STATISTICAL MODELLING FOR TEXT ANALYTICS WITH APPLICATIONS TO FINANCE, ACCOUNTING AND ECONOMICS

Abstract

Textual analytics has been widely performed in finance, accounting, and economics, and other disciplines. From hand-coding or rule-based approaches in early stages to recently presented use of neural networks, there has been a high demand on developing statistical models and learning methods for capturing contextual dependency in languages in realistic settings. The prime concern toward that goal is lack of interpretability in the sense that it is hard to justify or explain the learning outputs. In this thesis, we propose several methods that allow application of neural network based textual analytics in an interpretable manner for studies of finance, accounting, and economics. First, we construct a topic tone classification model based on neural networks in a supervised setting. Second, we offer a word importance weight system and a lexicon of sentiment words based on the attention mechanism. Finally, we propose a model that incorporates texts into the regime-switching model.