Automated real estate valuation with machine learning models using property descriptions

Automated and accurate real estate valuation benefits buyers and sellers in real estate markets. So far, the literature on expert systems for real estate valuation has primarily focused on structured features like the age of the building or the number of rooms. The description of the property presents another rich source of information, which received comparably less attention. In this study, we evaluate several machine learning models in predicting real estate prices using different numeric representations of the property descriptions. Our empirical evaluation, based on rental apartments offers in Berlin (N = 30,218) and house purchase offers in Los Angeles (N = 33,610), shows that the best approach achieves mean absolute errors (MAE) of 1.01e monthly rent per square meter and 114.84$ per square foot, respectively. Including the property description into the best model reduces the MAE by up to 17.09 percent over the respective baseline models. In addition, we find that the benefit of including textual features of real estate descriptions only weakly depends on the description length. However, the benefit is comparatively less pronounced for rental apartment offers of low prices per square meter. We finally shed light on how the models arrive at decisions by visualizing description embeddings and presenting Shapley additive explanations.