Detecting Outliers in Time Series Data

Abstract

Occasional large errors in data can have drastic effects on estimates for such quantities as correlation coefficients. regression coefficients, and spectral density estimates. In this article we investigate the effect of outliers on time series data by considering the influence function for the autocorrelations p ( k ) of a stationary time series. This influence function matrix is applied to simulated data, to power plant data, and to inventory data on nuclear materials.

A New Nuclear Material Safeguards Method Using Fractal Dimension

“…Chernick et al [11] proposed one of the earliest of the material balance time series outlier detection methods. In their paper, they gave seven nuclear loss data sets.…”

Assessing influence in time series

“…Potential influential points are then examined through a diagnostic plot. Specific measures of influence for the autocorrelation function are given in Chernick, Downing and Pike (1982), Lattin (1983), Li and Hui (1987), and Lefrancois (1991). Abraham and Chuang (1989), and Peiia (1990) study influential observations for autoregressive coefficients in AR models.…”

Maximum studentized score tests for the detection of outliers in time series regression models

“…For example Chernick et al [10], Lattin [11], Li and Hui [12], and Lefrancois [13] have considered certain diagnostic techniques for the detection of outliers that are the result of the shifted autocorrelations. But, these tests do not appear to accommodate the detection of a variance shifted outlier which may or may not cause any shift in the autocorrelation structure.…”