An Investment Strategy Based on News Sentiment Words and Its Empirical Performance

Information published by news media affects stock prices through two channels. One is the information therein, which can cause a permanent price change. Pinpointing general durations and correlations for every piece of information and price change with a machine learning algorithm is difficult, however. The second channel is investor sentiment, which can result in a temporary price change. Much literature contains general rules for predicting price changes based on investor sentiment. This study uses machine learning's automatic text classification algorithm to train a classifier with a sufficient amount of past data (news articles and stock returns). Using an article's positive and negative words as its features, the author labels the corresponding return from three days before to three days after the article's release. After the classifier has been trained, each news article can be classified into the appropriate category by the classifier, and the investment strategy can be implemented accordingly. The empirical evidence suggests that the proposed strategy has a short-term positive return and does not perform significantly differently for various levels of sentiment.