Herald-Motu Research Poll of Polls: NZ’s 2026 election result forecaster

The Herald has joined forces with Motu Research to create the Election 2026 Poll of Polls. Head of data Chris Knox explains how it works

On October 28 the Herald and Motu-Research released the Poll of poll electorate predictor. This article has been updated to make it easy to find those the latest electorate prediction interactive. The original article continues below. Read about the electorate predictor here.

Electorate predictions are more difficult than party vote predictions, but electorate outcomes could be more important in this election than in any other MMP election.

The Herald-Motu Research electorate predictor uses the overall party vote, candidate history, electorate polls and way previous national-level party votes swings played out in individual elections to predict the outcome for almost every electorate this election.

Use the interactive below to modify the outcome of races and see how they impact the election outcome.

The Poll of polls model can predict the likely spread of seats – and their probability of being able to govern – for different potential coalitions. The original version of this graphic was based solely on predicted party vote.

The possibility of overhangs changes the range of potential outcomes. The newer seat distribution graphic takes this into account, but it also has a higher level of uncertainly.

The Herald-Motu Research Poll of polls model is not an polling average – it is a model that predicts the party vote in a hypothetical election held on the day the model was run. The table below shows the current differences between polling average estimates and the model’s estimates.

When making comparisons between polls and the model’s output it is also helpful to consider the model’s prediction range and not just a single number.

Original article

Politicians like to say the only poll that matters is the one on election day.

They’re right - but our new aggregated poll model gets you closer to that result than any single poll is likely to do.

It goes beyond the snapshot view of individual polls, providing insights into trends in voter preference and is designed to simulate the election outcome.

Motu Research’s senior fellow Stuart Donovan told the Herald he was inspired by the success of overseas poll aggregation models to develop a New Zealand version.

“The New Zealand political environment is almost easier to model than overseas because of MMP,” Donovan said.

“In the model, we feed in all the polls as well as all the general elections going back to 2014.

“We include general elections and we treat them as a large poll. And they are obviously the poll that we are most interested in.

“So what that model’s trying to do is predict the party vote that an individual party would receive if an election was held on a given day.”

The model has simulated 4000 hypothetical elections every day since September 15, 2014. Comparing the outcomes of those simulations provides an estimated voter preference on any given day.

If one bloc of parties has more than 61 of the standard 120 seats in Parliament for every one of the 4000 simulations, then we can be very confident that New Zealanders would have voted for that bloc on that day.

But if half of the simulations go one way and half the other then we know things were very close.

Each time a new poll arrives, the entire model is updated.

The chart below shows the modelled support for the Government coalition and the current opposition bloc from November 27, 2023 when the current government was sworn in until now.

Voters would have re-elected the coalition throughout its time in government, although the gap has narrowed. Support for the opposition was highest at the end of March 2025, but even then the model predicted the coalition had an 75% chance of being re-elected, despite greater fluctuations in support for each political party.

Donovan noted, “When you inspect the model results, you can see parties tend to attract support from other parties that you might think are similar to them.

“The net result is you can see in polls and in the model results that there’s quite a lot of variability in levels of ... the support for individual parties, but when you aggregate up to the government as a whole or the opposition as a whole, it’s much more stable.”

The most recent model run, following the August release of the latest Curia TPU poll, puts the coalition ahead in 61% of the 4000 simulations, giving them an average of 61 seats.

The Poll of Polls model doesn’t just make the single prediction that, given recent polling, the coalition would win 50.3% of the party vote and 61 seats. Rather, each of the 4000 simulations produces a result that is consistent with the model’s interpretation of current polling data. The more likely a scenario, the more often we will see it in the results. So 50.3% is the average combined party vote support across all those simulations for National, Act, and NZ First.

The most recently released poll, from Curia TPU, puts support for the coalition parties at 47.4%. This isn’t far from the model’s estimate of 50.3%. In fact, 50.3% is within the poll’s margin of error.

You could think of the range of outcomes produced by the model as taking the polling margin of error and computing it forward. For example, in just one of those 4000 simulations the coalition wins 74 seats. Essentially this is an outlier scenario where the model asks, “What if all three government parties are being under-polled and the real support for all for National, Act, and NZ First is at the top end of the margin of error, while Labour, Green and Te Pāti Māori are all being over-polled?” It’s possible, but unlikely given the previous performance of polls in New Zealand - so it is only produced in the model once.

The model looks at how well previous polls predicted the actual election outcome and builds what are called “pollster effects”.

Donovan said, “For each pollster, we have a measure of how close they have been.

“We also have a measure for every pollster of their relative statistical bias for or against different parties.

“We can see, that some pollsters - for whatever reason, whatever aspect of their methodology - tend to oversample support from a particular party.

“It’s not a criticism because polling’s really hard. It’s just something that we want to control for when we’re interpreting the polls.”

Asked how he knew the model was correct, Donovan responded with a common statistical mantra, “All models are wrong, but some are useful”.

He added, “The way we tested the model’s performance was we estimated it on data up to, but not including, the election result.

“Then we used the model to predict the election result for the different parties using all of the polls up until the last poll before the election.

“The model was very good at predicting the actual outcome for each party. We did that for every election, so 2017, 2020, 2023 and in each election the model performed very well.”

Asked about so-called “rogue polls” - results so different from other polls that they are dismissed as unreliable - Donovan said the model took a wait-and-see approach.

“The response to an unusual poll will be relatively muted initially.

“But if there is subsequently a second poll that is similar to that poll, then that will start to really bend the curve.”

On June 19, a second version of the Poll of polls model was released. This version included estimates for the Opportunity Party.

Motu Research is an independent economics and policy research institute that helps decision-makers in Aotearoa New Zealand grapple with complex social, environmental and economic issues. The model for the Herald-Motu Research poll of polls was designed by Motu Research. Motu Research advises only on technical matters related to the model; political analysis of the results is the domain of the NZ Herald and not Motu Research.

Chris Knox is a scientist turned data journalist who investigates the stories behind the numbers, and creates interactives for Herald readers to explore them.