City Research Online

Abstract

Real-world time series data often change their dynamic behavior, such as means and variances, across different time periods. Sometimes changes may occur in the form of “jumps”, which are flash, or in the form of “breaks”, which are permanent. A more frequently observed form of these changes is “regime shifts”, which are both persistent and recurrent so that data seems to cycle between periods of behavior. Such changes are prevalent in the long-term trends of many macroeconomic and financial series, and are useful to glean a range of economic insights related to the cycling of the economy between recession and expansion periods (see, e.g., Hamilton, 1989; Ang and Bekaert, 2002), “bull” and “bear” markets in equity returns (see, e.g., Guidolin and Timmermann, 2006; Maheu et al., 2012), and various contagion phases experienced by financial institutions over time (see, e.g., Billio et al., 2022). Therefore, modeling and forecasting time series subject to regime shifts constitute a crucial area of econometrics. The dissertation contributes to this area in time series econometrics by advancing the application of Markov-switching models; developing Bayesian inference procedures for efficient model estimation; and providing simulation and empirical applications to understand modern economic and financial systems. The contributions are presented in three pieces of self-contained chapters.

Chapter 1 presents a modeling framework and estimation methods for detecting the regime shifts in the currency-liquidity-timing behavior of globally-diversified funds. Our approach builds upon the recent developments of regression models with endogenous Markov switching parameters (Hwu et al., 2021; Kim and Kang, 2022), which allows for capturing the potential regime shifts in both direction and strength of funds’ timing behavior together with underlying drivers that lead to such shifts. An effective Bayesian inference procedure is implemented for model estimation and selection. By analyzing a sample of 382 international fixed income mutual funds, we find that these globally-diversified funds on average engage in currency liquidity timing by adjusting their currency exposure in response to the underlying liquidity movement; however, this timing behavior exhibits significant regime changes across varying market conditions: funds time currency liquidity negatively (adjust their currency exposure in the opposite direction to the liquidity movement) in normal times but switch to aggressively positive timing (adjust their currency exposure in the same direction to the liquidity movement with increasing aggressivity) in turbulent market conditions. We present evidence that these regime-switching timing behaviors are possibly driven by currency liquidity deterioration and negative shock on fund returns.

Chapter 2 concerns the problem of measuring connectedness in financial systems, which is central to modern risk management, including market risk (return connectedness and volatility connectedness); credit risk (default connectedness); counter-party and gridlock risk (bilateral and multilateral contractual connectedness); and systemic risk (system-wide connectedness). The literature on connectedness analysis to date has shown some notable empirical features of financial connectedness, such as the mixture of contemporaneous and temporal dependences, high-dimensionality, and regime shifts. However, existing econometric methods at most capture two of the above features. The novel model we introduce in the second chapter, which is refereed to as a Markov-Switching Graphical Structural Vector Autoregressive (MS-GSVAR) model, facilitates a ”full sweep“ of the list of features. An efficient Bayesian graph inference method is developed to address the computational complexities arising from inference on graph structures in the context of high model dimension, numerous lags, and multiple regimes.
Simulation studies validate the effectiveness of the proposed framework in recovering many empirically relevant dependence structures, and in handling large datasets with changing dependence structures. Our model applied to the volatility series of 96 global banks detects different connectedness states, identifies systemically important individuals, and uncovers the frequency-specific source of connectedness, which are relevant to systemic risk management.

Together, the first and second chapter have inspired the focus of the third chapter. Markov-switching panel models face a major challenge in practical implementation, which is determining how many regimes are necessary to adequately characterize the observed data. Existing solutions typically rely on the assumption that the regime dimension is homogeneous in the cross-section. Such an assumption may be restrictive as individuals are likely to be characterized only by one or a subset of regimes identified from the panel. Chapter 3 proposes a general framework to estimate the number of regimes in Markov-switching panel models, allowing possible heterogeneity in cross-sectional regime dimension. We model individual heterogeneity via a binary matrix where its column dimension and configuration indicate respectively the regime dimension of the whole panel and the units. We develop new Bayesian nonparametric inference to jointly estimate the latent binary matrix and the other model parameters. Simulation studies validate the effectiveness of the proposed framework under different panel settings. An application to US state-level macroeconomic indices illustrates the empirical gains of considering likely heterogeneous regime dimension in the cross-section.

Publication Type:	Thesis (Doctoral)
Subjects:	H Social Sciences > HG Finance
Departments:	Bayes Business School > Bayes Business School Doctoral Theses Bayes Business School > Finance Doctoral Theses

Export

Downloads