Characteristics and predictability of Twitter sentiment series

In this paper we generate Twitter sentiment indices by analysing a stream of Twitter messages and categorising messages in terms of emoticons, pictorial representations of facial expressions in messages. Based on emoticons we generate daily indices. Then we explore the time-series properties of these indices by focusing on seasonal and cyclical patterns, persistence and conditional heteroscedasticity. In particular, we find significant day-of-the-week effect present in all indices, high persistence and significant degree of conditional heteroscedasticity. Then, using individual emoticon-based indices we generate an aggregate Twitter sentiment index and demonstrate that the index is good in capturing major world event such as major festive days and natural disasters. Using tests for linear and nonlinear Granger-causality we investigate whether the Twitter sentiment index contains extra information which could be used in a real-time prediction, but fail to detect any predictability at the moment. Our approach is inspired by two recent papers by Bollen et al. [2011] and O'Connor et al. [2010] who explore the relationship between Twitter sentiment and stock markets and economic indicators and find certain predictability. We significantly simplify the computational feasibility of the existing methodologies by experimenting with alternative sentiment classification processes, which may be the reason for reduced predictability of the index.