ARAÚJO JÚNIOR, J. G.; http://lattes.cnpq.br/1026450683366163; ARAÚJO JÚNIOR, José Gildo de.
Resumo:
For some time researchers and market analysts have shown evidence of predictability of
stock markets. Although many investors believe that the stock market is unpredictable, predictability analysis in China, Turkey, Hong Kong, Italy, Tehran and the US stock markets
has shown the opposite situation. The Efficient-Market Hypothesis (EMH) was designed in
1970 and could not anticipate the cultural and technological changes that affected the world, such as the increased processing power of computers, the development of machine learning techniques, real time publication of news and opinions of investors in social media platforms, such as twitter and facebook, for example. The combination of these elements enabled investors to perform more complex analysis of sectors, índices and stocks in almost real time, thus increasing their understanding of the stock market dynamics and improving their likelihood of success. his study aimed to shed light on the relationships and impacts that economic news published in online Brazilian newspapers, have with the national stock market in two leveis of analysis: Bovespa Index and sectors. Initially, we collected economic news published in high-circulation newspapers in Brazil between 2000 and 2015, their comments and their repercussions on social medias like Twitter, Facebook, Linkedln and GooglePlus. The correlation analysis between the Bovespa index and the amount of news shared on social networks showed a negative correlation of 48%. Furthermore, using sentrment analysis it was found that the amount of positive news reported is in average of 4.5 times higher than the negative, and, nonetheless, the negative news are more rebound on the social media than positive news. For the sectors, it was found that the most predictable sector by economic news is the Oil, Gas and Biofuels while the less predictable is the Industrial Goods sector. Finally, the variables drawn from the news were used as as input for the definition of prediction models for both the Bovespa Index and for the sectors of BM& FBOVESPA. In general, the results overperformed baselines such as the random classifier in ~ 20%.