Today was one of those perfect Autumn days in which I found myself pondering: I wonder what the Q1 2012 print for GDP will look like?
So I fished out my calculator to see what it would tell me. I feed it the data (in log format) and asked it to find the best seasonal, autoregressive, integrated, moving average model it could. The auto.arima() function is such a marvel. In seconds it said it was an ARIMA(1,1,2)(2,0,0) with drift. Before I knew it, I had the following graph in my hands.
If my ARIMA model is to be believed, the most likely print for Q1 2012 GDP growth is 1.1 per cent, quarter on quarter. The through the year result is a fabulous, faster than trend growth figure of 3.8 per cent. Though to be fair, the through the year figure benefits from Q1 2011 dropping off with its -0.3 per cent. Indeed, we need to do worse than -0.3 quarter on quarter for the through the year GDP result to decline from its Q4 2011 read of 2.3 per cent (an unlikely result according to the confidence intervals from the ARIMA model).
Do I believe it? Well ... it is as good a guess as any. It is remarkably close to the Government's pre-budget guess for 2012-13.
But ARIMA models are just a sophisticated way of saying the whether tomorrow will be much like the weather today. They're right most of the time, but they can miss the big turning points. They do a bit better during the storm, because they know the storm will eventually clear up and the weather will return to what it is like usually at this time of the year.
Notwithstanding their naivety, ARIMA models often do a reasonable job at predicting the next GDP print. It takes quite some hard work (and some devilish statistics) to beat them.
Nonetheless, I am up for the challenge of developing an unobserved components model with Kalman filtering. It is just going to take me a little longer that dropping the data into the auto.arima function.