On Aspects of Changepoint Analysis Motivated by Industrial Applications

Grundy, Tom (2022) On Aspects of Changepoint Analysis Motivated by Industrial Applications. PhD thesis, UNSPECIFIED.

[img]
Text (2022GrundyPhD)
2022grundyphd.pdf - Published Version
Restricted to Repository staff only until 14 March 2023.
Available under License None.

Download (2MB)

Abstract

In numerous industrial applications, organizations wish to monitor time series data to better understand the past and make predictions for the future. To achieve this, appropriate analysis and modelling of time series data needs to be performed. However, it is common for time series to undergo abrupt structural changes, known as changepoints, which can cause major challenges during model fitting. Identifying changepoints is crucial to appropriately understand, analyze and model time series data in a wide range of applications including finance, genomics and countless others. In this thesis, we introduce new methodologies and frameworks for detecting changepoints in two scenarios; when there are multiple series to analyze simultaneously (multivariate setting) and when we receive time points sequentially and aim to detect changes as soon as possible (sequential setting). These methodologies provide novel and innovative ways to detect different types of changepoints in a wide range of data structures. Firstly, we introduce a novel bivariate test statistic for detecting changes in mean and variance simultaneously in multivariate data. Our attention then turns to changes in covariance, specifically, we introduce a cost function and changepoint framework for identifying changes in subspace. Finally, we switch to the sequential changepoint setting and provide a framework for quickly identifying mean and variance changes in more complex data structures using forecast models. For all the methods we demonstrate their strong empirical performance in both simulated examples and industrial applications.

Item Type:
Thesis (PhD)
ID Code:
167468
Deposited By:
Deposited On:
15 Mar 2022 09:30
Refereed?:
No
Published?:
Unpublished
Last Modified:
03 May 2022 00:18