QUANTIFYING THE IMPACT OF LOCATION DATA FOR REAL ESTATE APPRAISAL – A GIS-BASED DEEP LEARNING APPROACH

Appraising real estate is challenging since a property\u27s characteristics, location, and neighborhood make up for substantial shares of its price. Current Computer-Assisted Mass Appraisal (CAMA) approaches rely on hedonic models and structured real estate data. Recent work proposed to include geospatial or image data to improve CAMA’s predictive accuracy. However, the value of these data sources and their preparation costs remain unclear. In this study, we use Geographic Information Systems (GIS) and Convolutional Neural Networks (CNNs) to appraise 71,000 properties in Philadelphia. We find that geospatial data requires substantial preprocessing but increases CAMA’s predictive accuracy by 12%, while image data come with higher computational complexity and an improvement of 16%. Using both datatypes is complex to optimize but offers increased robustness towards extreme outliers. Beyond real estate appraisal, our study is one of the first to quantify the value of geospatial data for data-driven services in the information systems discipline