Machine Learning for Mass Valuation of Residential Real Estate

Mass valuation is a tool necessary to determine the cadastral value and form the tax base of the regions. At the same time, the state cadastral valuation methods currently used in Russia are based on traditional statistical models without considering the opportunities provided by state-of-the-art information technologies. The first part of this study gives a brief overview of modern approaches to mass valuation of urbanized territories with GIS technologies and artificial intelligence. The main part of work is devoted to demonstrating how the mass valuation quality can be improved through innovative machine learning technology by the example of residential real estate in Rostov-on-Don (Russia). For this purpose, LGBoost (Light Gradient Boosted Machine), XGBoost (eXtreme Gradient Boosting) and CatBoost software tools are used. The dependent variable is the price of one square meter of total floor area, independent variables include kitchen area, total floor area, type of the floor (first, last, others), number of floors in the house, year when house was built, type of the real estate (multi-storey new building, secondary housing), the house facade material, city district and suburb. The classical regression model accuracy characterized by the coefficient of determination is 0.68, while advanced machine learning models give the accuracy of 0.85. Thus, machine learning models can be recommended as a practical toolkit to improve the quality of cadastral valuation of residential real estate.