Forecasting real house price of the U.S. : an analysis covering 1890 to 2012

Abstract

This paper evaluates the ability of Bayesian shrinkage-based dynamic predictive regression models estimated with hierarchical priors (Adaptive Jefferys, Adaptive Student-t, Lasso, Fussed Lasso and Elastic Net priors) and nonhierarchical priors (Gaussian, Lasso-Lars, Lasso-Landweber) in forecasting the U.S. real house price growth. We also compare results with forecasts from bivariate OLS regressions and principal component regression. We use annual dataset on 10 macroeconomic predictors spanning the period 1890 to 2012. Using an out-of-sample period of 1917 to 2012, our results based on MSE and McCracken (2007) MSE-F statistic, indicate that in general, the non-hierarchical Bayesian shrinkage estimators perform better than their hierarchical counterparts as well as the least square estimators. The Bayesian shrinkage estimated with Lasso-Landweber is the best-suited model for forecasting the U.S. real house price. Among the least square models, the individual regression with house price regressed on the fiscal policy variable outperforms the rest. Also results from Lasso-Landweber portray the fiscal policy variable as the best predictor of the U.S. house prices especially in the recent times while the short-term interest rate and real construction cost also did well at the beginning and middle of the sample.