Wavelet based Segmentation in Detecting Multiple Mean Changes in Time Series

Aims/ Objectives: Multiple mean break detection problem in time series is considered. A segmentation based on detecting turning points is applied to the original time series and its scaling coefficients series resulting from the maximal overlapped discrete wavelet transform (MODWT). Using a segmentation level along with a minimal distance parameter between two successive turning points we select a small number of segments within each series. A change point statistical test is then run separately within each series and over each segment. The simulation experiment shows that the multiple mean break detection procedure offers very good practical performance. The test procedure is applied to a real set of data.