Evaluating the impact of social-media on sales forecasting: a quantitative study of world's biggest brands using Twitter, Facebook and Google Trends

In the world of digital communication, data from online sources such as social networks might provide additional information about changing consumer interest and significantly improve the accuracy of forecasting models. In this thesis I investigate whether information from Twitter, Facebook and Google Trends have the ability to improve daily sales forecasts for companies with respect to the forecasts from transactional sales data only. My original contribution to this domain, exposed in the present thesis, consists in the following main steps: 1. Data collection. I collected Twitter, Facebook and Google Trends data for the period May 2013 May 2015 for 75 brands. Historical transactional sales data was supplied by Certona Corporation. 2. Sentiment analysis. I introduced a new sentiment classification approach based on combining the two standard techniques (lexicon-based and machine learning based). The proposed method outperforms the state-of-the-art approach by 7% in F-score. 3. Identification and classification of events. I proposed a framework for events detection and a robust method for clustering Twitter events into different types based on the shape of the Twitter volume and sentiment peaks. This approach allows to capture the varying dynamics of information propagation through the social network. I provide empirical evidence that it is possible to identify types of Twitter events that have significant power to predict spikes in sales. 4. Forecasting next day sales. I explored linear, non-linear and cointegrating relationships between sales and social-media variables for 18 brands and showed that social-media variables can improve daily sales forecasts for the majority of brands by capturing factors, such as consumer sentiment and brand perception. Moreover, I identified that social-media data without sales information, can be used to predict sales direction with the accuracy of 63%. The experts from the industry consider the results obtained in this thesis to be valuable and useful for decision making and for making strategic planning for the future.