We have built a functioning Proof of Concept for the Appreciate the Object project (AtO), an App which uses Machine Learning algorithms to predict the price of a house in the Netherlands both at the current time and forward into the future. The overall accuracy of the AtO is already roughly parallel to an unassisted human real estate agent (16% Median Absolute Percentage Error) using just the internal area of the property, the property type (apartment etc) and information on the surrounding neighbourhood. Importantly this model also estimates the uncertainty on its predictions, allowing users to better understand the range of values their house might be expected to sell at.
The AtO Apps house price estimation model makes use of the cutting edge XGBOOST machine learning algorithm which was trained on only ~10% of the sales data available for 2017, so substantial improvements in accuracy can be expected with the full sample size, as well as further gains from the inclusion of internal property state data (years since renovations, room numbers etc) and other external variables (GDP, inflation, unemployment etc). The current version shows a slight bias towards overestimating low value properties and underestimating high-end properties but this is expected to improve substantially with the inclusion of the full data set.
AtO is deployed on a parallelised, distributed and completely scalable cluster of servers and is supplied with new data as soon as it becomes available, by a completely automated data ingest pipeline.
Finally the POC AtO front-end allows for the evaluation and comparison of any existing property registered with the Netherlands Property Registry; BAG. This front-end also allows for the displaying of relevant neighbourhood and property data alongside the identification of top features in the surrounding area.
Next steps include user testing and further development of the front-end UI, sourcing of a larger dataset of house-sales data and further model tuning to make the predictions truly competitive. Finally a partnership with existing real estate agents should be arranged in order to source high-quality internal house state data for inclusion in the model. By undertaking these steps we aim to provide a highly usable, accurate house price prediction service with true estimations of the predictions uncertainty, a key feature missing in many similar products on both the domestic and international market.
This project is supported by the Europees Fonds voor Regionale Ontwikkeling