When can social media lead financial markets?

Social media analytics is showing promise for the prediction of financial markets. The research presented here employs linear regression analysis and information theory analysis techniques to measure the extent to which social media data is a predictor of the future returns of stock-exchange traded financial assets. Two hypotheses are proposed which investigate if the measurement of social media data in real-time can be used to pre-empt – or lead – changes in the prices of financial markets. Using Twitter as the social media data source, this study firstly investigates if geographically-filtered Tweets can lead the returns of UK and US stock indices. Next, the study considers if string-filtered Tweets can lead the returns of currency pairs and the securities of individual publically-traded companies. The study evaluates Tweet message sentiments – mathematical quantifications of text strings’ moods – and Tweet message volumes. A sentiment classification system specifically designed and validated in literature to accurately rank social media’s colloquial vernacular is employed. This research builds on previous studies which either use sentiment analysis techniques not geared for such text, or which instead only consider social media message volumes. Stringent tests for statistical-significance are employed. Tweets on twenty-eight financial instruments were collected over three months – a period chosen to minimise the effect of the economic cycle in the time-series whilst encapsulating a range of market conditions, and during which no major product changes were made to Twitter. The study shows that Tweet message sentiments contain lead-time information about the future returns of twelve of these securities, in excess of what is achievable via the analysis of Twitter message volumes. The study’s results are found to be robust against modification in analysis parameters, and that additional insight about market returns can be gained from social media data sentiment analytics under particular parameter variations