Analyzing Massive Data Sets: An Adaptive Fuzzy Neural Approach for Prediction, with a Real Estate Illustration

Abstract

Drawing useful predictions from vast accumulations of data is becoming critical to the success of an enterprise. Organizations’ databases grow exponentially from transactions with external stakeholders in addition to their own internal activities. An important organizational computing issue is that, as they grow, the databases become potentially more valuable and also more difficult to analyze. One example is predicting the value of residential real estate based on past comparable sales transactions. This is critical to several important sectors of the US economy including the mortgage finance industry and local governments that collect property taxes. The common methodology for dealing with such property valuation is based on multiple regression, although this methodology has been found to be deficient. Data mining methods have been proposed and tested as an alternative, but the results are very mixed. This article introduces a novel approach for improving predictions using an adaptive, neuro-fuzzy inference model, and illustrates its application to real estate property price prediction through the use of comparable properties. Although neuro-fuzzy–based approaches have been found to be effective for classification and estimation in many fields, there is very little existing work that investigates their potential in a real estate context. In addition, this article addresses several common problems in existing studies, such as small sample size, lack of rigorous data sampling, and poor model validation and testing. Our model is tested with real sales data from the assessment office in a large US city. The results show that the neuro-fuzzy model is superior in all of the test scenarios. The article also discusses and refines a unique technique to defining comparable properties to improve accuracy. Test results show very promising potential for this technique in mass appraisal in real estate and similar contexts when used with the neuro-fuzzy model.