The main purpose of this dissertation is to introduce and critically assess some novel statistical methods for change-point detection that help better understand the nature of processes underlying observable time series. First, we advocate the use of change-point detection for local trend estimation in financial return data and propose a new approach developed to capture the oscillatory behaviour of financial returns around piecewise-constant trend functions. Core of the method is a data-adaptive hierarchically-ordered basis of Unbalanced Haar vectors which decomposes the piecewise-constant trend underlying observed daily returns into a binary-tree structure of one-step constant functions. We illustrate how this framework can provide a new perspective for the interpretation of change points in financial returns. Moreover, the approach yields a family of forecasting operators for financial return series which can be adjusted flexibly depending on the forecast horizon or the loss function. Second, we discuss change-point detection under model misspecification, focusing in particular on normally distributed data with changing mean and variance. We argue that ignoring the presence of changes in mean or variance when testing for changes in, respectively, variance or mean, can affect the application of statistical methods negatively. After illustrating the difficulties arising from this kind of model misspecification we propose a new method to address these using sequential testing on intervals with varying length and show in a simulation study how this approach compares to competitors in mixed-change situations. The third contribution of this thesis is a data-adaptive procedure to evaluate EEG data, which can improve the understanding of an epileptic seizure recording. This change-point detection method characterizes the evolution of frequencyspecific energy as measured on the human scalp. It provides new insights to this high dimensional high frequency data and has attractive computational and scalability features. In addition to contrasting our method with existing approaches, we analyse and interpret the method’s output in the application to a seizure data set.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:700961 |
| Date | January 2016 |
| Creators | Schröder, Anna Louise |
| Publisher | London School of Economics and Political Science (University of London) |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://etheses.lse.ac.uk/3421/ |
Page generated in 0.0024 seconds