Title: Change point estimation for high-dimensional time series
Authors: Sayar Karmakar - University of Florida (United States) [presenting]
Abstract: In this talk, we discuss two projects where synchronization of change-point is discussed for multiple time series. We primarily focus on multiple series observed simultaneously who are all susceptible to have one significant change-point. If we can accept that all series change simultaneously after a proper statistical testing procedure, one can then look for reasons that stimulates this change. On the other hand if synchronization fails one can cluster these series based on the changepoint location. We show multiple applications as motivation.In the first project we focus on low-dimensional case where number of observed series does not grow. Here we use a moving sum based detection technique and derive the threshold distribution of our statistic through a Gaussian distribution. In the second one (ongoing) we focus on high-dimensional scenario where we do not have a similar Gaussian approximation. Instead we establish some results for high dimensional quadratic forms that allow us to test the synchronization in high dimension. We provide some simulation studies for both the scenario.