The explanatory power of a topic-weighted measure of sentiment: The application of topic modeling to sentiment analytics

Details

Supervisors

Thewissen, James

Faculty

Louvain School of Management

Degree label

Master [120] : ingénieur de gestion, à finalité spécialisée

Abstract

The literature examining the qualitative information of corporate disclosures shows that net sentiment – the difference between the number of positive and negative words – has explanatory power for the market reaction around the publication of financial disclosures and future stock performance. However, there is only limited work focusing on finding better proxies than equally weighted net sentiment for the tone of financial disclosures. The work that does, generally ignores the context in which positively and negatively connoted words occur. I address this research gap by using topic modeling to uncover 30 themes discussed in a sample of 60,662 earnings press releases of S&P 1500 companies for the period of 2004 to 2015. I find that there are significant differences in the importance of the discussed topics for investors and derive a weighting scheme reflecting this. The results indicate that a topic-weighted sentiment proxy possesses a significantly higher information value to predict investor reaction and future firm performance than its equally weighted counterpart. These results are insensitive to the wordlist selection but sensitive to the prespecified number of topics used to run the topic modeling algorithm. The information value of the topic-weighted proxy increases with the level of information asymmetry between a firm and its investors. Thus, a topic-weighted sentiment proxy should be used to get more accurate predictions of the market reaction and future firm performance.