Abstract: A Volterra adaptive filter is proposed for the multivariate chaotic time series.The initial Volterra kernel of the filter is determined by the least square method based on singular value decomposition,and the real time filter coefficient is determined by an NLMS(normalized least mean square) algorithm.The simulations clearly show that one-step prediction performance of this Volterra adaptive prediction method of multivariate chaotic time series improves 102 times on the prediction accuracy over those of univariate time series under the noise-free Lorenz time series,which indicates that this model has better convergence performance,and its structure is easy to implement.The multi-step prediction performance of this multivariate Volterra predictor is superior to the one of the local polynomial prediction method of multivariate chaotic time series under the noise added Lorenz time series.