The model

How Quasimodo forecasts New Zealand elections

Quasimodo is the forecasting model behind Kiwi Polls. Rather than making a single prediction, it works through the whole range of plausible results under MMP: seats for each party, each party's chance of getting into Parliament, the electorate seats, and each bloc's chance of a governing majority.

This page follows one forecast from the published polls to a Parliament, one stage at a time.

100,000
simulated elections in every model run
3
layers of uncertainty: today, drift to election day, and the day itself
2
electorate pools modelled separately: general and Māori
120+
seats: Parliament can grow past 120

See today's forecastThe forecast is on the home page and is updated as new polls are published.

From published polls to a Parliament

Six stages, from the polls going in to the probabilities coming out.

  1. Stage 1Weighting the published polls

    Quasimodo starts with the published polls of party support. It doesn't treat every poll as equal. It builds its own weighted average of them, and three things decide how much each poll counts:

    Track record
    how the polling house has performed in past elections
    Recency
    how recently the poll was taken
    Sample size
    how many people the poll asked
  2. Stage 2How sure can we be?

    The weighted average is an estimate, not a fact, and an election months away could land a long way from today's numbers. Quasimodo builds its uncertainty from three layers and adds them together:

    Today
    how tightly the polls pin down party support right now
    Drift
    how far support could move before election day. This layer grows the further away the election is.
    Election day
    a remaining allowance for the result on the night itself

    So a forecast 18 months out is deliberately vague, and one in the final week is sharp, with no manual adjustment.

    The forecast range narrows as the election nearsIllustration, not a forecast
    How the forecast range narrows as election day gets closerA horizontal band around a central line, running from 18 months before the election on the left to the final week on the right. The inner layer (how tightly today's polls pin down support) and the middle layer (election day) stay the same width. The outer layer (drift) is widest 18 months out and shrinks to nothing in the final week, so the whole band narrows from left to right. Illustration, not a forecast.
    • TodayThe polls' own uncertainty. Steady.
    • Election dayA fixed allowance for the day.
    • DriftWide far out, gone by the final week.
    Read left to right: the further out the forecast is made, the wider the range of plausible results for a party.
  3. Stage 3100,000 simulated elections

    Next comes a Monte Carlo simulation: Quasimodo plays out election day 100,000 times. In each run, a logistic-normal model nudges every party's vote away from the weighted average by an amount drawn from the uncertainty in stage 2.

    The nudges aren't independent of each other. The noise is bloc-aware, and each run's national swing follows the pattern real campaigns do: parties within a bloc tend to rise and fall together, and a gain for one side usually comes with a loss for the other. A National surge tends to arrive with a Labour slide.

    Every run is then turned into electorate results (stage 4) and a full Parliament under the actual rules (stage 5). Stage 6 is what you get when you add them all up.

  4. Stage 4Electorates, general and Māori

    Electorate seats are modelled separately from list seats. Quasimodo uses calibrated relationships that map a party's vote strength onto electorate results, including the edge the stronger of the two major parties tends to have in first-past-the-post seats. It runs those relationships over two pools, treated as distinct contests:

    General electorates
    the larger pool, covering the whole country
    Māori electorates
    modelled as a contest of their own, not as more general seats

    The electorate winners matter twice. They decide who holds those seats, and in stage 5 they can carry a party into Parliament even if its party vote is under 5%.

  5. Stage 5Turning votes into seats under MMP

    Each simulated election is converted into a Parliament using the statutory rules:

    • The 5% threshold. A party needs 5% of the party vote to share in the seats…
    • The electorate-win exception. …unless it wins at least one electorate, in which case it shares in them anyway.
    • Seat retention. An electorate winner whose party doesn't qualify still keeps the seat.
    • Sainte-Laguë. The 120 seats are shared out between the qualifying parties in proportion to their party vote, using the Sainte-Laguë method. Each party's electorate seats count first, and list seats fill the rest.
    • Overhang. If a party wins more electorates than its party vote entitles it to, it keeps them all and Parliament grows past 120.

    So Parliament's size is an output, not a fixed number, and the majority line moves with it: 61 seats in a 120-seat Parliament, 62 in a 122-seat one.

