Deep Learning for Real Estate Price Prediction

Abstract

In this article, deep learning is applied to the task of real estate mass appraisal. To the best of our knowledge, we are the first to systematically evaluate a large collection of neural network architectures and tuning parameters for real estate price data. We compare the deep learning based approach to a classical linear regression model with manual feature engineering, gradient boosted trees, as well as a meta model which combines the prediction of the other models. Using transaction data for residential apartments in Switzerland, we find that a deep learning model results in significantly better predictive accuracy for real estate prices compared to a linear model. However, the difference is of a relatively small magnitude from an economic point of view. Further, the combined meta model results in substantially and significantly better predictions than each of the individual models.