Degree Type

Dissertation

Date of Award

2020

Degree Name

Doctor of Philosophy

Department

Statistics

Major

Statistics

First Advisor

Zhengyuan Zhu

Abstract

This dissertation is focused on time series analysis, particularly the high-dimensional time series and its application in traffic network data and economic forecasting. In the first project, we propose a weighted l1 regularized vector auto-regressive (VAR) model for spatio-temporal data. We compared the proposed method with commonly used LASSO estimation by extensive simulation studies and an application on a traffic network dataset. Both the simulation and real data application demonstrate the advantages of our method. In addition, we explored the theoretical properties of the proposed method under both exactly sparse and weakly sparse scenarios. In my second project, we provide a thorough analysis of the dynamic structures and forecasting of the Chinese consumer price index (CPI), with a comparison to those of the US CPI. We studied the similarities and differences between the Chinese CPI and US CPI in terms of their decompositions, dynamic structures and predictabilities. We found that both series can be well modeled by a class of seasonal auto-regressive integrated moving average model with covariates (S-ARIMAX), which precisely captures the dynamic structures of these two series, including trend, seasonality, outliers and lunar holiday effects. In addition, the Chinese CPI series possesses stable annual cycles and strong Spring Festival effects, with fitting and forecasting errors largely comparable to their US counterparts. Finally, for Chinese CPI, the factor augment (FA) approach can improve prediction over S-ARIMAX models for short-term forecasting. In the third project, we survey several cutting-edge variable selection methods and apply them to the field of economic forecasting with many predictors. Four groups of variable selection methods are reviewed and applied in real data to forecast three important economic indices. Their performances are compared with the popular FA approach. We also conduct several simulation studies to compare the four groups of variable selection methods in terms of their variable selection accuracy and out-of-sample prediction.

DOI

https://doi.org/10.31274/etd-20200902-161

Copyright Owner

Zhenzhong Wang

Language

en

File Format

application/pdf

File Size

139 pages

Available for download on Sunday, August 28, 2022

Share

COinS