Predictability in International AssetReturns: A Reexamination

Abstract

This paper argues that inferring long-horizon assetreturn predictability from the properties of vectorautoregressive (VAR) models on relatively shortspans of data is potentially unreliable. Weillustrate the problems that can arise byreexamining the findings of Bekaert andHodrick(1992), who detected evidence of in-samplepredictability in international equity and foreignexchange markets using VAR methodology for a varietyof countries from 1981–1989. The VAR predictions aresignificantly biased in most out-of-sample forecastsand are conclusively outperformed by a simplebenchmark model at horizons of up to six months.This remains true even after corrections for smallsample bias and the introduction Bayesian parameterrestrictions. A Monte Carlo analysis indicates thatthe data are unlikely to have been generated by astable VAR. This conclusion is supported by anexamination of structural break statistics. We showthat implied long-horizon statistics calculated fromthe VAR parameter estimates are very unreliable.