Statistical Inference with High-Frequency Financial Data : New Perspectives from Alternative Observation Schemes

Yu, Shifan and Nolte, Ingmar and Nolte, Sandra and Li, Yifan (2024) Statistical Inference with High-Frequency Financial Data : New Perspectives from Alternative Observation Schemes. PhD thesis, Accounting and Finance.

Text (2024yuphd)
2024yuphd.pdf - Published Version
Download (4MB)

Abstract

This thesis is a compilation of three main studies with the common theme: statistical inference with high-frequency financial data under alternative observation or sampling schemes. The increasing availability of high-frequency financial data has motivated the development of new statistical inference tools, such as (i) advanced volatility estimators with better robustness in the presence of local extreme phenomena in intraday asset prices, and (ii) techniques for identifying these unusual extreme events, which contribute to a comprehensive understanding of their impact on financial markets. The investigation on both topics in this thesis adopts an innovative perspective by exploring potential enhancements from utilizing high-frequency financial data under alternative observation schemes, which stands in contrast to traditional methods that rely on equidistantly sampled transaction price records. In Chapter 2, we introduce a novel nonparametric high-frequency jump test for discretely observed Itô semimartingales. Based on observations sampled recursively at first exit times from a symmetric double barrier, our method distinguishes between threshold exceedances caused by the Brownian component and jumps, which enables the construction of a feasible, noise-robust statistical test. Simulation results demonstrate superior finite-sample performance of our test compared to classical methods. An empirical analysis of NYSE-traded stocks provides clear statistical evidence for jumps, with the results highly robust to spurious detections. In Chapter 3, we develop a new nonparametric estimator of integrated variance that utilizes intraday candlestick information, comprised of the high, low, open, and close prices within short time intervals. The range-return-difference volatility (RRDV) estimator is robust to short-lived extreme return persistence hardly attributable to the diffusion component, such as gradual jumps and flash crashes. By modelling such sharp but continuous price movements following some recent theoretical advances, we demonstrate that RRDV can provide consistent estimates with variances about four times smaller than those obtained with the differenced-return volatility (DV) estimator. Monte Carlo simulations and empirical applications further validate the practical reliability of our proposed estimator with some finite-sample refinements. In Chapter 4, we introduce an innovative semiparametric framework for duration-based volatility estimation. We filter out daily volatility dynamics from intraday price durations by employing a nonparametrically predicted threshold that dynamically adapts to the volatility of each day. This enables the application of parametric models to price durations collected over various days, which greatly enhances the flexibility of model estimation and facilitates the construction of more accurate volatility estimators. Simulation results demonstrate superior finite-sample performance of our duration-based estimators for both spot and integrated volatility compared to some established methods. An empirical application based on intraday data for the SPDR S&P 500 ETF highlights the improved forecasting accuracy of our integrated volatility estimator within a standard daily volatility forecasting framework. Furthermore, an intraday analysis based on our spot volatility estimator reveals an immediate and substantial impact of FOMC news announcements on market volatility.