Property Valuation in the Age of Machine Learning: Enhancing Public Assessments Through Predictive Modelling

Abstract

In light of public criticism and technical limitations surrounding Denmark’s current property valuation system, this thesis explores whether machine learning techniques can offer a viable and improved alternative to the valuation model used by the Danish Tax Agency (SKAT). Drawing on publicly available data, we simulate SKAT’s current approach to both understand and recreate SKAT’s valuation method. We also develop regression-based machine learning models, including Linear Regression, Decision Trees, Random Forests, Gradient Boosting, and K-Nearest Neighbors designed to explore whether modern machine learning can improve upon the current public valuation approach. Our methodology follows the CRISP-DM framework and applies established preprocessing and feature engineering techniques to ensure a robust evaluation. We find that machine learning models consistently outperform the public valuations in terms of predictive accuracy across key evaluation metrics. In addition to performance, the thesis critically assesses interpretability, transparency, and regulatory compatibility, identifying both the strengths and the constraints of adopting machine learning in public-sector decision-making. We find that machine-learning methods have the potential to make property taxation more consistent and accurate. Based on our findings, we outline how such approaches could complement or improve the current system in practical ways.