Abstract: In this paper, a practical algorithm of modeling and forecasting is presented for the extended Box-Jenkins multivariable time series transfer-function model. It contains an Exponent-Weighted Limited Memory Adaptive Estimator for statistic property of noise, and a measuring method Of double believable intervals for outliers as well as the estimating and forecasting formulas. Compared with other algorithms, its calculation is simpler and the speed of convergence is faster because the algorithm is linear and both the system parameters and the noise parameters of the transfer-function model are estimated simultaneously. The efficiency of the algorithm has been proved by simulation and application.