Two days ago, we released our forecast for yesterday's election in New Zealand. Its electoral system is pretty similar to the one in Germany, and we were curious to see how our model performs in another multi-party system with proportional representation. As all votes in New Zealand are counted, it is time for an assessment.

Our forecast of the New Zealand election.

For New Zealand, we predicted the vote shares of the National Party, Labour Party, New Zealand First Party, Green Party. All other parties were grouped under "Others". As no forecast is certain, we predicted the vote share for each party plus a range (indicated by the bars) where we expect the actual value to be with an 83% probability - the same 5-in-6 chance as rolling anything other than a six on a die.

The following graph shows the actual election results and our forecasts.

For all parties except for the Green Party, the actual election result is within the range of the bars. This means we correctly forecasted the results of three out of four parties (as well as the other parties) correctly. Why did our model fail at getting the Greens right? A preliminary look at the data shows that the fundamental component of our model was overpredicting the Greens. The polls, on average, also overestimated the Greens. In addition, the Greens had started to tank in the polls only a couple of weeks before the election (before July, they had consistently polled over 10%), which was difficult to pick up for our model, which searches for a reasonable compromise between the fundamentals component and the increasingly precise polling component. More generally speaking, our fundamental model worked quite well and is fairly generic. In fact, we did not put any effort in adapting it from the German context.

Besides the raw election result for each party, we also forecasted a couple of interesting scenarios. Here, we were right that the National Party becomes the strongest party, and that both, the Greens and New Zealand First, clear the five percent hurdle.

A comparison with the last opinion polls before the election.

Did our model better than the polls? For this, we compare how much we were off on average in comparison with the last four opinion polls. On average, across all parties, we were only off by 1.16 percentage points. While the average of the last four polls was off 1.75 percentage points on average across all parties. We were better in predicting the vote share of the Labour party than any of the latest opinion polls.

Party	Result	Prediction Zweitstimme.org, 22 Sep. 2017	Newshub Reid Research, 12 Sep, 2017	One News Colmar Brunton, 14 Sep. 2017	Newshub Reid Research, 20 Sep. 2017	One News Colmar Brunton, 21 Sep. 2017
NAT	46%	44.7%	47.3%	40%	45.8%	46%
LAB	35.8%	36.2%	37.8%	44%	37.3%	37%
NZF	7.5%	6.4%	6.0	6%	7.1%	5%
GP	5.9%	8.4%	4.9%	7%	7.1%	8%
OTHERS	4.8%	4.3%	4%	3%	2.7%	4%

We are excited to see how our model performs tonight in Germany.