Sentiment Analysis in The Norwegian Housing Market: Evaluating the inferential and predictive power of sentiment scores on housing price using linear modelling and machine learning

Abstract

In this thesis, we investigate how information and sentiment provided through news media affect prices in the Norwegian housing market. Our analysis is based on news articles from selected Norwegian news outlets, transaction data from the housing market in Oslo and macroeconomic data. We derive sentiment values from the news articles using a recurrent neural network algorithm. We infer on the data using an OLS regression model and study the predictive ability of sentiment using XgBoost on models with and without sentiment data. We observe that the variation in measured sentiment values explains almost half the variation in the housing price index for Oslo. This suggests that people respond to the information provided in the newspapers, and that the price development is not a random walk. Further, we observe that the sentiment coefficient is significant both in statistical and economic terms after we control for fundamentals, suggesting that people react to sentiment more excessively than what is justified by the fundamentals. The implication is that the housing market is not fully efficient. This is supported by data showing that an increase in sentiment values also widens the difference between asking price and final price. With the introduction of the XgBoost model, we decrease predictive error present in linear regression predictive benchmark by 14.96 percent. Our best sentiment model causes a decrease in prediction error of 2.52 percent relative to the reference model. This leads us to conclude that both fundamental information and sentiment is associated with price developments in the Norwegian housing market.