Frequency-Domain Methods for Change-Point Detection

• Activity Number: 175
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Physical and Engineering Sciences
• Author(s): Karim Rahim*+ and David J. Thomson
• Companies: Queen's University and Queen's University
• Keywords: Change-point Detection ; Spectral Estimation ; Time Series ; Physical Sciences ; Climate Science ; Multitaper

Abstract

Change-point detection generally occurs in the time domain and focuses on change in statistical distribution parameter such as a change in mean, variance, shape, or change in rank statistics. We introduce a frequency-domain change-point detection method, and propose its use in climate time series. This method makes use of multitaper spectral estimation. In this talk we: (1) introduce the estimation method, (2) study it with simulation, and (3) present an example application using climate time series.

Our method proposes using ave(log(S1(f)) - log(S2(f))) where S1(f) and S2(f) are spectra at frequency f taken at different time sections. We explore both averaging over the frequency range 0 < f < 1/2, when the time spacing is one, and over a localized frequency range. We demonstrate using simulations that when the spectra S1(f), and S2(f) are estimated using the lower bias, lower variance, higher computational cost multitaper spectral estimator, the proposed method is successful in change-point detection in the frequency-domain, even given a homoscedastic times series with no noticeable change in mean over time.