“Our study examines the overall ‘vibe’ or sentiment conveyed in the speech”
What is the focus of the paper?
The article studies whether the cues conveyed in the voice of CEOs at S&P 500 companies can predict future stock returns from a few days to several months ahead. The authors analyze short segments of the opening remarks and Q&A sessions from quarterly earnings calls using two deep learning models: TRILLsson and W2V2. The models provide a set of vocal characteristics, such as how dominant or emotionally aroused a CEO sounds. The authors examine how such signals relate to future stock returns and compare these results with predictions based on traditional macroeconomic models.
What are the key findings?
The most notable result is that the W2V2 model can predict stock returns better than chance (54–59%), while TRILLsson shows no predictive power. W2V2 uses two different methods: a complex method that is difficult to interpret directly and a simpler method that analyzes emotions. Only the complex method works well for the analysis of the CEO’s opening remarks; in the analysis of the Q&A session, both methods deliver predictions better than chance. In the method that directly assesses emotions, arousal and dominance are followed by positive returns, while valence is followed by negative returns. Finally, only the complex method increases the predictive power when vocal characteristics are added to a macroeconomic model.
What are the implications?
- Which deep learning model is used—and for models with multiple methods, which method is chosen— is crucial for predictive power.
- Which segments of the opening remarks or the Q&A session are used for extracting voice characteristics influences the findings.
- Models that include macroeconomic variables as well as the complex, multidimensional method yield better predictions than models based solely on financial data. This underscores the importance of further research into the role of voice characteristics in the prediction of stock returns.