Studies on break detection in financial time series volatility

PhD Thesis
Start Year: 
2009
Year of Completion: 
2015

The aim of this thesis is to provide an econometric analysis on volatility dynamics by examining the implications of structural changes. Specifically, it analyses how the existence of structural changes may influence the volatility persistence and/or long memory in financial time series.

In the second chapter a Monte Carlo simulation experiment it is employed to examine the performance of a CUSUM type statistic for break detection. In particular, we study the statistical properties of a non-parametric approach for single and multiple breaks detection by employing a large number of different long run variance estimators and different types of breaks. However it cannot be supported the adoption of a single long run variance estimator to inflate the algorithm for the break detection.

In the third chapter we examine the effects of structural changes on the volatility dynamics of a market has been basically unexplored in the context of structural change, the Athens Stock market. A CUSUM type statistic is used for the break detection. Once the breaks are accounted for the volatility persistence is substantially reduced.

The fourth chapter focuses on the interaction between structural changes and outliers on the volatility. To this end, it employs a wavelet-based outlier detection method along with a non-parametric break detection approach. The examination of GARCH models reveals that the series are highly persistent, if breaks are not accounted for, while ignoring outliers induces biases to GARCH parameters estimates.

The last chapter provides empirical evidence whether long memory in daily log-range series could be explained by the presence of structural changes. A multiple break model that assumes abrupt shifts or jumps in volatility and a smooth transition model that allows abrupt shifts, smooth shifts or a combination are employed to detect and characterize breaks. The analysis revealed that long memory could be considered to some extent spurious. When accounting for the level shifts, the evidence in favour of long-memory in the log-range series is no longer supported.