Abstract
We examine managerial behavior during conference calls vis-à-vis the informational impact on firm stock prices. Implementing an unsupervised machine learning algorithm, we document statistically and economically meaningful relationships between impromptu soft information divulged during calls and stock prices. Managers who choose to divulge more impromptu soft information in the Q&A session experience improved liquidity and less volatility, accompanied with lower abnormal returns. Conditioned on earnings surprise, large positive-surprise impromptu soft information results in larger returns and positive drift, indicating that investors do not completely trust the positive signal initially, with confidence growing over time evidenced as prices adjust to equilibrium.
Acknowledgments
I thank Jim Brau, Charles Cao, Jeremiah Green, Matt Gustafson, David Haushalter, Peter Iliev, Jason Kotter, Jed Neilson, Giang Nguyen, Tim Simin, Alexey Zhdanov, and participants of the Pennsylvania State University Finance Department seminar series for constructive feedback and insights. Contact: 1155 Union Circle #305339, Denton Texas, 76203. Stephen.Owen@unt. 940.369.7202
Notes
1 We adopt the verbiage of works such as Brau, Cicon, and McQueen (Citation2016) and Liberti and Petersen (Citation2019) in the classification of textual information as soft information.
2 Unsupervised means we estimate the topic distributions without imposing any priors or external bias, e.g., a dictionary of words. Therefore, the model learns from the data and estimates the relationships between words and phrases to form topics. Those topics are then used as the basis to calculate the similarity between the prepared remark section and Q&A section of an earnings call, returning a model of topical similarity between the two corpora.
3 Machine learning techniques have gained traction in various fields of study, as well as the use of different embedding models in economic analyses (see Hanley and Hoberg Citation2010, Citation2012; Bollen et al. Citation2011; Cicon et al. Citation2012; Netzer et al. Citation2012, Citation2019; Tirunillai and Tellis 2012, Citation2014; Akansu et al. Citation2017; Brau et al. Citation2022).
4 Refer to Appendix B for a more detailed explanation of the process for fitting and estimating the topic modelling algorithm.