Thursday, November 12, 2009

September Industrial Production: Down But Not Out

September's IPI report shows production falling in most sectors, even on a seasonally adjusted basis - no biggie, since we're still looking at a Hari Raya effect here (log monthly changes):

I'll only admit to being worried if the numbers don't turn up strongly next month.

Here's an interesting question to ponder while we wait: can we use the IPI to provide an advance forecast for current quarter GDP? The Q3 numbers will be coming out very soon, so this is an exercise with a quick payoff, and since we're interested in a forecast rather than in the structural relationships there's no need to do a full scale model. The following specification works pretty well:

log(rgdp) = constant + D2 + D3 + D4 + log(ipi) + AR(1)

The IPI figure used here is the quarterly average, the D* terms are seasonal dummy variables, while the AR(1) term models the sample error as a 1 period autoregressive scheme. Sample period is 2001:1 to 2008:4:

The R2 is a suspiciously high 99.7%, which really warrants some cointegration testing, especially with a DW stat close to 2. But since I'm not interested in structure I'm ignoring those for now, especially since all the diagnostics check out ok. The full sample forecast against realization shows a fairly close relationship:

...but note that the estimated forecast is overstating GDP over the last three quarters. There lies the weakness of most forecast models when you're trying to foresee trouble: because of the techniques used, models like this will show a rosier picture than the reality during a downturn, and be more pessimistic when the economy is in a bubble. In any case, I'm only interested in the 3Q2009 forecast, so using a 1-step forecast yields:

Point Forecast:133572.56, Interval forecast: 135834.37-131310.76

Which is about -1.7% growth for the point forecast and 0.0% to -3.3% growth for the interval forecast (95% confidence level), using a year-on-year comparison. For quarter-on-quarter annualised, the equivalent numbers are 22.7% (!!!!), and 31.3% to 14.6%.

Technical Notes:
1. All data from DOS

No comments:

Post a Comment