The predictive power of tweets: an exploratory study

Abstract

We describe a system that estimates when an event is going to happen from a stream of microtexts on Twitter referring to that event. Using a Twitter archive of 60 known football events, this problem is transferred into a classiffcation problem. Different training procedures were followed, such as varying the training data and hierarchical classiffcation. The best performing method was on average 52.3 hours off, and especially the tweets that referred to an event that was still far away appeared to be hard to predict. Comparing the performance of the system to the performance of humans on the same task, it appeared that there is room for improvement for the system.