Outlier Detection by Forecasting

The Consumer Expenditure Quarterly Interview Survey collects data from consumer units (CUs) about their expenses during the previous 3 months. The purpose of the survey is to gather information about large purchases, such as those of vehicles and appliances, and expenditures that are made on a regular basis, such as rent and utility payments. These data are collected by the U.S. Census Bureau and then transferred to the U.S. Bureau of Labor Statistics (BLS), Division of Consumer Expenditure Surveys (CE). The branch of Production and Control (P&C) screens and processes the raw data for their eventual use in publications and in the weighting of the BLS Consumer Price Index. P&C’s final data-editing procedure for the Interview Survey is the Monthly Tabulation of Expenditures (MTAB), which maps or assigns expenditures to a specific month and a Universal Classification Code (UCC).1 The MTAB Review procedure then evaluates the created data for suspicious values. To improve the existing review procedure, P&C initiated a research project in August 2005. The goals of this project were to make the MTAB Review more efficient, focus analysts’ attention on outliers, create more informative reports, and provide more accurate data to end users. Three techniques for improving the process of selecting outliers were investigated during the modernization of the MTAB Review. The method that was chosen, which compared forecasted with reported values, was implemented in February 2006. With this technique, the analyst detects outliers by using forecasted prediction intervals created by SAS and comparing them with current means.2 This article summarizes the forecasting technique adopted for the MTAB Review.