Testing for Unconditional Predictive Ability

Abstract

AbstractThis article provides an overview of methods for constructing tests of unconditional predictive ability. Section 2 begins by describing the necessary notation for delineating these results. Sections 3 and 4 provide a discussion of the two main forms of the null hypothesis: population-level unconditional tests of predictive ability and finite-sample unconditional tests of predictive ability. For each, the main methodological approaches to conducting inference are discussed, emphasizing results in Clark and McCracken (2001, 2005a, 2009a), Giacomini and White (2006), McCracken (2007), and West (1996). In both instances, the discussion is in the context of using estimated parametric models to form a point prediction of a scalar-dependent variable y. Other types of forecast are referenced when relevant. Section 5 provides a brief overview of other developments in this literature and suggests new directions for future research. Section 6 provides Monte Carlo evidence on the efficacy of selected methods, and Section 7 concludes.