Machine Learning for Real Estate Time Series Prediction

Several researchers have demonstrated that real estate investments have improved the risk-adjusted performance of mixed-asset portfolios belonging to institutional investors. In order for these portfolio strategies to be more effective, one could use price predictions (instead of historical data) to optimize weights. The goal of this paper is to investigate the predictive performance on price time series of REITs (real estate investment trusts), stocks and bonds, of five different machine learning (ML) algorithms. These algorithms are: linear regression; support vector regression; gradient boosting; long short-term memory neural networks; and k-nearest neighbour. We run experiments on 90 datasets and compare the ML results to those of an ARIMA model, which is a popular econometric benchmark used in financial time series predictions. Our results show that machine learning algorithms statistically outperform ARIMA. In addition, we find that all machine learning algorithms are able to produce very low root mean square errors, with linear regression and long short-term memory obtaining the lowest error values.