Date of Completion


Embargo Period



break detection, high frequency data, time series, logarithmic autoregressive conditional duration

Major Advisor

Nalini Ravishanker

Associate Advisor

Jun Yan

Associate Advisor

Kun Chen

Field of Study



Doctor of Philosophy

Open Access

Open Access


The accurate learning of the underlying structure in high-frequency data has become critical in the analysis of time series for capturing valuable information that facilitates decision-making. The time series data in finance often is large, dynamic, heterogeneous and even structural unstable. Each aspect of these characteristics will add a degree of difficulty in efficient analysis. The goal of this dissertation is to discover the latent structure of dynamic high-frequency data that may have structural breaks, from both univariate and network perspective. We focus our analysis on durations between user-defined events in transaction-by-transaction stock prices from the Trade and Quotes (TAQ) data base at Wharton Research Data Services (WRDS). Our proposed approach can be easily adapted to other models.

The dissertation has three main contributions. First, we propose a fast and accurate distribution-free approach using penalized martingale estimating functions on logarithmic autoregressive conditional duration (Log ACD) models. We discuss three approaches for parameter estimation. Our approach employs effective starting values from an approximating time series model and provides investigators accurate fits and predictions that can assist in trading decisions. Second, we propose a sequential monitoring scheme to detect structural breaks in the estimated parameters of a univariate piecewise Log ACD model. Based on martingale estimating function, this scheme does not require any distributional assumption. This monitoring scheme can detect structural breaks and choose model orders at the same time. Assuming data is given, we compare the performance of our scheme with that of a state-of-the-art offline scheme via simulation studies. Third, we propose a framework for detecting structural breaks in dynamic networks of a large number of stocks. In particular, we discover unobserved dynamic network structure from nodal observations governed by both the latent network and time. Our empirical analysis on the 30 most liquid stocks in S&P100 is an exploratory study. Such an analysis would be useful to economists studying the structural breaks in financial networks.