Abstract/Summary

This thesis deals with the extraction, construction and analysis of commercial real estate (CRE) sentiment within Europe and the U.K. especially. The three empirical studies in this thesis may contribute to our understanding of the discipline. As I establish in the literature review, the analysis of commercial real estate sentiment still offers a lot of potential for further research. Since real estate markets are subject to sentiment swings, scholars and market participants should consider them in their market analysis. The first study establishes the need for sentiment consideration within the European real estate market. In order to justify the research of sentiment analysis, I have used different indirect and direct sentiment proxies and applied them in yield models for 80 different commercial property (sub-)markets within Europe. The statistical modification of different sentiment proxies is needed since not all European property markets offer direct sentiment measures. The results suggest, that the consideration of sentiment in a yield model framework adds significant information. I found, that CRE markets, which are assumed to be more liquid and developed, show a larger exposure to property specific sentiment measures. Markets, which are assumed to be less developed (i.e. Eastern European markets) on the other hand, have a larger exposure to more general macroeconomic sentiment indicators. The second study introduces a new method, which can be used to extract sentiment from text documents. The primary motivation for the use of text documents and the application of Natural Language Processing (NLP) methods lies in the fact that these documents are published much faster than other sentiment proxies. This allows extracting a much more accurate market sentiment. The second study should be understood as an introductory chapter to the method and the field of NLP. In total four different wordlists (AFINN, BING, NRC and TM) are used to extract the sentiment form various market reports for the CRE market in U.K. The study reveals that sentiment extracted from those documents, can be used to improve autocorrelated models. The last study uses those findings and applies different supervised learning methods. While the second study has produced sufficient results, the underlying text corpus of market reports has shown a series of insufficiencies. I have therefore, used a large dataset of more than 120,000 news articles, all concerning the British CRE market. Findings suggest, that the main issue of supervised learning algorithms is the appropriate classification of the different entities. I offer two approaches in order to construct robust sentiment indicators.