Spatial Bayesian methods of forecasting house prices in six metropolitan areas of South Africa

Abstract

This paper estimates Spatial Bayesian Vector Autoregressive (SBVAR) models, based on the First-Order Spatial Contiguity and the Random Walk Averaging priors, for six metropolitan areas of South Africa, using monthly data over the period of 1993:07 to 2005:06. We then forecast one- to six-months-ahead house prices over the forecast horizon of 2005:07 to 2007:06. When we compare forecasts generated from the SBVARs with those from an unrestricted Vector Autoregressive (VAR) and the Bayesian Vector Autoregressive (BVAR) models based on the Minnesota prior, we find that the spatial models tend to outperform the other models for large middle-segment houses; while the VAR and the BVAR models tend to produce lower average out-of-sample forecast errors for middle and small-middle segment houses, respectively. In addition, based on the priors used to estimate the Bayesian models, our results also suggest that prices tend to converge for both large- and middle-sized houses, but no such evidence could be obtained for the small-sized houses.