An Influence Method for Outliers Detection Applied to Time Series Traffic Data

The applicability of an outlier detection statistic developed for standard time series is assessed in estimating missing values and detecting outliers in traffic count data. The work of Chernick, Downing and Pike (1982) is extended to form a quantitive outlier detection statistic for use with time series data. The statistic is formed from the squared elements of the Influence Function Matrix, where each element of the matrix gives the influence on pk, of a pair of observations at time lag k. Approximate first four moments for the statistic are derived and by fitting Johnson curves to those theoretical moments, critical points are also produced. The statistic is also used to form the basis of an adjustment procedure to treat outliers or estimate missing values in the time series. Chernick et al's (1982) nuclear power data and the Department of Transport's traffic count data are used for practical illustration.