A comparison of automatic outlier detection methods for

A comparison of automatic outlier detection methods for time series Tariq Aziz Office for National Statistics

Research motivation • Outlier identification can have a significant impact on seasonal adjustment and forecasting • ONS uses X-13 ARIMA-SEATS for automatic detection of additive outliers and level shifts • Indicator Saturation offers another approach to outlier detection • The aim of the study is to compare these two approaches 2

X-13 ARIMA-SEATS • Uses reg. ARIMA approach for modelling the time series • Specific-to-general approach based on sequential addition and followed by backward deletion • Fits AO and LS regressors at all points of a time series • Uses large critical t-values based on the length of the series. 3

Indicator Saturation • A general-to-specific approach proposed by Hendry (1999) • Used reg. ARIMA approach for modelling the time series • Impulse indicator saturation(IIS) for automatic detection of additive outliers (AOs) • Step indicator saturation(SIS) for automatic detection of level shifts (LSs) • Uses critical t-values based on the length of the series. 4

Conclusions • IS method detects higher proportion of true outliers. • X-13 ARIMA-SEATS method detect lower proportion of incorrect outliers. • Forecast performance is slightly better for X 13 ARIMA-SEATS when detecting AO and IS performs better when detecting level shifts • Revisions analysis showed that IS method produces consistently higher revisions 5