This page shows the results of combining the simulations of the two different models I use for predicting the New Zealand election results:
- my “main” model, Model A, which is based on a generalized additive model with smoothing splines, fit after correcting for polling house effects, to forecast election day party vote - which allows recent trends in changing voting intention (eg the systematic growth or decline in support for a party) to continue towards election day
- a “state space” model, Model B, which simultaneously estimates house effects and latent voting intention for each party and assumes that voting intention is a random walk, so the level will not (on average) change in any structural way between now and election day although it will wander around at random.
See the page on Model B for more detail on the differences between the two approaches.
High level results - combined models
How we get there
Model A - splines that assume recent trends will continue
Model B - state space model with random walk latent voting intention