Trading on online social mood: A machine learning strategy based on Twitter sentiment

Abstract

This paper examines the potential of using online social sentiment data in algorithmic trading strategies. Several machine learning models are tested to produce a trading signal from the sentiment data to forecast the trend of a stock’s price. The algorithms are trained on the features extracted from PsychSignal data (containing bullish and bearish sentiment from Twitter). One most popular model, Random Forest (RF) classifier, is selected to generate a signal for the trading strategy. After backtesting on 1386 stocks listed in both NYSE and NASDAQ, the results show that the proposed model outperforms the baseline model, a simple moving average (SMA) strategy. We use the GridSearchCV to fine-tune the parameters of the classifier and compare the performance with the SMA baseline and the SPY benchmark, showing that our model generates 114.5% return on investment from January 2013 through October 2015. Additionally, the portfolios constructed by the RF classifier appear to produce a higher return than portfolios constructed by an SMA strategy. The results show that Twitter sentiment data is a valuable technical trading indicator for specific sectors.