Ciência de dados e aprendizado de máquina para predição em séries temporais financeiras

Detalhes bibliográficos
Ano de defesa: 2019
Autor(a) principal: Caio Mário Henriques Silva da Rocha Mesquita
Orientador(a): Não Informado pela instituição
Banca de defesa: Não Informado pela instituição
Tipo de documento: Dissertação
Tipo de acesso: Acesso aberto
Idioma: por
Instituição de defesa: Universidade Federal de Minas Gerais
Brasil
ICX - DEPARTAMENTO DE CIÊNCIA DA COMPUTAÇÃO
Programa de Pós-Graduação em Ciência da Computação
UFMG
Programa de Pós-Graduação: Não Informado pela instituição
Departamento: Não Informado pela instituição
País: Não Informado pela instituição
Palavras-chave em Português:
Link de acesso: http://hdl.handle.net/1843/30444
Resumo: Throughout history several forecasting models have emerged with the objective of understanding the behavior of asset price series in the financial market. The advancement of computational power has facilitated the creation of new, increasingly complex models that arise for this purpose. However, even with the use of advanced machine learning techniques using a large volume of historical data, this task remains quite challenging, remaining an open problem. The objective of this work is to create automated strategies of operation in the market, based on a forecast model of trends in the prices of financial series, through machine learning. A recurrent neural network Long Short Term Memory is used as the predictive model. The paper also aims to demonstrate that several of the financial series have a temporal correlation, even if small, which allows the construction of forecasting models that are based on historical data. In order to demonstrate this correlation, the statistical properties of the series are analyzed and hypothesis tests are applied to them. The work presents a robust methodology from the data collection to the simulation of operation in the market involving the operating costs for 38 assets of the Brazilian stock exchange. The methodology further presents a method for creating a more correlated attribute with future values by means of a linear combination of the historical series in different time lags. The results obtained are promising since the best forecasting models obtained Accuracy values of up to 63% and financial return values of up to 47%. The best cases outperformed both in terms of prediction and in terms of financial return compared to baselines techniques as random classifier, Buy and Hold strategy, SELIC and CDI rates.