This paper considers a U.S. institutional investor who is implementing a long-term portfolio allocation using forecasts of financial returns. We compare the predictive performance of two competing macro-finance models ---an unrestricted Vector AutoRegression (VAR) and a fully-structural Dynamic Stochastic General Equilibrium (DSGE) model--- for horizons up to 15 years. Although the performances are similar for short horizons, the DSGE model outperforms the VAR at forecasting financial returns in the long term. This model also generates substantially higher Sharpe ratios. Although it contains fewer unknown parameters, it benefits from economically grounded restrictions that help anchor financial returns in the long term.