Mutual fund shareholder letters are a relatively unregulated form of communication between investment managers and investors. They often include a narrative description of the state of the economy. I train a Hierarchical Attention Network (Yang et al., 2016) to differentiate letters written during booms and busts. To achieve good performance on this task, the model needs to take into account word and sentence context, filter out irrelevant text (e.g., boilerplate language), and be able to handle relatively large documents. I discuss how HAN addresses these challenges. I then explore possible ways to use the outputs of the model to gain insights about investor communication.
Виталий Мерсо (Vitaly Mersault) выпускник факультета экономики ЕУСПб, Ph.D. student in Financial Economics, Tepper School of Business, Carnegie Mellon University