Factor Model for Forecasting with Multi-Collinearity and Nonlinear Dependence

• Activity Number: 38
• Type: Contributed
• Date/Time: Sunday, July 29, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Business and Economic Statistics Section
• Author(s): Joseph Egbulefu*+
• Companies: Rice University
• Keywords: multicollinearity ; non-linear least squares ; dynamic factor model ; principal component analysis ; partial least squares ; PCA PLS

Abstract

Empirical analysis of financial time series has identified non-linear dependence properties inherent in financial variables. Factors based on Principal Component Analysis and Partial Least Squares, empirical methods used for forecasting under multi-collinearity, can be deficient in extracting certain non-linear properties. We construct a dynamic factor model for asset prices and returns using non-linear least squares to identify dependencies inherent in variables and constructing factor loadings from singular vectors of the data matrix under a suitable non-linear transformation. The method has been shown to outperform PCA and PLS forecasting when applied to high frequency exchange rates.