This thesis tries to answer the question how to predict the reaction of the stock market to news articles using the latest suitable developments in Natural Language Processing. This is done using text classiffication where a new article is matched to a category of articles which have a certain influence on the stock price. The thesis first discusses why analysis of news articles is a feasible approach to predicting the stock market and why analysis of past prices should not be build upon. From related work in this domain two main design choices are extracted; what to take as features for news articles and how to couple them with the changes in stock price. This thesis then suggests which different features are possible to extract from articles resulting in a template for features which can deal with negation, favorability, abstracts from companies and uses domain knowledge and synonyms for generalization. To couple the features to changes in stock price a survey is given of several text classiffication techniques from which it is concluded that Support Vector Machines are very suitable for the domain of stock prices and extensive features. The system has been tested with a unique data set of news articles for which results are reported that are signifficantly better than random. The results improve even more when only headlines of news articles are taken into account. Because the system is only tested with closing prices it cannot concluded that it will work in practice but this can be easily tested if stock prices during the days are available. The main suggestions for feature work are to test the system with this data and to improve the filling of the template so it can also be used in other areas of favorability analysis or maybe even to extract interesting information out of texts.
|Place of Publication||Universiteit Twente|
|Publisher||University of Twente|
|Number of pages||95|
|Publication status||Published - 30 Jan 2004|