News-based sentiment analysis in real estate: a machine learning approach

ABSTRACT

This paper examines the relationship between news-based sentiment, captured through a machine learning approach, and the US securitised and direct commercial real estate markets. Thus, we contribute to the literature on text-based sentiment analysis in real estate by creating and testing various sentiment measures by utilising trained support vector networks. Using a vector autoregressive framework, we find the constructed sentiment indicators to predict the total returns of both markets. The results show a leading relationship of our sentiment, even after controlling for macroeconomic factors and other established sentiment proxies. Furthermore, empirical evidence suggests a shorter response time of the indirect market in relation to the direct one. The findings make a valuable contribution to real estate research and industry participants, as we demonstrate the successful application of a sentiment-creation procedure that enables short and flexible aggregation periods. To the best of our knowledge, this is the first study to apply a machine learning approach to capture textual sentiment relevant to US real estate markets.

KEYWORDS:

• Textual analysis
• News-based sentiment
• Machine learning
• US commercial real estate
• Support vector machine

Acknowledgements

The authors would like to acknowledge the support of the S&P Global Market Intelligence database as provider of the news dataset. The authors would also like to thank the four anonymous referees and the 2017 ARES and 2017 ERES conference participants for their valuable comments as well as Brian Bloch for his language support.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. Controlling for the CPI did not substantially alter the results in terms of sign, size and significance of the sentiment indicators and autoregressive components when running the direct commercial real estate market models. As we adapted the controls to better reflect markets without losing to many degrees of freedom we report regression equations without the CPI.

2. Note that only the remaining headlines are used to calculate the sentiment indicators afterwards. This makes sure that algorithm ‘tuning’ does not influence classification results.

3. For ease of reading, we stick to the common notation of matrices using bold characters.

4. Because it is mathematically more convenient, the optimal hyperplane can be derived by minimising $0,5\mathit{w}\ast \mathit{w}$ subject to $y_i\left(\mathit{w}{x}_i+b\right)-1\ge 0,i=1,\dots ,l$.

5. Note that our results do not automatically indicate that market participants are ignoring past market performance in terms of their sentiment about the future or that past market performance is not relevant for our constructed sentiment indicators at all. As our text corpus does not only contain news about past market movements, but also many other possible aspects concerning the real estate industry, past market performance is most likely only one factor driving sentiment indicator changes. Furthermore, different news might incorporate different levels of textual sentiment, are reported at different frequencies and can be forward- or backward-looking. Hence, this heterogeneity might be the reason why our models do not capture a statistically significant relationship between (pure) market performance in the past and future indicator changes.

6. Note that when using the OI and the SQ instead of the PI in Models 4 and 5 of Table 5, the OI is still insignificant and the SQ drives returns one month ahead similar to the findings of Table 4.

7. Note that when further controlling for lagged returns of the securitised real estate market, the findings of Table 6 do not change.

8. When substituting the PI by the OI or SQ in Table 7, the results of Table 6 with respect to the respective significance of the OI and SQ still hold.

Additional information

Notes on contributors

Jochen Hausler

Jochen Hausler is a research associate and doctoral candidate at the Honorary Professorship of Real Estate Development at IRE|BS International Real Estate Business School, University of Regensburg. He holds a Master’s degree in Real Estate from the University of Regensburg as well as in Real Estate Finance from the Henley Business School, University of Reading, graduating from both universities in 2016. His research interests are focused mainly on the application of machine learning and artificial intelligence in real estate.

Jessica Ruscheinsky

Jessica Ruscheinsky is a doctoral candidate at the Chair of Real Estate Management at the International Real Estate Business School (IRE|BS), University of Regensburg. In 2014, she graduated from the MSc Real Estate programme at the Henley Business School, University of Reading as well as from the IRE|BS. At ERES 2016, she and Marcel Lang won the Award for the Best Paper in the PhD Session. Jessica Ruscheinsky works part-time for a German-wide acting real estate development company.

Marcel Lang

Marcel Lang is a doctoral candidate at the Chair of Real Estate Management at the IRE|BS International Real Estate Business School, University of Regensburg. In 2015, he graduated with a MSc in Real Estate at the IRE|BS International Real Estate Business School, University of Regensburg. Together with Jessica Ruscheinsky he helds the 2016 ERES PhD award. In addition to his scholarly work, Marcel Lang gained practical experiences in the area of real estate transactions and works currently part-time for a real estate development company.