Temporal Aggregation and Seasonal Adjustment
Due to user demand, the Australian Bureau of Statistics (ABS), in some instances, publishes original, seasonally adjusted and trend estimates at different observation frequencies for the same indicator. Hence sometimes at the quarterly level, original time series estimates are simply a temporal aggregate of their monthly counterpart. Suppose a time series of quarterly seasonally adjusted estimates is desired from such an equivalent time series pair. These estimates can be obtained via two approaches. Either by (1) seasonally adjusting the quarterly original time series directly, or by (2) seasonally adjusting the monthly original time series and then temporally aggregating to the quarterly level (referred to as the temporal aggregation approach hereafter). The ABS currently uses method 1 for seasonally adjusting equivalent time series pairs. This leads to quality and consistency issues along with duplicate work.
The ABS knows from previous research that estimating calendar-related effects (e.g. trading day effects) is more accurate when performed at the monthly level and then applied to the quarterly case. This idea of using a monthly time series to estimate a component of its quarterly equivalent is taken further by the temporal aggregation approach. The aim becomes to completely derive the quarterly seasonally adjusted series from its monthly seasonally adjusted counterpart. Hence significant improvements in quality and consistency are expected to be made.
For the Census X11 method, the literature suggests that seasonal adjustment first and temporal aggregation second is the more efficient approach in terms of mean squared error and forecast performance. However, the impact of temporal aggregation on current end revisions has not been assessed for the mixed X11/ARIMA forecasting method utilised by the ABS.
The ABS is about to undertake case studies to compare the quarterly seasonally adjusted estimates obtained via the two approaches using the ABS X11/ARIMA forecasting method, in terms of their relative efficiency, revisability and consistency. The aim of this work is to confirm that the temporal aggregation approach is more efficient, results in no worsening of current end revisions and improves consistency. The background and methodological basis of these studies will be presented in an ABS Methodological Advisory Committee paper in June 2008.
For more information on this project, please contact Lisa Apted on (03) 6222 5932 or Mark Zhang on (02) 6252 5132.