UC Santa Cruz

Stock prediction is a difficult task. Recently many studies focus on Natural Language Processing methods to draw information from texts and perform forecasting. This paper purpose to build a stock prediction system which utilize textual data from multiple resources such as financial news and social media feeds. The system applies different NLP methods like VADER analyzer and Word2vec representation, to extract sentiment scores and investor mood from texts. Then it combines the data with historical prices and volumes and uses Random Forest and SVM algorithms to train models, which make predictions on short-term stock price movements. Experiment results show good accuracy and F1 scores for some of the stocks and for the combined model of all stocks. The highest accuracy is for Apple shares, 75.68%. I also simulate two trading strategies based on the news sentiment and investor mood indicators respectively. They both outperform simple buy-and-hold strategy.