RDP 2007-03: Forecasting with Factors: The Accuracy of Timeliness 5. Conclusion

This paper shows that factor-based forecasts can outperform standard time-series benchmarks for key Australian macroeconomic series, as has been found for many other countries. This is perhaps not surprising. Using a relatively large number of series can produce a less noisy estimate of the current state of the economy and also makes the forecasts less susceptible to structural change in any given explanatory variable.

We find almost uniformly that simple models which use a fixed number of factors outperform more complex models that select a different number of factors at each forecast iteration. Further, for most series, the simplest model we present (which excludes lags of the series being forecast) tends to outperform those that include lags. There are two important exceptions to this: forecasts of CPI inflation and the unemployment rate need to include autoregressive lags to account for structural change and the long cycles in these series over our sample period of 45 years.

The use of a broad data panel to estimate the factors may enhance forecast accuracy but at the cost of including series with late publication dates, so resulting in less timely forecasts. We conduct an out-of-sample forecasting exercise that iteratively uses less timely data panels, which contain more information to estimate the factors, in order to assess the extent of this possible trade-off. With the exception of CPI inflation, the forecasts do not become more accurate when they utilise a broader but less timely selection of series. While this is an important result, it is probably not surprising as the factors, which capture economic cycles, are highly persistent. Consequently, the factors derived from adjacent quarters' data tend to be very similar and so are the forecasts. So while factor forecasts have large data requirements, we show that these should not prevent their practical use in a policy setting in which timely forecasts are needed.