This thesis mainly discusses how to use news headlines as features to predict stock price directions. It is well-known that some investors believe in news analysis strategy, which is highly depending on the news. Thus we will use the popular method Doc2vec, which uses vectors to represent words based on neural networks and probability, to represent each headline as a vector. Then we use the classical machine learning model to predict individual stock price directions. Our method best perform 70% accuracy for stock price directions prediction and expect to help investors making the right strategy.