A heuristic search procedure for detecting sudden shifts in stationary time series data

A new heuristic search procedure is proposed for retrospectively detecting shifts (defined as sudden changes in the process mean) within a stationary time series subject to substantial white noise. After identifyng the first, most significant shift, the search procedure is applied progressively to detect further shifts and also to define the timing, size and statistical significance of such shifts. Prior to the application of the procedure, the time series under review is evaluated to determine whether it is consistent with the shifting-mean model that underlies the heuristic. A feature of the search procedure is that it can be operated automatically, with searches terminated either when the segment of the data series within which the next identified shift occurs is shown not to be suitable for the application of the heuristic, or when the latest identified shift proves not to be statistically significant.