News Sentiment in Volatility predictions : Exploring the effect of news sentiment on stock volatility using machine learning regression models

Abstract

In this thesis, we explore how sentiment from financial news could affect stock volatility. Using financial data from the S&PlO0, volatility data from the Volatility Index (VIX) and sentiment data collected with web scraping we make five different machine learning models with different covariates. Analyzing the effect in both individual sectors and a combination of all sectors, with a total of 240 different models. In order to isolate the effect of sentiment, we create datasets with and without the information and look at how the results differ. We found little proof that the additional information from the news sentiment affects the result significantly. The reason for this is complex, but we believe that using sentiment would be better suited for classification of volatility direction. Our best attempts to predict volatility on index level came from the LSTM model that got an score of 43,6% using sentiment as a covariate. The best result on an individual sector came from the random forest model that got an R2 score of 62.5% using sentiment to predict volatility in the energy sector. Although these scores isolated are acceptable, for the majority of the models, those without sentiment data performed as well, if not better.