Robust Estimator of Persistence in Financial Time Series

Abstract

The goal of this thesis is to develop a novel robust log-periodogram regression method to detect the presence of long memory in time series. By the use of the Least Trimmed Squares regression we obtain an estimator that is less sensitive to outliers and leverage points, which is highly desirable particularly because the Periodogram estimator itself is prone to such inhomogeneities. In a Monte Carlo study, the new estimator provides smaller bias than the classical Least Squares log-Periodogram estimator. On the other hand the variability of estimation is increased. The proposed estimator is compared to existing long memory estimators on a case study of international currency exchange rates.