The time-varying nature of social media sentiments in modeling stock returns

The broad aim of this paper is to answer the following query: is the relationship between social media sentiments and stock returns time-varying? To provide a satisfactory response, a novel methodologya symbiosis of Bayesian Dynamic Linear Models and Seemingly Unrelated Regressions is introduced. Two sets of Dow Jones Industrial Average stock data and corresponding social media data from Yahoo! Finance stock message boards are used in a comprehensive empirical study. Some key findings are: (a) Affirmative response to the above question; (b) Models with only social media sentiments and market returns perform at least as well as models that include Fama-French and Momentum factors; (c) There are significant correlations between stocks, ranging from 0.8 to 0.6 in both data sets. Is the relationship between social media sentiments and stock returns time-varying? How does one capture the inherent cross-correlation between stocks to better model the time-varying relationship?Answers to the above queries are tackled via a novel methodology: Bayesian Dynamic Linear Models and Seemingly Unrelated Regressions.The impact of social media sentiments on future stock returns vary over time.The posterior distributions of correlations range from -0.8 to +0.6.The time-varying social media sentiments coefficients are more stable in 2011 when compared to 2009; the latter was a period of high turbulence due to recession.