Construction of Financial News Sentiment Indices Using Deep Neural Networks

With the progress in natural language processing and machine learning, using deep learning techniques to extract valuable information from financial texts has gained popularity among researchers. However, as deep learning models require large amounts of labelled data, current models for financial text mining are sensitive to noise and prone to overfitting due to the lack of labelled data in financial fields. We address this issue by using pretrained word representations and BERT model to leverage knowledge acquired from huge text corpus. In particular, we design a deep neural network model that combines CNN, Att-BLSTM and BERT for predicting financial news sentiment polarity. Then, the news sentiment indices are constructed based on the predictive model. In terms of predicting news sentiment polarity, we show that our model outperforms state-of-the-art competitors. Furthermore, we show that the news sentiment exhibits a significant relationship with stock market returns