Rede de unidades recorrentes chaveadas e transformadas discretas de ondaletas à previsão e operação no mercado financeiro
Carregando...
Arquivos
Citações na Scopus
Tipo de produção
Dissertação
Data
2021
Autores
Biazon, V.
Orientador
Bianchi, Reinaldo Augusto da Costa
Periódico
Título da Revista
ISSN da Revista
Título de Volume
Citação
BIAZON, V. Rede de unidades recorrentes chaveadas e transformadas discretas de ondaletas à previsão e operação no mercado financeiro. 2021. 109 f. Dissertação (Mestrado em Engenharia Elétrica) - Centro Universitário FEI, São Bernardo do Campo, 2020. Disponível em: https://doi.org/10.31414/EE.2021.D.131312.
Texto completo (DOI)
Palavras-chave
Redes neurais (Computação),Bolsa de valores
Resumo
Operar no mercado de ações sempre trás consigo o desafio de escolher a melhor decisão a ser tomada a cada passo no tempo. O problema é intensificado pela teoria de que não é possível prever uma série temporal do mercado financeiro uma vez que toda informação relacionada ao preço da ação já está contida em si, teoria conhecida como Hipótese dos Mercados Eficientes - Efficient Market Hypothesis (EMH). Embora o mercado em geral não tenha tendências distinguíveis, portanto sendo consistente com a EMH, existem várias janelas de tempo onde há alguma previsibilidade de certa extensão nos dados caso consideremos o uso de indicadores técnicos. Neste trabalho é proposto um novo modelo que busca se beneficiar de tais períodos
operando para escolher suas decisões e aguardando o melhor momento para executá-las. Este modelo, chamado Rede de Transformada Discreta de Ondaletas e Unidade Recorrente Chaveada - Discrete Wavelet Transform Gated Recurrent Unit Network (DWT-GRU), é dividido em três módulos, sendo eles, o pré processamento dos dados pela transformada de wavelets, o treinamento e a predição do preço de fechamento da ação para dois dias no futuro e a tomada da
decisão baseada na avaliação do gradiente do preço de fechamento. O modelo proposto foi comparado com outras arquiteturas de Rede Neural Recorrente - Recurrent Neural Network (RNN), com e sem o uso de pré processamento de ondaletas, e a estratégia de "comprar e segurar". Os resultados mostraram que o modelo proposto superou todos as métricas estatísticas, de acurácia, precisão, revocação e F1, e retorno financeiro de todos os modelos de comparação estabelecidos nas ações analisadas do mercado financeiro brasileiro. As ações analisadas como base para o estudo foram as blue-chips do índice do IBOVESPA sendo elas a PETR4, VALE3, ITUB4, ABEV3, e a Fundo de Índice Negociado em Bolsa - Exchange-Traded Fund (ETF) que espelha
o próprio índice, BOVA11. Para dados de treinamento foram utilizados dados desde 2001 para as ações e desde 2008 para a Fundo de Índice Negociado em Bolsa - Exchange-Traded Fund (ETF) BOVA11. Por fim é apresentado o resultado financeiro da aplicação do algoritmo em operações em tempo real de swing-trade comprovando sua eficácia e vencendo a estratégia de "comprar e segurar".
Trading in the stock market always comes with the challenge of choosing the best decision to take on each time step. The problem is intensified by the theory that it is not possible to predict stock market time series as all information related to the stock price is already contained in it, which theory is known as Efficient Market Hypotesis. Although the market, in general, has no distinguishable tendencies, thus being consistent to the EMH, there are several time windows where there is some predictability, to some extent in the data, if we consider the use of technical indicators. In this work, a novel model is proposed to seek benefit from said periods operating to choose its actions and waiting for the best moment to execute them. This model, called Discrete Wavelet Transform Gated Recurrent Unit Network) (DWT-GRU), is divided in three modules, them being, the preprocessing of the data by the wavelet transform, the training and prediction of the closing price two days in the future and the decision making based on the evaluation of the gradient of the closing price. The proposed model was compared to other RNN architectures, with and without the use of wavelet preprocessing, and the "buy-and-hold" strategy. The results shown that the proposed model surpassed all the statistical metrics of accuracy, precision, recall, F1 and financial return of all the estabilshed comparisson models in the analysed stocks of the Brazilian Stock Market. The analysed stocks as the base for the study were the blue-chips of the IBOVESPA index, them being, PETR4, VALE3, ITUB4, ABEV3, and the ETF that mirrors the index itself, BOVA11. As training data the analysed period was since 2001 for the stocks and 2008 for the Fundo de Índice Negociado em Bolsa - ExchangeTraded Fund (ETF) BOVA11. At last, it is presented the financial results of the application of the algorithm in real time swing-trade operations validating its efficiency and winning over the buy-and-hold strategy.
Trading in the stock market always comes with the challenge of choosing the best decision to take on each time step. The problem is intensified by the theory that it is not possible to predict stock market time series as all information related to the stock price is already contained in it, which theory is known as Efficient Market Hypotesis. Although the market, in general, has no distinguishable tendencies, thus being consistent to the EMH, there are several time windows where there is some predictability, to some extent in the data, if we consider the use of technical indicators. In this work, a novel model is proposed to seek benefit from said periods operating to choose its actions and waiting for the best moment to execute them. This model, called Discrete Wavelet Transform Gated Recurrent Unit Network) (DWT-GRU), is divided in three modules, them being, the preprocessing of the data by the wavelet transform, the training and prediction of the closing price two days in the future and the decision making based on the evaluation of the gradient of the closing price. The proposed model was compared to other RNN architectures, with and without the use of wavelet preprocessing, and the "buy-and-hold" strategy. The results shown that the proposed model surpassed all the statistical metrics of accuracy, precision, recall, F1 and financial return of all the estabilshed comparisson models in the analysed stocks of the Brazilian Stock Market. The analysed stocks as the base for the study were the blue-chips of the IBOVESPA index, them being, PETR4, VALE3, ITUB4, ABEV3, and the ETF that mirrors the index itself, BOVA11. As training data the analysed period was since 2001 for the stocks and 2008 for the Fundo de Índice Negociado em Bolsa - ExchangeTraded Fund (ETF) BOVA11. At last, it is presented the financial results of the application of the algorithm in real time swing-trade operations validating its efficiency and winning over the buy-and-hold strategy.