A0502
Title: Univariate properties and the limits to multivariate predictability
Authors: Donald Robertson - University of Cambridge (United Kingdom) [presenting]
Stephen Wright - Birkbeck College - University of London (United Kingdom)
Abstract: Low-order univariate ARMA processes fit the data nearly as well as multivariate models, and frequently do better at out-of-sample prediction. Multivariate VAR/ ABCD models also typically imply much higher-order univariate reduced forms. These observations are reconciled by establishing conditions for multivariate systems to improve upon (fundamental) ARMAs. A prior study shows that, for any given series, the maximal-predictive model arises from a univariate representation with hidden (i.e., nonfundamental) univariate shocks. If these representations are treated as the DGP, it is shown that multivariate models can only improve on the fundamental ARMA if a) the nonfundamental univariate shocks are sufficiently correlated and b) the univariate representations of different series have sufficiently different moving average parameters. Empirically, these conditions do not appear to apply to a standard set of macro variables.
