Using Twitter Sentiment for Stock Movement Prediction and Portfolio Optimization

Abstract

The rise of social media has given voices to many people to express their opinions and ideas. Similarly, much of the stock market is driven by news, innovations, and ultimately, the opinions of investors and consumers. Using tweets, we create a simple classification model utilizing sentence-BERT and financial time series data to explore the predictive power of the tweet, and construct a risk parity portfolio by combining our tweet data and historical returns data to modify our covariance matrix. We are able to demonstrate, using our stock-price prediction framework, that incorporating aspects of tweet data can predict economic activity with a 55% accuracy rate for our selected companies and average MCC score of .5684. However, while we find that a sentiment-enhanced risk-parity portfolio tends to be less volatile, it also has a lower expected return and a lower Sharpe Ratio, and we do not find that enhancing the covariance with sentiment analysis data results in a stronger-performing risk parity portfolio.