Laimonis Kavalieris

University of Otago

Counting breakpoints in time series

Structural breaks in time series include abrupt changes in mean, variance or autocorrelation either singly or in any combination. They have attracted a great deal of attention in the econometrics literature, for example time series with level shifts are used to model apparent long memory. Other recent work uses breakpoint models as approximations for non-stationary time series. We are concerned with the estimation of the number of structural breaks.

In this talk we discuss the Minimum Description Length (MDL) principle as a criterion for selecting the number of breaks in a level shift model in a positively correlated time series. Some new theoretical results will be mentioned together with links to smoothing time series with a drifting mean. The procedure will be illustrated using examples from the natural sciences.