Empirical information criteria for time series forecasting model selection

In this paper, we propose a new empirical information criterion (EIC) for model selection. It is applicable to situations involving a large number of time series to be forecast. For example, it can be applied to a large inventory of products for which sales need to be forecast on a monthly basis. Our new criterion provides a data-driven model selection tool which can be tuned to the particular forecasting task. The penalty function for each series is chosen based on the other series. We compare the EIC with other model selection criteria including Akaike’s Information Criterion (AIC) and Schwartz’s Bayesian Information Criterion (BIC). The comparison show that for the M3 forecasting competition data, the EIC outperforms both the AIC and BIC.