Stock Prediction from Unlabeled Press Releases using Machine Learning and Weak Supervision

Abstract

This thesis examines the effect of press releases on the Nordic stock market. A weak supervision approach is utilized to estimate the short-term effect on stock returns given press releases of different categories. By utilizing the data programming framework as implemented in the Snorkel library, approximately 24% of all press releases are categorized into a set of 10 distinct categories. Further, a collection of machine learning models for stock price prediction is developed, where simulation is conducted to determine how press releases may be used to forecast stock price movement. Stock price prediction is performed for large stock price movements and for stock price direction, where the result shows that the best performing model achieves a 53% F1-score and 54% accuracy respectively for the tasks. Finally, it appears that the labeled press releases can be used to increase the predictability of stock movements in the Nordic stock market.