Essays in Financial Econometrics

Abstract

This dissertation presents three essays in financial econometrics, addressing questions in forecasting in financial markets, high-frequency factor models, and spot covariance estimation. The first chapter provides an overview of these topics. The remainder of this dissertation is summarized as follows.In the second chapter, which is joint work with Andrew Patton, we propose methods to improve the forecasts from generalized autoregressive score (GAS) models (\citet{creal2013generalized}; \citet{harvey2013dynamic}) by localizing their parameters using decision trees and random forests, which exploit information in state variables from within and possibly beyond the model. These score-driven models are widely used in econometrics literature for forecasting volatility, correlation etc. The proposed methods allow the researcher to draw on information from multiple state variables simultaneously and avoid the curse of dimensionality faced by kernel-based approaches. We apply the new models in four distinct empirical analyses, and in all applications the proposed new methods significantly outperform the baseline GAS model. In our applications to stock return volatility and density prediction, the optimal GAS tree model reveals a leverage effect and a variance risk premium effect. Our study of stock-bond dependence finds evidence of a flight-to-quality effect in the optimal GAS forest forecasts, while our analysis of high-frequency trade durations uncovers a volume-volatility effect.Next, the third chapter, which is also joint work with Andrew Patton, analyzes variation in the factor structure of asset returns within a trade day by combining non-parametric kernel methods with principal component analysis. We estimate the model on a collection of over 400 high frequency U.S. equity returns over the period 1996-2020 and show that the proposed model has superior explanatory power relative to a collection of well-known observable factor models and standard PCA. We present a stylized model of asset prices and information flows and show that the factor structure of asset returns varies with the arrival of news. Using data on individual firm earnings announcements, FOMC announcements, and other macroeconomic announcements, we provide evidence consistent with our stylized model, that the superior performance of the proposed model is due to time variation in the factor structure of asset returns around times of information flows.The fourth chapter concerns about the precise estimation of spot covariance matrices. Spot covariance estimation is commonly based on high-frequency open-to-close return data over short time windows, but such approaches face a trade-off between statistical accuracy and localization. In this chapter, I introduce a new estimation framework using high-frequency candlestick data, which include open, high, low, and close prices, effectively addressing this trade-off. By exploiting the information contained in candlesticks, the proposed method improves estimation accuracy relative to the return-based approach while preserving local structure. I further develop a test for spot covariance inference based on candlesticks that demonstrates reasonable size control and a notable increase in power, particularly in small samples, compared to the test built on returns. Motivated by recent work in the finance literature, I empirically test the market neutrality of the iShares Bitcoin Trust ETF (IBIT) using 1-minute candlestick data for the full year of 2024. The results show systematic deviations from market neutrality, especially in periods of market stress. An event study around FOMC announcements further illustrates the new method's ability to detect subtle shifts in relatively mild information events.Finally, the fifth chapter concludes.