A Jump in the Wayback Machine
Posted by Possum Comitatus on December 2, 2007
Way back in the annals of history, when the blog first started in May 2007, one of the first things I did was build a very rough election prediction model – and it’s probably worth going back to have a quick squiz at it because we’ll be dealing with it here. What was interesting about that model was that it was more explanatory rather than predictive and it looked at relationships between variables over the entire 11 year period of the Howard government.
As a result it turned up some pretty interesting results, the most interesting being that the standard variable mortgage rate over the first 6 years of the Howard government actually had a positive statistical relationship with the government’s primary vote, meaning that as the standard variable mortgage rate increased, so did the government’s primary vote. It wasn’t until the second half of the Howard government’s reign that the relationship reversed itself, whereby the relationship between the cash rate and the ALP vote became positive. Yet, over the entire period of the Howard government up to May, the relationship between the mortgage rate and the government primary vote was still positive, albeit getting weaker with every rate rise (because, as we know – for the previous couple of years the actual relationship between rates and votes had reversed itself)
Back in that very first election model, we forecast ahead using some assumptions about the variables involved. We assumed that the satisfaction ratings of the two leaders would remain the same (which they basically did) , and assumed that the interest rate to disposable income ratio would continue on its medium term growth path (which it basically did too).
I just went over the model again, but this time feeding in the changing values of the variables as they occurred. So as the satisfaction ratings changed each month, those new values were fed into the model, but I maintained the internally generated vote estimates as the lagged values.
As a result, if we compare the original May forecast, with the updating forecast and the actual Newspoll monthly average results over the period of May to the election we get firstly on the ALP primary vote:
And secondly on the two party preferred forecast:
Currently, the AEC has the ALP two party preferred election result on 52.9% and the ALP primary on 43.7%
What I find amazing here is that this very early explanatory model projected forward produced far more accurate prediction results than the last model I did just before the election. Both the updating model and the original model in May that used the assumptions about satisfaction ratings, both ended up being close to the actual election result with the May model predicting 53.7 and the updating model predicting 52.8. The Newspoll monthly average for November ended up at 53.5
On the primaries, the May model ended up with 45.6, the updating model with 44.5 and Newspoll at 46.2. The updating model ended up being only 0.1% out on TPP terms, and 0.8% out on primary vote terms. The May model with the assumptions was only out by 0.8% on TPP terms and 1.6% on primary vote terms.
I’m actually quite amazed that it worked so well, considering it forecast out 6 months into the future.