Housing Market Forecasting with Factor Combinations

In this paper we take a computational approach to forecasting a macroeconometric model of housing markets across six original datasets with large cross-sectional dimensions. We compare a large number of models which vary by the choice of factors, 'observable endogenous variables' and the number of lags in addition to classical and modern (factor based) specifications. We utilize various optimal model selection and model averaging techniques, comparing them against classical benchmarks. Within a 'pseudo real-time' out of sample forecasting context, results show that the approximate BMA method is the best weighting and selection technique, generating forecasts able to outperform the automated univariate benchmark of Hyndman and Khandakar (2008) upwards of 58% of the time. However, the average forecast error is lower in magnitude over all recursions and countries for the benchmark compared with all models for all variables. We also provide results on the biased nature of this class of models in general, in addition to the forecast error increasing as a function for the underlying variance of the series being forecast.