A1271
Title: High dimensional changing region detection through distributional test
Authors: Ting Tin Ma - The Chinese University of Hong Kong (Hong Kong) [presenting]
Kin Wai Chan - The Chinese University of Hong Kong (Hong Kong)
Abstract: Different components of high-dimensional time series may respond to structural changes at varying rates, leading to change points clustering within an interval rather than occurring precisely at isolated points. A test for change intervals, or more broadly, change regions, is studied. The discrepancy from the null hypothesis is measured by the asymptotic proportion of coordinates that experience a change point over an unspecified change region. The test is constructed by comparing the empirical distribution of estimated change points across coordinates to the null distribution. This form of testing is highly scalable for high-dimensional time series and is tuning-free. The proposed framework covers the traditional change point test as a special case and avoids long-run variance estimation and self-normalization employed in mainstream testing approaches. The test statistic is proven to be pivotal asymptotically, and Monte Carlo simulations confirm good size control and high power even in finite samples.
