Abstract: Briefly describes the modeling procedure of the threshold autoregressive model and its disadvantages. Presents an improved modeling method and procedure. The model is transformed and the new parameters are estimated based on the differential data and some initial data by means of the linear least square rnethod. Then the original model parameters are obtained. Equation matrices and its inversion and the model parameters can be recurrently calculated when the model order is inereased. A set of real sample data is processed on an IBM-PC/XT and the results show that the improved method can save ca. 60% in time of computation attaining the same precision