A novel neural network technique for the valuation of residential property

Abstract

A number of published studies have investigated the application of neural network technology to residential property appraisal. The majority of these studies have concentrated on homogeneous areas (that is areas where properties are subject to the same environment and locational factors). This is generally done to restrict the data set to one local sub- market. However, the models created are specialised and not locationally portable. This paper presents a methodology, which builds on research reported by James [1], in which a Kohonen map is used to uncover sub-markets within a large data set that are subsequently independently used to train a series of back-propagation networks. (The paper also introduces a novel boundary detection algorithm for a Kohonen self organising map.) The study concludes that by modelling possible sub-markets an acceptable accuracy over a heterogeneous area can be achieved. The work presented in this paper is funded via a Realising Our Potential Award under the auspices of the ESRC.