Automatic outlier handling and model selection in seasonal adjustment

The first part of this paper presents the most important results from a history analysis of 52 Norwegian economic time series (Langsrud, 2011). It is illustrated how revisions is affected by two automatic ARIMA model selection methods (automdl and pickmdl). Furthermore, it is shown that straightforward re-identification of outliers (the concurrent method) leads to big revisions. From this knowledge the second part of the paper considers the problem of automatically dealing with outliers (Langsrud, 2012). How should potential outliers be handled before the final decision is made? Three algorithms are suggested which can be named as “jump in and out”, “jump in” and “jump out”. It is demonstrated how revisions and out-of-sample forecasts (quality of model) are affected by using the algorithms. The results are compared to the concurrent method. The results indicate, however, that the best improvements are obtained by increasing the outlier detection limit. The analyses were made by running X-12-ARIMA via the R programming language.