Information Criterion for Selecting Competing Models Using Different Weighting Schemes

Abstract

In this study, we propose a new information criterion for model selection in univariate time series analysis. This is done by weighting the penalty terms of the three most commonly used information criteria (i.e. Akaike Information Criterion (AIC), Schwarz Information Criterion (SIC) and Hannan Quinn Information Criterion (HQ)). The penalties of these three information criteria are considered as a linear function. Two different methods for calculating an appropriate weight (a model performance-based criterion and an eigenvector-based criterion) are considered for the new information criterion. A simulation of a univariate time series models with five different sample sizes (i.e. 15, 25, 45,100 and 500) is carried out to compare the performance of two proposed information criteria (Model Performance based weight information criterion (PWIC) and Eigenvectors based weight information criterion (EIC)) to three existing information criteria. Results obtained shows that the EIC and HQ outperforms SIC and AIC on the subject of the overall performance in the selection of the order of univariate time series model. Hence, the results indicate that the new information criterion based on the eigenvector weight work well in finite sample. The new approach shows outstanding performance with the HQ criterion in selecting the "best" model and this combination is recommended for model selection.