Information extracted from social media streams has been leveraged to forecast the outcome of a large number of real-world events, from political elections to stock market fluctuations. An increasing amount of studies demonstrates how the analysis of social media conversations provides cheap access to the wisdom of the crowd. However, extents and contexts in which such forecasting power can be effectively leveraged are still unverified at least in a systematic way. It is also unclear how social-media-based predictions compare to those based on alternative information sources. To address these issues, here we develop a machine learning framework that leverages social media streams to automatically identify and predict the outcomes of soccer matches. We focus in particular on matches in which at least one of the possible outcomes is deemed as highly unlikely by professional bookmakers. We argue that sport events offer a systematic approach for testing the predictive power of social media, and allow to compare such power against the rigorous baselines set by external sources. Despite such strict baselines, our framework yields above 8% marginal profit when used to inform simple betting strategies. The system is based on real-time sentiment analysis and exploits data collected immediately before the games, allowing for informed bets. We discuss the rationale behind our approach, describe the learning framework, its prediction performance and the return it provides as compared to a set of betting strategies. To test our framework we use both historical Twitter data from the 2014 FIFA World Cup games, and real-time Twitter data collected by monitoring the conversations about all soccer matches of four major European tournaments (FA Premier League, Serie A, La Liga, and Bundesliga), and the 2014 UEFA Champions League, during the period between Oct. 25th 2014 and Nov. 26th 2014.