This paper provides evidence on the economic significance of the predictability in U.S. stock returns using a real-time asset allocation framework. We examine the performance of a Bayesian investor who relies on conditioning information (dividend yield, T-bill yield, default spread, and term spread) to forecast future returns and contrast it with that of an otherwise identical investor who believes in i.i.d. returns. We find that the relative performance of the information-based strategy is unstable over time, being noticeably poor during 1989–2002. In marked contrast, the strategy performs significantly better when it relies on a model-based approach based on the CAPM.