How the numbers are calculated

The trend line

The line is not a poll. It is a weighted average of every poll near a given date, so a single surprising result moves it a little rather than a lot. Two things set a poll’s weight. A poll counts for more the closer its fieldwork sits to the date being estimated, fading smoothly over about 21 days on either side; and a larger poll counts for more than a smaller one, in proportion to how many people it surveyed.

Because larger samples carry more weight, the averaging has an “effective sample size” — the number of respondents the weighted average is really resting on. Where too few polls sit near a date to be meaningful, no point is drawn rather than a shaky one.

The confidence band

The margin of error a pollster quotes assumes a perfect random sample. Real polls are not: who answers the phone, how the sample is weighted, how undecided voters are handled, and voters changing their minds between fieldwork and polling day all add error that the quoted margin ignores. So the band drawn here is wider than sampling error alone — 2.0 times wider.

That figure is not borrowed and not a guess. It comes from comparing the final polls before every New Zealand election since MMP began in 1996 against the actual result — 39 polls across 10 elections. The typical miss was about twice what sampling theory alone predicts. Widening the band by less would leave it too narrow: at a factor of 1.7, the value a neighbouring project calibrated on Australian data, only 88.8% of past results fell inside a nominal 95% band. At 2.0 it is 94.4%.

Two things this band cannot do. It cannot be shrunk by averaging more polls: about 61% of the error is a miss every pollster makes in the same direction at the same election, which more polling does not touch. And it is built from final polls, the most accurate a campaign produces — so a band on a poll taken months out is, if anything, still a little too narrow.

One number is used for every party, large and small. It is tempting to treat the small parties differently, since they are the ones near the threshold — but measured against the record, polling error does not behave differently for them, and a two-tier rule made the answer worse exactly where it matters most.

Why so many 5% calls come back “too close”

Under MMP a party needs 5% of the party vote, or one electorate seat, to take list seats. The band above is wide enough that a party polling near 5% often has an interval reaching over the line on both sides. The table shows how close to the line a party has to be before a poll can place it on one side, and how that tightens as more polls accumulate near polling day.

A three-week window holding three polls of a thousand people each is worth roughly 3,000 in effective sample — so a party anywhere between about 3.7% and 6.6% cannot be told apart from the 5% line, and is shown as too close to call. The band tightens as more polls accumulate near polling day.

From votes to seats

The seat projection is the one modelled page on this site, and it is built to carry the uncertainty above all the way through rather than throw it away. Seat allocation under MMP is sharply non-linear — a party crossing 5%, or winning one more electorate than its party vote earns, moves seats in a jump — so there is no tidy formula for “the range of seats given the range of votes”. Instead the party vote is drawn from the same band described above and allocated 20,000 times.

Each run draws a party-vote share for every party from its current estimate and that estimate’s uncertainty, decides who qualifies — 5% of the party vote or an electorate seat — and shares out 120 seats among the qualifiers by the Sainte-Laguë method, the divisor method New Zealand law prescribes. A party that wins more electorates than its share earns keeps the extra seats, and the House grows past 120: that is an overhang, and it is why the projected House is usually a seat or two above 120. Across all the runs, each party’s seats become a range, not a point, and a coalition’s majority becomes a percentage of runs rather than a yes or no.

Electorate wins are assumed, because no poll measures them. The model floors the allocation only at the seats a party holds largely regardless of the national swing — Te Pāti Māori’s Māori electorates, ACT’s Epsom, the Greens’ urban seats. A major party’s general electorates move with its party vote and never exceed its proportional entitlement, so they are not floored; flooring them would manufacture an overhang that has never happened. This is why every overhang the model produces comes from a small party, as every real one has. The page lets you turn the pivotal electorates on and off.

Two honest limits. The draws are marginal: each party is drawn from its own band independently, so the model does not capture the shared miss every pollster makes together, nor the fact that the shares compete for the same voters. That makes a knife-edge coalition or wasted-vote probability approximate — a little more certain than it should read. And the whole thing is deterministic: a fixed seed (20261107) means the same inputs always give the same seat numbers, so the projection can be re-run and argued with rather than reshuffling on every load.

What the calibration rests on

The widening factor is derived from 39 final polls across the 10 MMP elections from 1996 to 2023. Official results were checked against Electoral Commission files for 9 of them — the exception is 1996, whose official result could not be read from the source file and was transcribed instead. Of the 39 polls, 18 carried a sourced sample size; the derivation of the factor itself rests on those. Where a sample size could not be found it is recorded as unknown, never estimated.

Because it rests on 5 elections with sourced sample sizes, the factor carries real uncertainty of its own: the underlying estimate is 1.98, with a 95% range of 1.69 to 2.24. It is rounded to 2.0 for use. It is an average over past elections and does not anticipate anything specific to the next one.