Do Tweet Sentiments Still Predict the Stock Market?

Abstract

In this paper, we document a strong predictive relationship between US equity markets as proxied by the capitalization weighted index excess returns and it’s corresponding tweet sentiments. We employ StockTwits' labels of "Bullish" and "Bearish" to train several standard machine learning classification algorithms. Armed with our sentiment engines, we document that Support Vector Machines (SVM) generates the most useful predictive sentiments vis-à-vis those generated by Single Layer Perceptron (SLP) and Naïve Bayes (NB). Tweet sentiments are strongly positively contemporaneously correlated with daily market returns, additionally we find a weaker negatively correlation with next day's market returns.

We report statistically significant evidence that the market returns Granger-cause next day's sentiment movements. Moreover, in the most recent period of 2015, our rolling analysis shows significant evidence that the tweet sentiments actually Granger-caused the market to move! While a fascinating discovery, we temper our enthusiasm, as the market appears to have quickly and efficiently digested the ability to convert tweets into sentiment and in turn predict market movements. Our results support the notion of a highly efficient market that has rapidly digested and processed such tweets sentiment data.

Keywords: Tweet, Sentiment, Market Efficiency, Stock Markets