Analysis of the Association Between Topics in Online Documents and Stock Price Movements

Abstract

This paper aims at discovering the topics hidden in the newspaper articles that have an impact on movements of stock prices of the corresponding companies. Document topics are characterized by combinations of specific words in documents and are shared across a document collection. We describe the process of discovering the topics, the creation of a mapping of the topics to stock price movements, and quantifying and evaluating the results. As the method for finding and quantifying the association, we use machine learning-based classification. We achieved an accuracy of stock price movement predictions higher than 70 %. A feature selection procedure was applied to the features characterizing the topics in order to facilitate the process of assigning a label to the topic by a human expert.