Predicting stock market movement using Bidirectional Encoder Representations from Transformers

Abstract

The essential motivation of this work was to find out if information found in news articles is relevant for predicting future price movement of stocks. A first part of the work consists of the extraction, processing, and storage of news articles gathered from the internet. A dashboard for monitoring the article collection process and a second one for browsing the collected data have been implemented. Bidirectional Encoder Representations from Transformers (BERT) form the basis of the solution for two major challenges. The first one was to detect organizations spoken of in the articles using a pre-trained Named Entity Recognition model. The second challenge consisted in the development of a model trying to predict the future stock price based on articles about the corresponding organization. The end performance of the latter model was not convincing, but several perspectives for improvement are presented for further studies.