A comparison of data mining methods for mass real estate appraisal

We compare the performance of both hedonic and non-hedonic pricing models applied to the problem of housing valuation in the city of Madrid. Urban areas pose several challenges in data mining because of the potential presence of different market segments originated from geospatial relations. Among the algorithms presented, ensembles of M5 model trees consistently showed superior correlation rates in out of sample data. Additionally, they improved the mean relative error rate by 23% when compared with the popular method of assessing the average price per square meter in each neighborhood, outperforming commonplace multiple linear regression models and artificial neural networks as well within our dataset, comprised of 25415 residential properties.