We aim to show the interest of the parareal method (based on the work of Y. Maday and J.L. Lions) for the approximation of discrete problems based on backward scheme using least-squares regression i.e. the Longstaff-Schwarz algorithm, approximation of a backward stochastic differential equations. We propose a parallelization of the time discretization of the backward dynamic programming principle. It allows us to approximate simultaneously the solution at several time-steps. Here, we analyze this algorithm and the convergence of the parareal scheme. We also give some performance results. Because of its parallel scalability the method is well suited to fast evaluation of CVAs. This is a joint work with G. Pagès and O. Pironneau.