Redes neurais artificiais e processamento de linguagem natural aplicados a previsão do minicontrato futuro do índice ibovespa

Abstract

Recently conducted scientific studies prove that there is a relationship between the information published on Social Networks and the variations in the prices of assets traded on the Brazilian Stock Exchange (BOVESPA). In these studies, Natural Language Processing (PLN) techniques are used for the treatment of textual data that enable the understanding of human language by machines, which, enriched with the historical information of the assets, generate indications for decision-making in the stock market negotiations. The relationships between the Social Network Twitter and Bovespa are approached through the use of PLN in the Social Network database with Word Embedding, performing a dichotomous classification for decision making, not taking into account the practices of greater returns with the earnings of the variations of asset in the small intervals between the day. The purpose of this work is to create a model for decision making in the financial market supported by messages related to BOVESPA on the Twitter, handled by Natural Language Processing (PLN) techniques. At this point, complete sentences are used for Word Embedding vectorization and classified with a Recurrent Neural Network (LSTM) to indicate trades in the BOVESPA mini-index asset with performance governed by the market trend plus the Word Embedding classification, aggregated into 5, 15 and 30 minutes, for actions in the sequence of minutes of day trade operations. The experiments carried out in this work demonstrated the validity of the hypothesis that messages from a social network can support decisions in the financial market, allowing to obtain profits in this domain.