Topic Tones of Analyst Reports and Stock Returns: A Deep Learning Approach
41 Pages Posted: 20 Aug 2018 Last revised: 23 Dec 2020
Date Written: August 8, 2018
We disassemble the text of analyst reports into multiple pieces that represent three types of topics - opinion type (overall assessments of analysts), corporate fact type (helping investors to better digest corporate facts) and justification type (explaining quantitative numbers). We extract the tone for the text of each topic via a deep neural network supervised learning methodology. A baseline model without using text information has an adjusted R squared of 2:3% in predicting the cumulative twoday abnormal returns. When we include the topic tones, the adjusted R squared increases to 15:4%. This significant increase of R squared is not much driven by the justification type topics since it is sort of redundant given the quantitative numbers (e.g., earnings forecasts). In contrast, the opinion type and the corporate fact type topics provide substantial information beyond the quantitative numbers and are the main drivers of the significant increase of R squared.
Keywords: Textual Analysis; DNN Approach; Topic Tones; Information Content
JEL Classification: C89, G11, G12, G14
Suggested Citation: Suggested Citation