Abstract

This study seeks to determine the possibility of a profitable trading strategy when examining the relationship at an hourly level across various centrality measures. Utilizing the existing dataset from Dr. Yang, we establish a methodology to further segment the data into hourly buckets, where we also present the empirical distribution of tweets throughout the day. Predicated on the definition of the financial community by Dr. Yang, we examine the top 2844 users’ tweets and construct a weighted sentiment measure that we can regress against hourly stock market returns. After expanding the analysis to sector ETFs of the S&P 500, we observe that the weighted sentiment measure is significantly correlated at the lag-1 level to various ETF returns. We construct a rolling regression and back testing methodology where we observe profitable trading strategies for at least one centrality measure for every ETF.

Research Topics:

Twitter, Sentiment Analysis, Regression Analysis, Trading Strategy

Researchers:

Research Group (2016 Fall):

Ronak Shah, Master in Financial Engineering, Graduated in Dec 2017
Dakota Wixom, Master in Financial Engineering, Graduated in Dec 2017

Advisor

Dr. Steve Yang

Main Results:

Overall, we see that across all indexes, at least one measure of centrality outperforms the corresponding index. With some indexes, centrality measures significantly outperform due to multiple trading signals from the strategy. Meanwhile, other indexes have centrality measures that closely track the benchmark, which is indicative of few trading signals from the strategy. We observe a 49-55% accuracy rate of the directional prediction across all centrality measures and indexes. Given the accuracy rate, we can definitively state that trading strategies that are predicated on Twitter sentiment and that outperform the benchmark security exist.

Conclusions:

Our study shows that sentiment analysis on Twitter messages from a financial community, weighted by different measures of centrality, can produce a statistically significant predictor of the hourly returns of broad market indexes. This sentiment index can then be used to create a trading strategy that buys or sells broad market indexes based on a trained model.

Other avenues that should be considered when extending this research are position sizing and risk management. Strategies can be weighted based on historical volatility to create a risk-adjusted portfolio. In addition, a rolling sentiment indicator can be utilized to improve prediction accuracy, as there is a strong possibility that sentiment regression models should only use the immediate past and not an extensive history.