Predicting sentence-level polarity labels of financial news using abnormal stock returns

Abstract

Expert systems for automatic processing of financial news commonly operate at the document-level by counting positive and negative term-frequencies. This, however, limits their usefulness for investors and financial practitioners seeking specific positive and negative information on a more fine-grained level. For this purpose, this paper develops a novel machine learning approach for the prediction of sentence-level polarity labels in financial news. The method uses distributed text representations in combination with multi-instance learning to transfer information from the document-level to the sentence-level. This has two key advantages: (1) it captures semantic information of the textual data and thereby prevents the loss of information caused by bag-of-words approaches; (2) it is solely trained based on historic stock market reactions following the publication of news items without the need for any kind of manual labeling. Our experiments on a manually-labeled dataset of sentences from financial news yield a predictive accuracy of up to 71.20%, exceeding the performance of alternative approaches significantly by at least 5.10 percentage points. Hence, the proposed approach provides accurate decision support for investors and may assist investor relations departments in communicating their messages as intended. Furthermore, it presents promising avenues for future research aiming at studying communication patterns in financial news.