Neural Networks and Linear Models in Real Estate Appraisal: the impact of sets selection procedures

The use of Neural Networks in real estate appraisal has been recently subject of renewed interest by the scientific community. Generally, their effective use requires the availability of a large database, otherwise facing the real risk, even with an excellent performance on the "training set", of obtaining unsatisfactory generalisation properties (the so called over fitting effect). The well-known linear regression models (ARMs), on the other side, require fewer parameters for their optimisation but are unable to capture complex nonlinear relationships. Since large databases are usually difficult to find in the real estate market, ARM models often provide better results than Artificial Neural Networks (ANNs). Furthermore, the latter require considerable effort to be effectively trained, both in finding the best structure and in estimating the characterising parameters. The optimisation process that leads to an efficient neural network requires a long job as well as considerable computational capabilities. This contribution, after outlining the state of the art in the use of ANNs and confirming that the scarcity of real estate market data often turned out to be a serious obstacle in their concrete application, proposed an innovative algorithm for selecting the data used in the training process. Such an algorithm seems to be able to improve predictive performance: networks that seek to take full advantage of the information available for learning and seem to have better abilities in generalising the behaviour of the underlying phenomenon than those that are trained with completely randomly selected data, as usually done in practice.