Out-of-Sample House Price Prediction by Hedonic Price Models and Machine Learning Algorithms

In an illiquid market like the real estate market, market values are not readily available. Transactions are scarce and do not always reflect market value. As a consequence, appraisal values play an important role to inform agents in decision making, financial reporting and for property taxes. For example, appraisal values are used for property investment decisions and for providing mortgage loans. In a recent report De Nederlandsche Bank raises concerns about the quality and independency of appraisal values (Van der Molen and Nijskens, 2019). The authors show that one third of all appraisal values exactly match the transaction price, and in almost 60% the appraisal value is higher than the transaction price. Automated valuation models (AVMs) are less prone to potential client influence. However, in order to be accepted by a broad audience, AVMs need to be transparent, robust, explainable and they need to provide reliable predictions. In this research we address these issues. We compare traditional hedonic price models to more advanced machine learning algorithms and analyse the accuracy of out-of-sample predictions and variable importance. The research is based on almost all residential transaction prices in the Netherlands in 2017.