Herald election forecasts explained

Herald's forecast model is built using all the polling data and previous election results.

The make-up of the next government will be determined by Winston Peters - that's the prediction of the Herald's exclusive election forecast.

The predictive model is the first of its kind from a New Zealand news organisation. It allows Herald readers to cut through the noise of individual polls and see a much clearer estimate of what is likely to happen on September 23.

It predicts there is no possible scenario for a government without New Zealand First.
Peters' party is predicted to win 12-18 seats, leaving National and Labour short of a clear path to a majority.

The model uses data from every major poll conducted in the past year and the results of every election since 1999.

It runs 10,000 simulations to forecast this year's results. The simulations create a range for each possible outcome. The more times the simulations lead to the same outcome, the more likely that outcome is.

The model is currently predicting the party vote at national and electoral levels. Once the electorate candidate lists are confirmed it will also predict the electorate results.

Click here to view the interactive

How is the model built?

The results of each poll are adjusted according to previous biases towards parties. They are then weighted based upon the age of the poll.

An adjustment is made for the amount by which a poll was off in predicting the result.
The party vote at electorate level is estimated by combining the adjusted poll results with the national trends.

The model was commissioned by the Herald and built by University of Auckland student Wil Undy.

How accurate is the model?

It allows for more context than individual polls, which might make for an 'explosive' headline on release but can take less account of the uncertainty inherent in predictions.

According to Professor Thomas Lumley from the Department of Statistics at the University of Auckland, there are two advantages of a model-based summary of polls, such as this.

"First, you can use more information than a single poll, which should lead to more accurate estimates that don't bounce around so much every time a new poll comes out.

"Second, and maybe more importantly, you can get better uncertainty estimates. The standard uncertainty estimates on a single poll, even though they seem very wide, are optimistic.

"Modelling doesn't guarantee good estimates: in last year's US election, many groups still seriously underestimated the uncertainty in their predictions. But modelling allows uncertainty to be taken seriously."

Will the model be updated?

The model will be updated after every public poll and the latest predictions will be available within four hours of polls being released.

Predictions for the electorate battles will be added once final candidate lists are available.

What are the limitations of the model?

It is hard to predict party vote for a newcomer like The Opportunities Party. Because it's new there is not enough polling data for the model. However, raw polling data is included in our visualisation instead.

Another limitation, according to Professor Lumley, is the lack of demographic information accompanying publicly available polling data.

"Knowing about party preferences by, say, gender, age, ethnicity and income lets you use Census data about each electorate to get even more accurate results, and with less concern about how representative the sample is."