Predicting housing prices with machine learning : a macroeconomic analysis of the Norwegian housing market

Abstract

This thesis explores the applicability of machine learning in macroeconomic housing price predictions in Norway. We apply three machine learning models Elastic Net, Random Forest and Neural Network on historical time-time series data and predict quarterly and yearly growth rates between 2013 and 2019. The performance is evaluated upon predictions from Norges Bank, DNB and SSB. Our results indicate that machine learning can produce predictions with the same accuracy as professional institutions. Among the machine learning models, Elastic Net produces the most accurate quarterly predictions. Compared to Norges Bank, Elastic Net’s predictions are more accurate in 29,6% of the quarters, but less precise in the overall evaluation. Large deviations during 2018 and 2019 are decisive for the lacking performance, after new mortgage regulations were introduced from Finanstilsynet. Random Forest predicts the most accurate yearly predictions but is outperformed by Norges Bank. Still, Random Forest surpasses both DNB and SSB throughout the evaluation process. The thesis contributes to the existing literature in several aspects. First, by outperforming housing experts, we challenge traditional macroeconomic approaches in the choice of predictive models. Second, our results indicate that linear models are more suited in shorter time spans, while nonlinear models perform better over longer horizons. Third, the machine learning models have identified household debt as the most influential variable to determine the housing prices in Norway. Overall, we believe machine learning approaches could become valuable in further academic and professional macroeconomic research. Keywords - Machine Learning, Prediction, Forecasting, Housing Market, Macroeconomics