[BHK08] Wavelets and principal components regression

Conférence Internationale avec comité de lecture : Interims meeting of the ISA RC33 on “Logic and Methodology in Sociology”, Naples, Italy, September 2008, pp.50,
Résumé: Principal Components Regression (PCR) is an applied regression on PCA factors beforehand carried out on initially strongly correlated variables. The use of the PCA allows replacing the initial variables by principal components which preserve the quasi-total information and which have the advantage of being not correlated. These components are taken as explanatory variables Xi for a multiple linear regression. The PCR modelling quality (better than that of a simple regression) remains affected by the existence of noise in initial variables. In this work, we propose a denoising procedure based on wavelet thresholding technique. This denoising allows to separate the signal from the noise without losing information. We applied this technique to French stock exchange data. We study profitability with respect to factors influencing stock price changes. We show that removing noise from the Xi variables using a wavelet-based soft thresholding improves the quality of adjustment of the regression model (PCR after denoising) as well as forecasting quality. KEY WORDS: Wavelets, Thresholding, PCR, PCA, French stock exchange, Returns.

Equipe: msdma


@inproceedings {
title="{Wavelets and principal components regression}",
author=" S. Benammou and N. Haouas and Z. Kacem ",
booktitle="{Interims meeting of the ISA RC33 on “Logic and Methodology in Sociology”, Naples, Italy}",