Abstract
Parameter estimation and model selection are two main steps in the procedure of modeling threshold autoregressive model. The latter one includes selection of threshold variable and AR order, which play key role in the performance of modeling. Selecting threshold variable according to F statistic proposed by Tsay (1989) or according to method of MLE are two simple and widely used methods. However, it is showed in present study by simulation as well as real data that threshold variables selected by these two methods not necessary brings the fitted model with good prediction performance. Modeling two threshold autoregressive models according to the first two threshold variables chosen by either of the two methods and then averaging the one-step prediction usually decrease PMSE. The ILI, Lynx and sunspot data are used for illustration.