Detecting outlier samples in multivariate time series dataset
Knowledge-Based Systems
Hi-index | 0.03 |
The standard methods of time-series analysis, widely applied in many disciplines, tend to require long series, yet many situations give rise to several or many short time series. The existing literature on such data focuses on testing and estimating serial correlation. Another interesting aspect of this situation is whether any series is an outlier in relation to the others. In this paper, some models for sets of short time series are described and likelihood ratio tests are derived in order to detect if the mean level of a series is an outlier. Through a simulation study, the critical values of the tests are determined. The powers of the tests are computed for selected sample sizes and numbers of series. Their performance is compared to simple approximate tests based on standard outlier tests for the normal distribution applied to the set of means of the series.