    Worked exampleIllustration, not a forecast
    1. Party E has 4.5% and no electorate, so it misses the threshold and gets no seats.
    2. Party F has only 3%, but it wins an electorate, so it qualifies.
    3. Sainte-Laguë shares the 120 seats between A, B, C, D, F and G on their party vote.
    4. Party G is entitled to 1 seat but wins 3 electorates. It keeps all 3, so there are 2 overhang seats: Parliament has 122 seats and a majority needs 62.
    Fictional parties A to G. 71 electorate seats plus 51 list seats make 122.
    PartyParty voteElectorates wonEntitled (of 120)List seatsTotal seats
    Party A39%38491149
    Party B33%27411441
    Party C11%2141214
    Party D8.5%0111111
    Party EOut: under 5% and no electorate4.5%0–00
    Party FIn: under 5%, but won an electorate3%1434
    Party GIn: won 3 electorates, entitled to 1, keeps all 31%3103
    Parliament100%7112051122
    Worked example: a Parliament of 122 seatsOne block per seat for the fictional parties A, B, C, D, F and G, in that order. Party G's two overhang seats are drawn as outlined blocks at the end. A line after seat 62 marks the majority. Illustration, not a forecast.
  6. Stage 6What comes out

    Stack the 100,000 simulated Parliaments together and count. That count is the forecast:

    Bloc majority
    each bloc's chance of a governing majority, and the chance of a hung Parliament, where neither bloc gets there
    Seats
    a central seat projection for each party, with the spread around it
    Getting in
    each party's chance of entering Parliament at all, by clearing 5% or winning an electorate
    Electorates
    the electorate seats each party is likely to win, general and Māori

    Every figure is a share of simulations: a majority in 57,000 of 100,000 runs is a 57% chance.

Has it worked before?

How Quasimodo does when it's fed the polls from past campaigns.

Back-tests on past elections. Kiwi Polls model estimates, not polls, each made from that campaign's final published polls.
ElectionQuasimodo's most common resultWhat happened
202367 seats for the right bloc, including NZ First68 seats for the right bloc. Off by one seat.
2020A Labour majorityLabour won a majority.

Electorates, checked separately

Run against recent elections, the electorate model reproduces the actual electorate results to within about a seat.

Calibrated, not hand-tuned

How electorates respond to the party vote, how parties move together, how fast uncertainty grows and how much weight each pollster gets are all estimated from past election and polling records using machine-learning and statistical calibration, then checked against past elections.

Two elections is a short track record, and election models have been badly wrong before.

How to read a probability

57% is not a prediction.

If the model gives one side a 57% chance, the other side still wins a bit over two times in five. That isn't a long shot. It's an outcome you should expect to see often.

Treat the forecast like a weather report: sometimes it's wrong, so don't base decisions on this or any other election model.

57 in 100, shown as a gridA ten by ten grid of 100 squares. 57 are filled in dark grey for side A and 43 are left white for side B. Illustration, not a forecast.
Illustration, not a forecast
  • Side A wins: 57 in 100
  • Side B wins: 43 in 100

What it does well, and what it can't do

Good at

  • Turning noisy, conflicting polls into one coherent picture
  • Putting an honest number on uncertainty
  • Applying MMP's seat maths consistently, every time
  • Staying current with new polls

Can't do

  • Foresee shocks like scandals, late swings and campaign moments
  • Escape shared polling error. If every pollster misses in the same direction, so does the model.
  • Call individual electorate races with confidence. There's little public polling on single electorates, so Quasimodo leans on calibrated relationships between the party vote and electorate results.
  • Predict coalition talks. A bloc majority is not yet a government.

Why “Quasimodo”?

Because “Quasimodo predicted all of this.” It's a nod to the Sopranos scene where Bobby Baccalieri gets Quasimodo and Nostradamus tangled up, which makes him a fitting patron saint for election forecasting, where prediction and hindsight are easily confused.

Quasimodo

See what the model says today

Sources: polls are dated by their final fieldwork date, in NZ time. The pollsters and how many of their published polls we use are listed in the sources on our About page.