Location, location, location: A comparison of the predictive validity of a neural network and regression methodology and their performance in the construction of a heterogeneous residential appraisal model

The appraisal of residential property is one of the most time consuming steps in the mortgage lending process. The financial lending industry has long recognized the need for an automated process. This study attempts to construct a heterogeneous appraisal model using two different automated methodologies; regression and a neural network. The results of the two methodologies are then compared as to forecasting accuracy.;This study examines 25,857 data observations which consists of properties sold in the greater Memphis, Tennessee area over the period November, 1985-December, 1994 and conducts a definitive and comprehensive analysis of the two methodologies.;Although much has been written on property appraisal using various automated techniques, many deficiencies exists in these studies including limited data sets, non-domestic data, non-comparable properties or time periods and models being built around a homogeneous data set such as a particular neighborhood or census tract. Furthermore, previous research is inconclusive as to which methodology, regression analysis or a neural network offers a superior appraisal technique.