An online Change-Point Detection Procedure in Time Series

DOI: 10.5176/2251-1938_ORS69

Authors: Yong (Jimmy) Jin and Chun-yip Yau

Abstract:

This paper proposes a new and fast sequential change point detection procedure for a change point in a time series. We develop a new proxy for the change in the covariance structure, the Time Average Variance Constant (TAVC), which could be expressed as the summation of all the auto-covariance functions. Wu (2009)' s algorithm is implemented to compute the TAVC recursively with an efficient updating computational and memory complexity O(1). In this article, we construct a confidence band to monitor changes in a time series based on the Strong Invariance Principle. Good size and power properties are demonstrated on Monte Carlo Simulation. Our algorithm only takes ~2' s on a single 1.86GHz processor with a sequence of length 10000.