Specification of Economic Time Series Models Using Akaike's Criterion

Abstract

In this article we use the procedures introduced by Akaike (1974, 1976) and Schwarz (1978) to investigate the optimal specification of autoregressive moving average (ARMA) models for a group of major economic time series (see also Fine and Hwang 1979; Sawa 1978; Kashyap 1977; Leamer 1979; Chow 1979). The basic question we ask is whether the short AR representations that are estimated on an ad hoc basis in economics are validated by these more objective criteria. The article also compares the results given by Akaike and Schwarz procedures and studies the sensitivity of these results to the preliminary transformations used to reach stationarity. The main conclusion is that the two procedures designate significantly different models as optimal, the Schwarz procedure in general favoring low-order AR specifications. The reasons for this difference are mentioned and the crucial role played by the initial transformations is emphasized. In particular, we show that optimal specification designated by Schwarz criteria is sensitive to the initial transformation selected to reach stationarity. The ad hoc nature of these prefilters is probably the most inadequate aspect of the article-although our transformations are widely used (see, e.g., Pierce 1977).