In recent years data science had provided us many opportunities to uncover new social phenomena and behaviors in online social networks, and to utilize such information for business applications. One such interesting phenomenon is the use of hashtags to emphasize important content. In this paper, we evaluate the information content of hashtags for sentiment analysis applications. Specifically, we focus on multi-word hashtags, which challenge automated sentiment analysis methods. For this purpose, we develop a new algorithm to split multi-word hashtags into individual terms. We then compare the predictive accuracy of sentiment analysis with and without this finer-grained representation. We find that breaking down hashtags into multiple terms significantly improves the predictive accuracy of sentiment analysis procedures, and more generally, that hashtags are highly informative for sentiment analysis purposes.