Redes neurais convolucionais e máquinas de aprendizado extremo aplicadas ao mercado financeiro brasileiro

• Ano de defesa: 2018
• Autor(a) principal: Faria, Elisangela Lopes de
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio de Janeiro; Brasil; Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa de Engenharia; Programa de Pós-Graduação em Engenharia Civil; UFRJ
• Palavras-chave em Português: Redes Neurais de Aprendizado Profundo; Word2vec; Mercados Financeiros; CNPQ::ENGENHARIAS::ENGENHARIA CIVIL
• Link de acesso: http://hdl.handle.net/11422/11418

This work presents a predictive study of the main index of the Brazilian stock market (Ibovespa) by using economic and financial information extracted from important financial news providers in Brazil. A hybrid model is proposed combining the potentialities of two distinct artificial neural network architectures, a convolutional neural network (deep learning architecture) and an extreme learning machine, together with techniques of Natural Language Processing, more precisely techniques of distributed representation of words (word embeddings). The influence on the results of the different parameters intrinsic to the proposed model is show and discussed. The maximum accuracy obtained with the predictive model was 60,2%. A trading strategy was developed according to the model predictions, reporting superior profitability as compared to the buy and hold strategy. Finally, the results obtained on this emerging market are similar to the other found in studies reported in the literature, although performed on well developed markets.