Novel Methods for Efficient Changepoint Detection

Romano, Gaetano and Fearnhead, Paul and Eckley, Idris (2021) Novel Methods for Efficient Changepoint Detection. PhD thesis, Lancaster University.

Text (2021romanophd)
2021romanophd.pdf - Published Version
Available under License Creative Commons Attribution.
Download (3MB)

Abstract

This thesis introduces several novel computationally efficient methods for offline and online changepoint detection. The first part of the thesis considers the challenge of detecting abrupt changes in scenarios where there is some autocorrelated noise or where the mean fluctuates locally between the changes. In such situations, existing implementations can lead to substantial overestimation of the number of changes. In response to this challenge, we introduce DeCAFS, an efficient dynamic programming algorithm to deal with such scenarios. DeCAFS models local fluctuations as a random walk process and autocorrelated noise as an AR(1) process. Through theory and empirical studies we demonstrate that this approach has greater power at detecting abrupt changes than existing approaches. The second part of the thesis considers a practical, computational challenge that can arise with online changepoint detection within the real-time domain. We introduce a new procedure, called FOCuS, a fast online changepoint detection algorithm based on the simple Page-CUSUM sequential likelihood ratio test. FOCuS enables the online changepoint detection problem to be solved sequentially in time, through an efficient dynamic programming recursion. In particular, we establish that FOCuS outperforms current state-of-the-art algorithms both in terms of efficiency and statistical power, and can be readily extended to more general scenarios. The final part of the thesis extends ideas from the nonparametric changepoint detection literature to the online setting. Specifically, a novel algorithm, NUNC, is introduced to perform an online detection for changes in the distribution of real-time data. We explore the properties of two variants of this algorithm using both simulated and real data examples.