Classifying Press Releases and Company Relationships Based on Stock Performance

We classify press releases as “good” or ‘bad’ news for 3 companies based on whether the stock increases n minutes after publication. We tried different classifiers (Multinomial Naive Bayes, Regularized SVM, and Nearest Neighbors) and various feature representations (such as the TF-IDF of the words in the document). We do a few percent better than majority baseline with our best setup: nearest neighbor classifier with a cosine similarity metric, binary word-in-doc features, and n = 15 minutes. Stemming words to base forms helped significantly. Using the clustering to predict the stock price of related companies did not work. Overall a lack of sufficient press release data was the limiting factor of our research. Various suggestions for improvement are discussed in the conclusion.