HKUST Business School Magazine

Biz@HKUST Biz@HKUST 32 33 // Cover // Insight Sentiment of next content Sentiment of current content Negative Neutral Positive Negative 42.39% 40.52% 17.09% Neutral 32.09% 47.77% 20.14% Positive 29.30% 43.11% 27.60% Variable Model Sentiment Index 0.0481*** (0.0036) User fixed effects Yes Team x day fixed effects Yes Controls Yes Observations 3,608,654 R-squared 0.111 Adjusted R-Squared 0.095 FinSent – An Interactive Sentiment Analysis Dashboard Based on FinBERT The Center has been developing a first-of-its-kind sentiment analysis dashboard based on FinBERT, a BERT model that outperforms other language models in the sentiment classification of financial texts. Aptly named FinSent, the fully automated dashboard visualizes sentiments of financial texts using cutting- edge natural language processing technologies, enabling users to gain insights into the financial sentiments of publicly traded companies through charts, graphs, and color- coded sentiment scores. FinSent comprises an automated module that collects and processes corporate 10-K and 10-Q filings and earnings call transcripts of public companies on a daily basis. The sentiments towards the companies are presented via an interactive dashboard that allows users to freely customize the content. The following diagram shows an example of the FinSent interface: Spillover of Social Media Sentiments In recent decades, the rise of social media has aided the free flow of public information. It has also become a reliable platform for people around the world to interact and share their opinions in an affordable, anonymous, and uncensored manner. While countless benefits have been accrued from the use of social media, some complicated and worrying challenges have surfaced in recent years, catching the attention of researchers, policymakers, and members of the public. It is necessary to understand the difference between real-life and online communications, and how this affects our behavior and mental health. To address this issue, we study the spillover of social media sentiments in the English Premier League (EPL) soccer community using data from Reddit. We apply sentiment analysis with a machine learning model to all the contents in our dataset to capture the sentiment of the contents. The model outputs the likelihood of those contents being positive, neutral, or negative. We trace the contents created by each author and provide evidence of how their next sentiments will change from the current one with a transition matrix. The transition matrix shows that sentiments of the next content differ depending on their current sentiments. Users will more likely produce negative sentiment contents when they generate negative contents (42.39%). On the other hand, sentiments of the next contents are more likely positive when they produce positive sentiment contents (27.60%) rather than negative ones (17.09%). Thus, we consider the sentiment of the current content as one of the controls when we estimate the model. With 3.6 million pieces of content from 20 EPL subreddits in the 2017/2018 season, we run a regression model to check whether the sentiments of what users see would affect the sentiment of their next piece of content. The dependent variable of the model is the difference between the likelihood for the content to be positive and negative. A positive index means the content is more likely to be positive. Similarly, a negative index means the content is more likely to be negative. The main independent variable is a sentiment index, which measures the sentiments of contents that users see between their posts. Furthermore, we control various confounding factors, including the sentiment of their previous posts. We find strong evidence of spillover effects from the sentiments. The estimate of the sentiment index (0.0481) is significant and positive. It means that emotions on social media are contagious and can easily intensify. Social media users should recognize this spillover to their emotions and avoid being manipulated by other social media users. Learning how to deal with an increasingly polarized online environment and offensive conversations is important if we are to be responsible netizens. Notes. Robust standard errors clustered by user are in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01. The Center tackles complex business and social problems by listening to what the crowd actually says