Sentiment Analysis Applied to IBOVESPA Prediction.

Social media is increasingly being used as a source of news, a trend which has resulted in large amounts of data. This work presents an evaluation strategy for assessing the impact that social media has on the Bovespa Index (IBovespa), the benchmark index of the Brazilian stock market. A total of 105000 tweets were collected from the twitter profile of "G1 Economia", one of the main Brazilian finance portals. This data was processed using sentiment analysis methods which were then incorporated into the development of an artificial neural network whose objective was to predict the IBovespa. A hyperparameter optimization study is also presented. The experimental results show that of the 1279 topologies studied, 82.4% exhibited better performance when using sentiment analysis in conjunction with historical data, against the baseline of using only the latter. Curiously, even though the average performance was higher, the absolute best result was obtained without the use of NLP techniques. In the context of the method developed and the data used, it appears that approaches using sentiment analysis alongside historical records may be more effective than using only one or the other.