WIFO's New Seasonal Adjustment Procedure

Economic time series are submitted to seasonal adjustment procedures in order to facilitate the analysis of the data. Originally, deterministic approaches were adopted for the non-observable components like trend, seasonal factors, etc. Later on, they were replaced by the calculation of moving averages. Best known are "Census X–11" for seasonal and "HP filter" for trend adjustment. The advantage of such ad-hoc filters lies in their easy applicability, their disadvantage is that "they take a look at the world from a particular perspective", with possibly far-reaching negative consequences. Also, the lack of a statistical model detracts from a meaningful application of ad-hoc filters. In order to eliminate the problems attached to ad-hoc filters, new ways of estimating the non-observable components have been developed over the last 15 years. With the ARIMA model-based approaches, one first identifies a model of the observable time series and then derives from it models for the particular non-observable components (trend, season, irregular component) which are compatible with the global model. For the selection of an optimal seasonal adjustment method, both theoretical and empirical criteria should be applied. In a comprehensive study, Eurostat has examined five different seasonal adjustment procedures using more than 80 time series. TRAMO/SEATS – an ARIMA model based approach – proved to be the most sophisticated theoretical method and also performed best in most empirical tests. The TRAMO/SEATS program set allows the modeling of a time series via the combined estimation of ARIMA approaches and outlier as well as special effects, while also protecting against over- or underdifferentiation. It also allows a comparison between the components derived from the theoretical model and the empirical estimators, thereby facilitating the diagnosis. Furthermore, it offers a number of analytical tools which form the basis for statistically robust conclusions. Application of TRAMO/SEATS to time series of retail sales and consumption expenditure shows that it produces appropriate models for the explanation of these data. The outliers identified in these time series are readily explained by the expected effects of discretionary fiscal measures and changes in the statistical base. The theoretically obtained components correspond to a high degree to the empirically estimated ones. The series seasonally adjusted with TRAMO/SEATS exhibit markedly weaker variations in growth rates than those adjusted with X11, which makes interpretation of the data easier. The empirical record so far as well as test results of Eurostat suggest that this approach should be adopted for seasonal adjustment also in Austria.