Appraisal of Apartments in Belgrade Using Hedonic Regression: Model Specification, Predictive Performance, Suitability for Mass Appraisal, and Comparison with Machine Learning Methods

We present a Bayesian hedonic regression model for the appraisal of apartments located within the metropolitan area of Belgrade, Serbia. Data on 12.904 apartments were collected from a local classified website, and used to fit and validate the model. The response variable, which is the price per m 2 in euros, is assumed to be log-normally distributed. Nested random effects are used to model the hierarchical structure present in the location identifiers, and thin-plate spline functions are used to capture the nonlinear effects. Location explains around 78.62% of the total variation in advertised prices per m 2 in the training set, confirming its paramount importance. Further major factors affecting prices are: area in m 2 , floor number, the total number of floors in the building, the availability of an elevator and condition. The predictive functionality of the model is demonstrated through an open access online application that accompanies this study. The model achieves Mean Average Percent Error (MAPE) of 13.06% in the validation set. Its predictive performance is compared to that of three popular machine learning (ML) methods, and its suitability for mass appraisal is examined. This is the first empirical study to present the results obtained from a hedonic regression built using data pertaining to Belgrade.