An L1 smoother for outlier cleaning of time series

摘要

This paper introduces a new robust outlier cleaner specific for high-frequency time series data and provides guidelines for researchers whowish to use this procedure before the analysis process starts. The essenceof the method is a fully automatic, data-driven procedure based on fit-ting, by least absolute deviations, a reference function to the actual timeseries. Once the reference curve has been defined, it can be used to es-tablish bands such that all observations that deviate from the referencecurve by more than a prefixed amount will be replaced. Properties ofthe new screening tool are investigated through the accuracy of simul-taneous prediction intervals produced by Box-Jenkins models appliedto real data, before and after the outlier cleaner usage. It is shown thatthe new method can be validly used as a data preparation technique toensure that statistical analysis is supported by clear-cut data.