Recent research [e.g., DeMiguel, Garlappi and Uppal, (2009), Rev. Fin. Studies] has cast doubts on the out-of-sample performance of optimizing portfolio strategies relative to naive, equally weighted ones. However, existing results concern the simple case in which an investor has a one-month horizon and meanvariance preferences. In this paper, we examine whether their result holds for longer investment horizons, when the asset menu includes bonds and real estate beyond stocks and cash, and when the investor is characterized by constant relative risk aversion preferences which are not locally mean-variance for long horizons. Our experiments indicates that power utility investors with horizons of one year and longer would have on average benefited, ex-post, from an optimizing strategy that exploits simple linear predictability in asset returns over the period January 1995 - December 2007. This result is insensitive to the degree of risk aversion, to the number of predictors being included in the forecasting model, and to the deduction of transaction costs from measured portfolio performance.