One of the most important findings in empirical finance has been the fact that returns are not i.i.d. Predictability is one of the basic ingredients of asset pricing and portfolio choice models nowadays, but there is still high uncertainty about its true extent. This paper develops a careful evaluation of predictability of returns in a relevant context for portfolio management. First, multiple assets must be included in the analysis. In this paper, monthly U.S. excess returns of bonds and stocks are jointly studied. Second, density forecasts of returns are required, not only point forecasts. Predictive distributions in a Bayesian framework are used to take into account parameter uncertainty. Third, conclusions about predictability are drawn from evaluation of density forecasts using out-of-sample checks against realizations of returns. Density forecasts are computed for normal VAR, fat tails and Markov-switching models.