The effect of local news on the Maltese stock market : an empirical evaluation using machine learning

Abstract

Stock market predictability has been a longstanding issue in financial research and practice. This dissertation investigates the possibility of finding patterns within the Maltese stock market using a synthesis of linguistic and statistical methods. In particular, this work applies text mining and machine learning to determine the performance of different supervised prediction models and the predictive power of non-quantifiable news articles for future market movement. Accordingly, a set of web scraping tools are proposed to retrieve news data from local online sources. The corpus of articles is diversified with articles from the business and national news sections. Each news instance is represented in a feature space using sentiment and keywords as primary attributes. Article sentiment is determined using sentiment analysis techniques, where every individual article is categorised as positive or negative. Different keyword extraction methods are then used to identify terms that provide a clear description of the article content. Each news instance is labelled either as a rise or drop, based on the change in the Malta Stock Exchange (MSE) Index. Finally, a set of base learning algorithms are used to generate predicted labels for every instance in an unseen test set. Testing is iterated to evaluate various aspects of the predictive models. The results of this experiment contribute towards an empirical appraisal of the application of various learning algorithms in correlating news stories with fluctuations in the MSE index. This research also manages to provide a comparative view on the effectiveness of different feature types in building accurate models. Results are corroborated with experts from the finance domain and it is established that the proposed methodology is substantially reflective of the true dynamics of the Maltese market.