The blessing of dimensionality in forecasting real house price growth in the nine census divisions of the US

Abstract

This paper analyzes whether a wealth of information contained in 126 monthly series used by large-scale Bayesian Vector Autoregressive (LBVAR) models, as well as Factor Augmented Vector Autoregressive (FAVAR) models, either Bayesian or classical, can prove to be more useful in forecasting real house price growth rate of the nine census divisions of the US, compared to the small-scale VAR models, that merely use the house prices. Using the period of 1991:02 to 2000:12 as the in-sample period and 2001:01 to 2005:06 as the out-of-sample horizon, we compare the forecast performance of the alternative models for one- to twelve–months ahead forecasts. Based on the average Root Mean Squared Error (RMSEs) for one- to twelve–months ahead forecasts, we find that the alternative FAVAR models outperform the other models in eight of the nine census divisions.