We consider a kernel-based approach to nonlinear canonical correlation analysis and its implementation for time series. We deduce a test procedure of the reversibility hypothesis. The method is applied to the analysis of stochastic differential equation from high-frequency data on stock returns.

We decompose a stationary Markov process (Xt) as a linear combination of ARMA. These decompositions are deduced from a nonlinear canonical decomposition of the joint distribution of (Xt, Xt−1).