Predicting downturns in US housing market : a Bayesian approach

Abstract

This paper estimates Bayesian Vector Autoregressive (BVAR) models, both spatial and non-spatial (univariate and multivariate), for the twenty largest states of the US economy, using quarterly data over the period 1976:Q1 to 1994:Q4; and then forecasts one-to-four quarters ahead real house price growth over the out-of-sample horizon of 1995:Q1 to 2006:Q4. The forecasts are then evaluated by comparing them with the ones generated from an unrestricted classical Vector Autoregressive (VAR) model and the corresponding univariate variant the same. Finally, the models that produce the minimum average Root Mean Square Errors (RMSEs), are used to predict the downturns in the real house price growth over the recent period of 2007:Q1 to 2008:Q1. The results show that the BVARs, in whatever form they might be, are the best performing models in 19 of the 20 states. Moreover, these models do a fair job in predicting the downturn in 18 of the 19 states, however, they always under-predict the size of the decline in the real house price growth rate – an indication of the need to incorporate the role of fundamentals in the models.