We take a computational approach to forecasting real and nominal house prices, comparing a large number of models varying by the choice of factors, 'observable endogenous variables' and the number of lags, in addition to classical and modern econometric models. We utilize various optimal model selection and model averaging techniques, comparing them against classical benchmarks. Using six original datasets with large cross-sectional dimensions, we include recent developments in the factor literature as part of our model set. Within a 'pseudo real-time' out of sample forecasting context, results show that model averaging across a large set of candidate factor models is able to consistently generatye forecasts with favorable properties, but unable to consistently generate the singularly lowest forecast error. Other results include forecast error increasing in horizon, error magnitude varying by country (the variance of the underlying series) and that errors are lower for nominal, as opposed to real indexes.
