Sentimental Value: The Viability of Momentum and News Sentiment in Stock Price Prediction Models & The Impact of Retail Investors

Abstract

Stock market behaviour has long been a field of interest among hedge funds and asset managers. With the increase in machine learning techniques and access to large amounts of data, firms have invested huge amounts of resources and capital into exploring predictive models and their success in trading the market through back-testing hypotheses. These models can often be broken down into 'factor based' approaches, such as momentum. This paper explores the impact and value of using news data and sentiment data, as well as momentum, in the prediction of directional changes in stock price. Specifically, the success for both momentum based indicators and news sentiment based indicators from February 2018 - December 2019 (referred to as Pre COVID-19) is analyzed, and compared to the success from January 2020 - October 2020 (referred to as Post COVID-19). This attempts to capture the changes in the importance of both momentum and sentiment as retail investor percentage in the market increased in 2020 (Winck, 2020). Market sentiment surrounding the Standard & Poor's (S&P) 500 companies is extracted from news articles by RavenPack (WRDS, 2020a). Model success is judged by the ability to correctly predict stock movements, with 1 signifying an increase in stock price and 0 signifying a decrease in stock price, and accuracy is determined by the percentage of correct predictions. The results of these models are then compared for the two timeframes (Pre and Post COVID-19) to determine the importance of sentiment and momentum. The success rates for various modelling techniques, including support vector machines (SVMs) and neural networks are also compared to determine the highest accuracy predictions. Although the data for Post COVID-19 stock prices is limited, the results indicate that with a rise in retail investors, sentiment indicators become increasingly important, while momentum indicators perform better for lower retail investor percentage. With regards to modelling techniques, for linear stock data transformations, the simpler linear models such as regularized regressions and linear SVMs tend to have the highest prediction accuracy.