Granger causality analysis between twitter sentiment and daily stock returns

Abstract

Textual data potentially carries information not found in quantitative data but is equally invaluable for financial analyses. This paper utilises sentiment analysis to examine the predictive value that Twitter posts (tweets) have on US equity returns. We assess the sentiment of tweets that mention specific firms by counting their use of positive and negative words as categorised in predefined word lists. Granger causality analysis was then conducted on 517 NASDAQ-100 and S&P 500 constituents by modelling a Panel Vector Autoregression process. We found that the market overreacts to both positive and negative tweets on the first trading day, but slightly corrects the shock by the second trading day. These findings are robust to different definitions of excess returns, and the use of different word lists for sentiment analysis. This suggests that Twitter sentiment does provide information useful for forecasting stock returns, albeit only for the short-term.