A framework for Twitter events detection, differentiation and its application for retail brands

We propose a framework for Twitter events detection, differentiation and quantification of their significance for predicting spikes in sales. In previous approaches, the differentiation between Twitter events has mainly been done based on spatial, temporal or topic information. We suggest a novel approach that performs clustering of Twitter events based on their shapes (taking into account growth and relaxation signatures). Our study provides empirical evidence that through events differentiation based on their shape one can clearly identify clusters of Twitter events that contain more information about future sales than the non-clustered Twitter signal. We also propose a method for automatic identification of the optimum event window, solving a task of window selection, which is a common problem in the event study field. The framework described in this paper was tested on a large-scale dataset of 150 million Tweets and sales data of 75 brands, and can be applied to the analysis of time series from other domains.