M Kheradmandia - Isfahan University Of Technology

In choosing one model amongst several models with different dimensions, The maximum likelihood principle invariable leads to choosing the model with highest dimension. In order to overcome. This problem akaike (1973) introduced an information criterion (AIC). Later Schware (1978) introduced an alternative criterion (8IC). Various other similar sriteria have appeared in the literature. Smith and sptegelhalter (1980) introduced a unified view of number of choice criterion. In this paper we introduce a new criterion (OPA) which is based on calculation of posterior probilities of competing models using what we tern as “overlap data based priors” for model parameters. Through a Large number of simulated AR series we compare the extent to which different criteria can identify the true underlying generating mechanism of series.

Model selection, predictive ability, poscerior mode, AIC, BIC, OPA.