Please use this identifier to cite or link to this item: https://scholarbank.nus.edu.sg/handle/10635/226234
Title: HIGH-DIMENSIONAL NONSTATIONARY TIME SERIES MODELING WITH FUNCTIONAL DATA ANALYSIS AND DEEP LEARNING
Authors: XU XIUQIN
Keywords: High-dimensional, nonstationary, time series, functional data analysis, deep learning
Issue Date: 5-Jan-2022
Citation: XU XIUQIN (2022-01-05). HIGH-DIMENSIONAL NONSTATIONARY TIME SERIES MODELING WITH FUNCTIONAL DATA ANALYSIS AND DEEP LEARNING. ScholarBank@NUS Repository.
Abstract: This thesis contributes to the high-dimensional and nonstationary time series modeling based on functional data analysis and deep learning. It is challenging to develop accurate probabilistic/point forecasting methods for time series due to their complex characteristics, such as high-dimensionality, cross-dependence, non-linearity, and nonstationarity, especially when the sample size is small. In this thesis, we developed three models to deal with these challenges. First, to provide probabilistic forecasting for functional data (high dimensional data), we developed predictive predictors for curves based on a curve-to-curve regression framework. Second, we proposed a hybrid functional autoregressive and convolutional neural network method to efficiently model high-dimensional time series with both linear and non-linear patterns. Third, we proposed a deep switching state-space model for time series with nonstationary regime-switching patterns. We extend the linear switching statespace model by using deep learning to parameterize its transition and emission functions and developed an efficient estimation method.
URI: https://scholarbank.nus.edu.sg/handle/10635/226234
Appears in Collections:Ph.D Theses (Restricted)

Show full item record
Files in This Item:
File Description SizeFormatAccess SettingsVersion 
XuXQ.pdf6.3 MBAdobe PDF

RESTRICTED

NoneLog In

Page view(s)

38
checked on Sep 29, 2022

Download(s)

5
checked on Sep 29, 2022

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.