Spatial Determinants of Real Estate Appraisals in The Netherlands: A Machine Learning Approach

Abstract

With the rapidly increasing house prices in the Netherlands, there is a growing need for more localised value predictions for mortgage collaterals within the financial sector. Many existing studies focus on modelling house prices for an individual city; however, these models are often not interesting for mortgage lenders with assets spread out all over the country. That is why, with the current abundance of national geospatial datasets, this paper implements and compares three hedonic pricing models (linear regression, geographically weighted regression, and extreme gradient boosting—XGBoost) to model real estate appraisals values for five large municipalities in different parts of the Netherlands. The appraisal values used to train the model are provided by Stater N.V., which is the largest mortgage service provider in the Netherlands. Out of the three implemented models, the XGBoost model has the highest accuracy. XGBoost can explain 83% of the variance with an RMSE of €65,312, an MAE of €43,625, and an MAPE of 6.35% across the five municipalities. The two most important variables in the model are the total living area and taxation value, which were taken from publicly available datasets. Furthermore, a comparison is made between indexation and XGBoost, which shows that the XGBoost model is able to more accurately predict the appraisal values of different types of houses. The remaining unexplained variance is most probably caused by the lack of good indicators for the condition of the house. Overall, this paper highlights the benefits of open geospatial datasets to build a national real estate appraisal model.

Keywords

• real estate values modelling
• housing market
• housing price
• real estate appraisals
• hedonic model
• extreme gradient boosting