Publication: Exploring ICA for time series decomposition
Loading...
Download
Advisors
Tutors
Editor
Universidad Carlos III de Madrid. Departamento de Estadística
Publication date
Defense date
Journal Title
Journal ISSN
Volume Title
Publisher
publication.page.ispartofseries
UC3M Working papers. Statistics and Econometrics
11-11
11-11
Creative Commons license
To cite this item, use the following identifier: https://hdl.handle.net/10016/11285
Abstract
In this paper, we apply independent component analysis (ICA) for prediction and signal
extraction in multivariate time series data. We compare the performance of three
different ICA procedures, JADE, SOBI, and FOTBI that estimate the components
exploiting either the non-Gaussianity, or the temporal structure of the data, or
combining both, non-Gaussianity as well as temporal dependence. Some Monte Carlo
simulation experiments are carried out to investigate the performance of these
algorithms in order to extract components such as trend, cycle, and seasonal
components. Moreover, we empirically test the performance of those three ICA
procedures on capturing the dynamic relationships among the industrial production
index (IPI) time series of four European countries. We also compare the accuracy of the
IPI time series forecasts using a few JADE, SOBI, and FOTBI components, at different
time horizons. According to the results, FOTBI seems to be a good starting point for
automatic time series signal extraction procedures, and it also provides quite accurate
forecasts for the IPIs.