We examine the predictability of private and public real estate returns using recursive, out-of-sample, linear and Markov switching models, employing a rich set of predictor variables. We find considerable improved predictive power compared to simple regression models, especially at the intermediate horizon. Next, we test whether such improved forecasting accuracy translates into a positive risk-adjusted out-of-sample performance by performing a recursive mean-variance portfolio allocation analysis. We observe significant improvements in realized Sharpe ratios and mean-variance utility scores, especially when employing Markov switching models and exploiting predictability at intermediate horizons. Furthermore, our results are robust to the inclusion of transaction costs.