On the other side of things, National and Act’s chances continue to soar. They have a 95.1 per cent chance of forming a government if the election were held this weekend, and an 87 per cent chance of forming a government for the actual election date.
The Herald’s poll of polls takes account of a range of public polls. The most recent simulation takes into account recent Taxpayers Union-Curia, Talbot Mills, and Newshub Reid Research polls.
The simulation’s best estimate for National is that it will poll 36.1 per cent, followed by Labour on 28.6 per cent. It thinks the Greens will poll 11.8 per cent, and Act 11.3 per cent. NZ First would poll 4.7 per cent, Te Pāti Māori would poll 3.1 per cent, and Top 2.1 per cent.
How the Poll of Polls works
As part of its 2023 election coverage, the Herald has launched its Poll of Polls, which combines polling from different pollsters to predict the party vote for the 2023 election.
The model imagines that in any given week there exists an unobserved voting intention that is partially measured via opinion polls and is accurately measured once every three years by an election.
The model can also make predictions about how voting intentions can evolve between now and election day.
This approach enables us to estimate each polling organisation’s accuracy in previous elections and then use that to inform our predictions.
The Herald has based its model on a New Zealand election forecast developed by statistician and data scientist Peter Ellis. Ellis developed the model in a private capacity prior to taking on his current role as director of the Statistics for Development Division at the Pacific Community (SPC).
Ellis also used the model to forecast the Australian election. Similar models have been used to make predictions about the German and Swedish elections. Both Germany and Sweden have proportional representation electoral systems similar to New Zealand’s MMP system.
Our Poll of Polls combines results from a range of pollsters, who are signatories to the New Zealand Political Polling Code, specifically Curia, Kantar Public, Talbot Mills, and Reid Research.
Details of the model, including the source code, are available here. Between now and the election, we will be looking to improve some aspects of the model, in particular the handling of polls which do not provide polling data for some of the smaller parties.
Currently, in order to be included in the model, a party must have polled over 2.5 per cent in at least three polls.
It is assumed that parties currently holding an electorate seat retain them and no other parties win an electorate seat. Future versions will enable readers to modify this assumption.
