Social Media and Real Estate: Do Twitter Users Predict REIT Performance?

Abstract

This study investigates the impact of social media sentiment on indirect real estate market returns by utilizing a comprehensive natural language processing approach to identify relevant Twitter posts and extract sentiment from them. To handle the complex linguistic features inherent in social media messages, three different sentiment classifiers are compared. The findings suggest a significant relationship between monthly sentiment and REIT returns, which occurs in two phases: a short-term speculative reaction and a greater longer-term reaction related to actual changes in the real estate market. The study also highlights that while the conventional dictionary approach can identify this relationship, more sophisticated classifiers can achieve higher accuracy. Overall, the results demonstrate the valuable insights that can be gained from analyzing social media data and its potential impact on the real estate market.

Keywords:

• Social media sentiment
• textual analysis
• Real Estate Investment Trust
• machine learning

Disclosure Statement

No potential conflict of interest was reported by the author(s).

Notes

1 Matrix factorization techniques are used to decompose large matrices that contain statistical information about a text corpus, resulting in low-dimensional word representations and reduced computation time. On the other hand, context window methods use linear relationships between word vectors based on local context windows to predict linguistic patterns and improve word analogy tasks.

2 It was determined that filtering for the 25 closest words was optimal as a significant decline in the quality of the word representations occurred beyond this threshold.

3 The training dataset utilized in this study is sentiment140 and is classified by Go et al. (2009) using distant supervision. This balanced dataset contains 1.6 million tweets labeled as either positive or negative.

4 The same training data set and pre-processing steps as for the SVM have been applied.

5 Since the output is again probabilistic, with P indicating the probability of an input having positive sentiment, the same classification ranges as for the SVM have been applied.

6 Due to the feedback loop, RNNs encounter the challenge of vanishing or exploding gradients. As the weights are applied repeatedly to the values in the feedback loop, the magnitude of the input values either decreases (weights <1) or increases (weights >1) with every iteration, potentially leading to the input values either dropping to zero or increasing to infinity after a sufficient number of iterations. Vanishing or exploding gradients, however, impede the ability of the network to converge toward optimal weights and biases (Hochreiter, 1991).

7 For the SVM, only 25% of the full training dataset is used for training and testing, due to computational limitations.

8 The list of signal words can be found in Appendix 1.

9 The model accuracy of the SVM and LSTM increased from 76.22%, respectively, 78.02% when stop words were excluded, which confirms our assumptions described in the data cleaning section. Further note that the accuracy of the LSTM is given by the validation process after the final step of the training process in contrast to a separate testing as conducted for the SVM.

10 Since it is impossible to measure the accuracy of our unlabeled corpus, we compared the percentage of tweets that were classified identically by the models. The LSTM model and dictionary shared 40.44% similarity, the dictionary and SVM shared 48.80%, and LSTM and SVM shared 61.03%. For the dictionary approach, significantly more tweets have been classified neutral due to the limited size of the word lists.