Premium
Outlier identifiability in time series
Author(s) -
Battaglia Francesco,
Cucina Domenico
Publication year - 2020
Publication title -
stat
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.61
H-Index - 18
ISSN - 2049-1573
DOI - 10.1002/sta4.281
Subject(s) - outlier , identifiability , anomaly detection , series (stratigraphy) , computer science , data mining , object (grammar) , measure (data warehouse) , time series , statistical power , pattern recognition (psychology) , statistics , artificial intelligence , econometrics , machine learning , mathematics , paleontology , biology
The occurrence of undetected outliers severely disrupts model building procedures and produces unreliable results. This topic has been widely addressed in the statistical literature. However, little attention has been paid to determine how large an outlier has to be for correct detection of both time and magnitude to safely take place. This issue has been the object of research mainly in geodesy. In this paper, the minimal detectable bias concept is extended to vector time series data, and the risk of accepting an outlier as a clean observation is evaluated according to both the size and power of the statistical tests. This approach seems able to deal with the difficult issues known as masking and swamping. The proposed measure of outlier identifiability helps to determine if any configurations of multiple outliers, also occurring in patches, are easily detectable.