Abstract: ARMA models are widely used in the field of management science.Combined forecasting can impove the effect of forecasting.However,how to select the best model group is very important but not well done.In this paper we propose a best model group selection method based on the Kullback-Leibler (K-L) information.First we measure the so called K-L distances between every candidate model and the true model using the K-L information,and then derive the confidence intervals of the gap between the K-L distence of each candidate model and the best model using the central limitation theory.Furthermore based on the confidence intervals we identify agroup of models,which are not different significantly with the best model,as the best model group.Finally we compare the forecast ability of the best model group and the best model.The results show that the proposed method can improve the forecast with high probability when the best model is not the true model.

Key words: Kullback-Leibler information, the best model group, combined forecast