Machine learning for property valuation : an empirical study of how property price predictions can improve property tax estimations in Norway

Abstract

This thesis investigates whether machine learning methods can improve property price predictions, leading to more accurate property tax estimations in Norway. This study is important to ensure fair and trustworthy taxation for Norwegian taxpayers. The current method for predicting property values is a hedonic pricing model, developed by Statistics Norway using multiple linear regression. This model shows that 25% of all predicted property prices deviate by more than 20% of their observed price. These predictions are further used to estimate property tax, and the deviation in the current model suggests there is potential for improvement. The use of machine learning to improve property price predictions has yet to be explored by Statistics Norway. Consequently, this thesis investigates the predictive performance of more advanced machine learning methods on transacted properties, covering three districts in Oslo, from 2005 to 2020. These methodologies include decision trees, Random Forest, gradient boosting, and neural networks. All methodologies, except decision trees, performed better than multiple linear regression. Gradient boosting produced the best results, with an RMSE of 0.1140 compared to an RMSE of 0.2132 from the multiple linear regression. The total percentage of predictions deviating more than 20% of observed values were 6.4% using the gradient boosting approach, providing an improvement of 74% to the current method. The main conclusion drawn from this research confirms the superiority of machine learning methods for property valuation, capable of improving the current methods for estimating property tax in Norway. Additionally, the use of Local Interpretable Model-agnostic Explanations (LIME) can make the results transparent and compliant with current GDPR legislation for automated decisions. This thesis recommends the implementation of gradient boosting as the new method for property valuation in Norway. Keywords – Property tax, machine learning, LIME, GDPR, gradient boosting