Abstract:
This paper develops large-scale Bayesian Vector Autoregressive (BVAR) models, based on 268
quarterly series, for forecasting annualized real house price growth rates for large-, medium- and smallmiddle-
segment housing for the South African economy. Given the in-sample period of 1980:01 to
2000:04, the large-scale BVARs, estimated under alternative hyperparameter values specifying the
priors, are used to forecast real house price growth rates over a 24-quarter out-of-sample horizon of
2001:01 to 2006:04. The forecast performance of the large-scale BVARs are then compared with
classical and Bayesian versions of univariate and multivariate Vector Autoregressive (VAR) models,
merely comprising of the real growth rates of the large-, medium- and small-middle-segment houses,
and a large-scale Dynamic Factor Model (DFM), which comprises of the same 268 variables included
in the large-scale BVARs. Based on the one- to four-quarters ahead Root Mean Square Errors (RMSEs)
over the out-of-sample horizon, we find the large-scale BVARs to not only outperform all the other
alternative models, but to also predict the recent downturn in the real house price growth rates for the
three categories of the middle-segment-housing over an ex ante period of 2007:01 to 2008:02.
