Sentiment building from textual data content in quarterly reports

Abstract: Textual analysis is increasingly becoming a reliable tool for pattern assessing and forecasting especially statistically. Implementing such techniques in the financial field is still in an infantry stage and it represents a novel research area. While previous literature tends to rely on news articles as the source of data when examining the link with financial returns, internal information such as annual or quarterly reports drew small to no interest in academic research. This study aims to investigate the relationship between textual data content in quarterly reports and the corresponding stock price movement, to end up with a high-accuracy model when it comes to predicting the return. In this paper, we apply two different methods for sentiment building, first relying on python programming language and its lexicon-based approaches, second collaborating with Parlametric to establish document-specific sentiments. While the first method, combined with modeling done through MATLAB, lacked significance even with decent accuracy levels in sign prediction, the dictionary-free framework offered higher degrees of certitude. Our results confirm the hypothesis of an existing link between the asset returns and quarterly reports which is very encouraging especially for future studies.