Outlier detection with automatic modelling: TRAMO/SEATS versus X-12-ARIMA

Abstract

This paper presents a comparative study of the automatic outlier detection and model selection procedures of the most widely used ARIMA modelling and seasonal adjustment programs: TRAMO/SEATS (TSW) and X12-ARIMA, which are the most important software alternatives for automatically modelling economic time series. Recently X12-ARIMA's automatic modelling procedure has become more similar to TRAMO, but differences in outlier treatment remain. To assess the performance of both programs when modelling and detecting outliers, a comparison and a straightforward simulation exercise are performed to evaluate their default options. Judging by the outcome, and taking into account the great number of series (even hundreds) that frequently are handled in each time series analysis, TRAMO/SEATS should be preferred in automatic default use. Although the current X12-ARIMA main default automatic modelling procedure (automdl) records slightly better levels of success than TSW in selecting the proper model (among the ones tested), it maintains extremely high levels of non-convergence. An issue that entails a serious problem in (semi-)automatic applications. These findings will be of interest to statistical offices, central banks and corporate and academic users, and, more generally, to statisticians and economists involved with univariate modelling of seasonal monthly or quarterly time series.