| This research investigates both the individual and combined predictive capability of two investor sentiment indicators; one extrapolated from social media, text based, and one extrapolated from derivative data, market data based. Our findings show: 1) both microblogging message volume and sentiment can be used as features to predict continuation and reversal effects in asset prices; 2) specific market participant option trading volume is shown to be a predecessor to asset price movements; 3) short positions from specific market participants are a proxy of future performance; 4) combining both textual and market data features improves overall model performance. A significant contribution of this research to existing literature is made through the aggregation of two main sources of measurable sentiment, social media and market data. In addition, this research adjusts returns for risk, momentum and actual transaction costs (as a function of shares bought and sold) to properly capture a more realistic alpha. We use a predefined number of stocks (not company specific) which allows for a more practical approach that confines the number of daily positions to a reasonable count versus a large number that a quantile count would yield. We make no assumption that firms have unlimited capital or the means to invest in hundreds of stocks daily. Another contribution of our research is the use of a more recent data set that includes pre, during and post financial crisis, bringing us through varying market conditions. Such volatile market conditions (financial crisis) were not tested in previous research. The findings of this research also indicate investor overreaction to significant changes in crowd-sourced negative sentiment. |
