A1110
Title: Controlling FDR of change points in structural break time series
Authors: Wei Zhia Kua - The Chinese University of Hong Kong (Hong Kong) [presenting]
Chun Yip Yau - Chinese University of Hong Kong (Hong Kong)
Abstract: ScoreFDR is proposed, a procedure using global segmentation in a multiscale way with score statistics to find multiple change points in an autoregressive (AR) process. In contrast to traditional methods that are focused on achieving consistency on estimated changepoints, our approach is specifically designed for false discovery rate (FDR) control, i.e., controlling the expected proportion of falsely detected change points against all identified change points. The key to controlling the FDR involves constructing a multiscale quantile constraint that enables the control of local errors within individual segments. We proved that ScoreFDR is able to control the FDR asymptotically. The performance of FDR control in finite samples is illustrated via extensive simulation studies. Applications to real data examples are also illustrated.
