Detecting patterns in Time Series Data with applications in Official Statistics

Abstract

This thesis examines the issue of detecting components or features within time series data in automatic procedures. We begin by introducing the concept of Wavelets and briefly show their usage as a tool for detection. This leads to our first contribution which is a novel method using wavelets for identifying correlation structures in time series data which are often ambiguous with very different contexts. Using the properties of the wavelet transform we show the ability to distinguish between short memory models with changepoints and long memory models. The next two Chapters consider seasonality within data, which is often present in time series used in Offical Statistics. We first describe the historical evolution of identification of seasonality, comparing and contrasting methodology as it has expanded throughout time. Following this, motivated by the increased use of high-frequency time series in Official Statistics and a lack of methods for identifying low-frequency seasonal components within high-frequency data, we present a method for identifying periodicity in a series with the use of a simple wavelet decomposition. Presented with theoretical results and simulations, we show how the seasonality of a series is uniquely represented within a wavelet transform and use this to identify low frequency components which are often overlooked in favour of a trend, with very different interpretations. Finally, beginning with the motivation of forecasting European Area GDP at the current time point, we show the effectiveness of an algorithm which detects the most useful data and structures for a Dynamic Factor Model. We show its effectiveness in reducing forecasting errors but show that under large scale simulation that the recovery of the true structure over two dimensions is a difficult task. All the chapters of this thesis are motivated by, and give applications to, time series from different areas of Official Statistics.