Thesis
Bayesian inference in high-dimensional state space models with time-varying parameters and stochastic volatility
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2025
- Thesis identifier
- T17480
- Person Identifier (Local)
- 201678218
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- This thesis begins with an introductory chapter that outlines the inferential challenges of high-dimensional time series models with parameter changes, multiple dependent variables, numerous predictors, and substantial computational complexity, and motivates the methodological contributions that follow. The first objective is to evaluate a conventional Bayesian estimation approach in modelling the real effects of credit market disruption through a measure of financial distress, the financial external premium, which may vary over time. For this purpose, we specify an eight-variable structural vector autoregressive model with time-varying parameters and stochastic volatility (TVP-VAR-SV). We use the model to examine the nature and evolving features of the links between macroeconomic and financial variables in the U.S. economy. The second objective is to develop a Bayesian pairwise composite likelihood method to address a high-dimensional inference problem in time series models with parameter changes, many dependent variables, and computational complexity, in order to conduct structural analysis. While a larger macroeconomic and financial dataset could be analyzed using a suitable TVP-VAR-SV model, the computational burden of such a model becomes prohibitive in high dimensions. To address this, the method replaces the full likelihood function with a product of pairwise marginal likelihoods and then combines the results to make inference from the composite model. To efficiently aggregate information across bivariate models, we introduce the Direct Averaging Method, a novel approach that provides a computationally tractable approximation to the multivariate structure without requiring simulation from the joint model. An empirical study of time-varying pairwise composite impulse responses demonstrates the impact of an unexpected financial shock on fifty quarterly U.S. macroeconomic and financial variables, yielding economically meaningful results. The third objective is to develop a Bayesian dynamic graphical model approach to overcome a high-dimensional sparse inference problem in TVP-VAR models with volatility discounting, many predictors, and computational complexity, in order to conduct forecasting analysis. The approach incorporates pairwise conditional independence structures in both the coefficient states and the off-diagonal elements of the covariance states. The Bayesian dynamic graphical framework improves forecast combinations of multiple quarterly U.S. macroeconomic and financial variables.
- Advisor / supervisor
- Darby, Julia
- Koop, Gary
- Resource Type
- DOI
关系
项目
| 缩略图 | 标题 | 上传日期 | 公开度 | 行动 |
|---|---|---|---|---|
|
|
PDF of thesis T17480 | 2025-11-27 | 公开 | 下载 |