Title: An evaluation of automatic outlier detection methods
Authors: Jennifer Davies - Office for National Statistics (United Kingdom) [presenting]
Abstract: The Office for National Statistics (ONS) is the UKs largest independent producer of official statistical and the recognised national statistical institute of the UK. The Time Series Analysis Branch (TSAB) within the ONS is responsible for the seasonal adjustment and analysis of tens of thousands of time series measuring economic and social phenomena. A major challenge faced when seasonally adjusting time series is identifying and accounting for outliers; observations within a time series that differ distinctly from the general pattern of the trend and/or seasonal components. An outlier, if unidentified or misspecified, may cause issues in the analysis of a time series and lead to large revisions. Revisions are an inevitable part of the production process due to late returns and the process of seasonal adjustment. Analysts aim to minimise revisions to ensure reliability and retain public faith in published estimates, hence the importance in correctly identifying and specifying outliers. Outliers at the end of a time series can be particularly difficult to identify in a timely manner, due to the uncertainty around future observations.Techniques to account for outliers at the end of a series will be evaluated. It will consider the currently used automatic outlier detection method in X13-ARIMA-SEATS and compare this with other outlier detection methods found in the literature. The main alternative methods considered are indicator saturation and changepoint methods.