Incorporating business news sentiment into dynamic panel models to forecast stock prices

Abstract

Simultaneously observing the news and the development of financial markets suggests that both are interrelated in some way. In this dissertation, we explore this casual observation on the firm level by analysing the interactions between news sentiment, which we assess by Machine Learning techniques, and returns. Thereby, we base our analysis on all S&P 500 constituents’stock prices and news headlines published by the ‘Reuters Newswire’ between 1st of March 2019 and 30th of June 2020. Estimating dynamic panel models, we conclude that the causal relationship between the firm-specific daily news sentiment and a firms’ excess returns is mutual. News sentiment predicts next day returns that are not reversed within a trading week. Thus, we find evidence that newswires contain fundamental information. Further, excess returns predict sentiment, which indicates that newswires report about past events as well.These findings are in line with previous research of Ahmad et al. (2016). In addition, we investigate the out-of-sample accuracy of the fitted dynamic panel models by industry, level of media coverage and in a test set characterised by the outbreak of Covid-19. From these analyses, we obtain that industry and media coverage are not related to the prediction accuracy, which confirms the results of Hendershott, Livdan and Schürhoff (2015) and Tetlock (2010) respectively. Contrastingly to findings of Antweiler and Frank (2006) as well as García (2013),that suggest improved sentiment prediction accuracies during recessions, we find that the accuracy of our model is reduced following the outbreak of Covid-19.

Date of Award	15 Oct 2020
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Viktor Pekar (Supervisor) & José Faias (Co-Supervisor)