uses national polls that meet its rules for inclusion. It keeps two kinds of figures separate: the percentage of voters supporting each list (vote share), and the number of seats reported by the poll. Different scenarios from one survey share the weight of one poll. A missing figure is not treated as zero, and a seat count is not turned into a vote percentage that the poll never reported.
Each date uses only information that was available by that day. The model does not rank polling companies, exclude polls because of the company’s name, correct for a company’s tendency to favour particular parties, or adjust for expected turnout. The September 27 source set has 222 poll entries: 73 report vote shares and 220 report seats. Many report both, so these numbers should not be added together. They also do not mean every list uses all these polls.
Estimating vote share
For each list, the model follows its vote-share polls over time. It estimates how much support is changing and how much of the difference between polls may be ordinary polling error. This is called a local-level model. A poll’s weight depends on those estimates and a minimum allowance for polling error. It is not a simple average of recent polls, and older polls do not all lose weight at one fixed rate.
The September 27 update uses settings that account for this minimum allowance for polling error. Their performance is still being evaluated.
Simulating votes and seats
Vote Share Only creates many possible election-day results around the model’s best estimate for each list. The range of possibilities depends on the time left until the election and how uncertain each list’s support is. The current model also limits how widely the simulated support of small lists can vary. Each simulated election result is called a draw. In each draw, lists below 3.25% of valid votes receive no seats; seats are then divided under the legal rules.
September 27 uses 100,000 simulated election results; each earlier displayed date uses 20,000. Average seats (also called expected seats) average the seats a list receives across these results. The main seat figure is calculated differently: we first average each list’s simulated vote share, then divide the seats using those average vote shares. The two figures can differ.
For a list near the threshold, the 80% seat range uses only simulated results in which the list enters the Knesset. Of those results, 80% fall within the displayed range. The chance of getting no seats is shown separately. This describes the model’s simulations; it does not guarantee that real election results will fall in the range 80% of the time.
A bloc is a named group of lists whose seats we count together. Its 61+ chance is the percentage of simulated elections in which those lists win at least 61 seats in total. We count all lists from the same simulated election together, rather than adding separate best estimates. Reaching 61 gives the group a seat majority, but does not tell us whether its members would agree to form a government. These figures assume that no list outside the model passes the threshold.
Vote and Seat Fusion
Vote and Seat Fusion starts with the same simulated results as Vote Share Only. It gives more weight to results that better match additional polls reporting only seat counts. You can switch between the two approaches on the forecast page to see how the extra polling changes the picture. To check Fusion probabilities from the downloadable simulations, use their Fusion weights rather than counting every result equally.
What the model can and cannot tell us
Tests using past elections helped shape Vote Share Only. Limiting the simulated spread of small lists involves a trade-off in the test scores: an improvement on one measure does not mean an improvement on every measure. Some ways a small list might surge have not yet been tested. Some recent poll sources are also still being checked against filings with the Central Elections Committee.
The ten dates before September 27 were calculated afterward, using only information available on each day. They are reconstructions, not forecasts published on those dates. We leave dates unavailable when there is not enough evidence for the lists now in the model. Any result shown for a future date is a projection, not a forecast that was already published on that day.