This paper is concerned as to whether it is more appropriate to use aggregate or disaggregate models in forecasting house prices using hedonic modelling. It is accepted that the implicit pricing of some of the attributes is not stable between locations, property types and ages but it is argued that this can be effectively modelled with an aggregate method. The models are developed using a dataset of nearly 18,000 transactions in the UK Midlands region in 1994. The comparative performance of these models is then considered using two approaches. Chow tests of the error differences between actual price and the price predicted by the models suggest that the submarket models lead to statistically significant, though small, improvements. A second approach, using comparison of the root mean square errors, is conducted on the models' forecasts for a 10 per cent sample of nearly 2,000 transactions excluded from the modelling process. This shows little practical difference in the forecasting ability between the two approaches. Great care needs to be taken over sample size if a disaggregate model is used.
In published work on hedonic house price estimation it is not uncommon to examine some of the conditions required for the estimators to have desirable properties such as minimum variance and unbiasedness, in particular spatial autocorrelation. However, other conditions that can give rise to similar difficulties with the estimates are often ignored. If these technical conditions are not met, it is sometimes because the model is misspecified in some way. This paper argues that a wider range of diagnostic statistics should be used in the specification search for a good model, in particular, but not exclusively, those concerned with predictive stability. The paper illustrates this approach by examining both in-sample and out-of-sample diagnostic tests of various specifications of a hedonic house price model using data taken from the sale of over 1,600 properties in the Midlands of the UK in 1999/2000. The models investigated include various specifications of the dependent and independent variables, including models that are non-linear in the parameters. The paper concludes that such statistics can often help in model selection and should be more widely employed.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.