On factor models for high-dimensional time series

Sabharwal, Ragvir Singh (2023) On factor models for high-dimensional time series. PhD thesis, London School of Economics and Political Science.

Text - Submitted Version Download (1MB)

Abstract

The aim of this thesis is to develop statistical methods for use with factor models for high-dimensional time series. We consider three broad areas: estimation, changepoint detection, and determination of the number of factors. In Chapter 1, we sketch the backdrop for our thesis and review key aspects of the literature. In Chapter 2, we develop a method to estimate the factors and parameters in an approximate dynamic factor model. Specifically, we present a spectral expectation-maximisation (or \spectral EM") algorithm, whereby we derive the E and M step equations in the frequency domain. Our E step relies on the Wiener-Kolmogorov smoother, the frequency domain counterpart of the Kalman smoother, and our M step is based on maximisation of the Whittle Likelihood with respect to the parameters of the model. We initialise our procedure using dynamic principal components analysis (or \dynamic PCA"), and by leveraging results on lag-window estimators of spectral density by Wu and Zaffaroni (2018), we establish consistency-with-rates of our spectral EM estimator of the parameters and factors as both the dimension (N) and the sample size (T) go to infinity. We find rates commensurate with the literature. Finally, we conduct a simulation study to numerically validate our theoretical results. In Chapter 3, we develop a sequential procedure to detect changepoints in an approximate static factor model. Specifically, we define a ratio of eigenvalues of the covariance matrix of N observed variables. We compute this ratio each period using a rolling window of size m over time, and declare a changepoint when its value breaches an alarm threshold. We investigate the asymptotic behaviour (as N;m ! 1) of our ratio, and prove that, for specific eigenvalues, the ratio will spike upwards when a changepoint is encountered but not otherwise. We use a block-bootstrap to obtain alarm thresholds. We present simulation results and an empirical application based on Financial Times Stock Exchange 100 Index (or \FTSE 100") data. In Chapter 4, we conduct an exploratory analysis which aims to extend the randomised sequential procedure of Trapani (2018) into the frequency domain. Specifically, we aim to estimate the number of dynamically loaded factors by applying the test of Trapani (2018) to eigenvalues of the estimated spectral density matrix (as opposed to the covariance matrix) of the data.

Item Type:	Thesis (PhD)
Additional Information:	© 2023 Ragvir Singh Sabharwal
Library of Congress subject classification:	Q Science > QA Mathematics
Sets:	Departments > Statistics
Supervisor:	Moustaki, Irini
URI:	http://etheses.lse.ac.uk/id/eprint/4503

Actions (login required)

Record administration - authorised staff only