In this paper, a fine-grained supervised approach is proposed to identify bullish and bearish sentiments associated with companies and stocks, by predicting a real-valued score between − 1 and + 1. We propose a supervised approach learned by using several feature sets, consisting of lexical features, semantic features and a combination of lexical and semantic features. Our study reveals that semantic features, most notably BabelNet synsets and semantic frames, can be successfully applied for Sentiment Analysis within the financial domain to achieve better results. Moreover, a comparative study has been conducted between our supervised approach and unsupervised approaches. The obtained experimental results show how our approach outperforms the others.