Abstract: For an equivalent canonical form of multivariate time series transfer function model,a modeling and forecasting scheme,which consists of preliminary modeling and Extended Kalman Adaptive Estimation-Prediction (EKAEP) algorithm,is presented in this paper.In consideration of time-variation and outliers of physical systems,both the noise statistic property Limited Memory Adaptive Estimator (IMAE) and the outlier detection method based on the criterion of doube believable intervals are included in the EKAEP algorithm.The efficiency of the scheme has been proved by several simulations.