This paper deals with a key source of error when it comes to forecasting — that we often do not know where we are, or where we have been, when we are trying to gauge the outlook. The above chart shows that the one sigma confidence interval is around 100bps when a quarterly GDP print is released (so an initial 0.5%qoq result may turn out to be a -0.5%qoq result once all the revisions are in).
The above table shows that the errors are large for each component, and that the revisions go on for some time. The errors are smaller for the aggregate, as they cancel out when summed.
If pressed, it looks like GDP(P) is consistently the best single measure of output — particularly on a quarterly basis.
So what gets revised? Basically everything. Though the more volatile components tend to see the largest revisions — as you can see from the chart below.
Weighting these by their economic importance, you can see the problem with the GDP(E) and GDP(I) measures – relative to GDP(P). Both GDP(E) and GDP(I) have components that tend to be very heavily revised (inventories and profits respectively).
This is not just a problem for Australia. Similar nations also tend to have persistent problems with the estimation of GDP.
So what to do? The answer is to take averages of a large amount of data (so called factor models). At least in the US, research has shown that factor models defeat the real time / revision problem (see Bernanke and Bovin JME 2003)