Independent component analysis for time series

Abstract

El objetivo de esta tesis es aplicar el análisis de componentes independientes (ICA) sobre datos multivariantes de series temporales. También, se propone un nuevo procedimiento para predecir un vector de series temporales a partir de un número reducido de componentes independientes

The aim of this thesis is to analyze the performance of independent component analysis (ICA) when it is applied to a vector of non-Gaussian time series in order to find an "interesting" representation of the observations. First, we give an introduction to the ICA methodology and how it performs on estimating a set of non-Gaussian and statistically independent latent factors. Second, we review some basic ideas of multivariate time series analysis, paying special attention to well known dimension reduction techniques previously proposed in the literature. Third, we give an overview of the existing research that links ICA and time series data. Finally we outline the thesis