Sentiment Analysis Within a Deep Learning Probabilistic Framework – New Evidence from Residential Real Estate in the United States

Abstract

This paper is devoted to the relationship between news sentiment and changes in housing market movements. It provides a novel and straightforward approach to account for heterogeneous expectations of market actors within a probabilistic framework utilizing machine learning. Our novel sentiment index shows a persistent and statistically significant explanatory power for the prediction of the housing market, in contrast to common dictionary approaches. This holds for news headlines and abstracts and different definitions of sentiment indices. Our results can be regarded as the first sentiment-based evidence of heterogeneous actors in the housing market and underline the importance of different expectations for measuring non-fundamental drivers.

Keywords:

• Sentiment analysis
• neural networks
• housing market
• machine learning

JEL CLASSIFICATION:

• G40
• R310

Acknowledgements

I gratefully acknowledge many useful comments and discussions of the participants at the 37th ARES Conference in March 2021. I would like to thank Simon Stevenson, Old Dominion University and the Lucas Institute for Real Estate Development and Finance at Florida Gulf Coast University for honoring this paper with the manuscript prize in the category “housing”. Furthermore, I would like to thank Jeremy Gabe, University of San Diego and the James R. Webb ARES Foundation, for honoring this paper with the 2021 doctoral program manuscript prize.

Declaration of Interest Statement

The paper has not been published and is not under consideration by any other journal. There are no conflicts of interest to disclose. I approved the manuscript and this submission.

CRediT Author Statement

Cathrine Nagl: Conceptualization – Methodology – Software – Validation - Formal analysis

Data Curation - Writing - Original Draft -Writing - Review & Editing

Notes

1 The number of neurons of the two hidden layers is set to 512 and 256. The ANN is conducted with a gradient descent algorithm. The learning rate is 0.0001 and the dropout rate of 30% avoids overfitting. The batch size is set to 512. Although the model was set to train 100 epochs, early stopping allows 8-18 epochs.