Mining the emotional information in the audio of earnings conference calls : A deep learning approach for sentiment analysis of securities analysts' follow-up behavior

Abstract

In this paper, we propose a deep learning approach to extract emotional information from the audio of earnings conference calls and empirically examine the influences of these emotional variables on securities analysts' follow-up behavior. Our findings suggest that, in the statement section, positive emotional information tended to positively influence the analysts' willingness to issue rating reports, while the inverse was true for negative emotional information; non-negative emotional information in the question section had a positive influence, while negative emotional information in the response section had a negative influence. Secondly, for the specific rating of the issued reports, negative emotional information in the response section tended to result in a lower rating, and neutral emotional information might also have caused a lower rating. Thirdly, in terms of rating adjustments, non-negative emotional information in the question section tended to cause an upgrade revision, while the inverse was true for the negative emotional information in this section. Positive emotional information in the response section also caused an upgrade revision. The approach we proposed provides new insight for understanding analysts' follow-up behavior and offers practical implications for analysts, management, investors, and regulators.