Title: VAR estimation impacts on frequency causality measures
Authors: Thibault Soler - University Paris 1 -- Pantheon-Sorbonne (France) [presenting]
Philippe De Peretti - University Paris 1 -- Pantheon-Sorbonne (France)
Christophe Chorro - University (France)
Emmanuelle Jay - Fideas Capital (France)
Abstract: Granger non-causality tests have received a great deal of attention over recent years. In time domain, tests have been extended to deal with co-integration, near unit roots, mixtures of variables with different integration and structural breaks. In the frequency domain, methodologies have been developed to quantify causal relationships. These approaches are two-step ones consisting in first estimating a Vector AutoRegressive model (VAR), and then computing coherence of transfer function. Nevertheless, the first step may be incorrectly performed for many reasons: incorrect model order selection, insufficient number of observations regarding the number of series, the omission of zero-lag effect, etc., then we are likely to have cascading errors and flawed results. The goal is, therefore, twofold: first, evaluate the impact of a non-efficient VAR estimation on causality measures in the frequency domain, then suggest an efficient two-step methodology to overcome cascading errors in estimating a multivariate model. The first step consists in using the modified backward in time selection method (mBTS), and the second one makes use of the top-down strategy (TD). The impact of error diagonal covariance, time series lengths, and instantaneous effects on causality measures is studied first on simulated data, and then on financial data. The results are compared between our methodology (mBTS-TD) and existing methods of variable subset selection (mBTS, TD, bottom-up (BU)).