Abstract: We propose procedures designed to uncover structural breaks in the co-movements of financial markets. A reduced form approach is introduced that can be considered as a two-stage method for reducing the dimensionality of multivariate heteroskedastic conditional volatility models through marginalization. The main advantage is that one can use returns normalized by volatility filters that are purely data-driven and construct general conditional covariance dynamic specifications. The main thrust of our procedure is to examine change-points in the co-movements of normalized returns. The tests allow for strong and weak dependent as well as leptokurtic processes. We document, using a ten year period of two representative high frequency FX series, that regression models with non-Gaussian errors adequately describe their co-movements. Change-points are detected in the conditional covariance of the DM/US$ and YN/US$ normalized returns over the decade 1986-1996.
Key words and phrases: Change-point tests, conditional covariance, high-frequency financial data, multivariate GARCH models.