Financial High-Frequency Data

In finance, stock prices, exchange rates and commodity prices are recorded with each transaction or bid/ask offer resulting in intraday high-frequency data. Such time series have a very fine time scale (e.g. seconds or even fractions of seconds). High-frequency data have several specifics distinguishing them from low-frequency data (e.g. daily time series). The observations of prices are irregularly spaced, the values of prices are discrete and the price process consists of both bid and ask side. The latter two specifics are often captured by the market microstructure noise which conceals the information about the theoretical efficient price process. Over the past 20 years, a vast number of econometrical methods have been developed to address such market microstructure issues and to understand dynamics of financial markets. We focus on three aspects of high-frequency data analysis. First, we model durations between successive transactions using the autoregressive conditional duration model in a discrete framework with a special attention to split transactions. Second, we estimate and forecast quadratic variation of the price process by a variety of methods robust to the market microstructure noise. Third, we model prices as the Ornstein-Uhlenbeck process contaminated by the market microstructure noise with an application to the pairs trading strategy. In the empirical part of the thesis, we analyze selected stocks traded on the NYSE and NASDAQ exchanges.