Redes neurais artificiais para predição do preço de vale3 na bolsa de valores brasileira.
Data
2021-08-06
Autores
Barros, Gustavo Cavalcante de 
Orientadores
Martins, Tiago Dias
Tipo
Trabalho de conclusão de curso
Título da Revista
ISSN da Revista
Título de Volume
Resumo
O trabalho aqui apresentado tem por objetivo caracterizar um modelo estatístico-preditivo de redes neurais artificiais, cuja finalidade está na obtenção probabilística de preços futuros, utilizando-se de dados passados, do ativo financeiro VALE3, na bolsa de valores brasileira. A escolha para utilização dessas redes neurais é assentada no fato de que, essas, conseguem classificar e agrupar uma grande quantidade e complexidade de dados amostrais. Para realizar a análise dos prováveis futuros preços do ativo, fora demarcado uma amostragem do ativo de maior peso da Bolsa de Valores Brasileira na data de 25/01/2021. O primeiro impasse para estabelecer o método preditivo é conseguir desmentir, utilizando-se do chamado “Coeficiente de Hurst”, a teoria financeira “Random Walk”, onde afirma que os preços evoluem de forma aleatória, não podendo ser previstos. Posteriormente, com a teoria “Random Walk” desmentida, tornou-se probabilisticamente possível realizar a previsão dos preços futuros dos ativos com o auxílio de redes neurais, estabelecendo-se um horizonte de, no máximo, uma semana de operação, isto é, 5 dias. Utilizando-se do software MATLAB e variando-se as camadas internas das redes neurais, variáveis de entrada e funções de ativação, pode-se obter diferentes resultados de eficiência das predições e, para verificar o desempenho dos preços preditos com os preços reais, serão utilizados medidores de desempenho, como por exemplo o coeficiente de determinação (R²) e o Erro Médio Quadrático (EMQ). Após a realização de todos os testes, permutando-se entre as funções de ativação, com a faixa de neurônios ocultos de 5 à 35, variando-se de 5 em 5, foi possível identificar que a melhor configuração de rede neural obtida, isto é, aquela que possuía menor erro de resultado previsto versus resultado real, foi a que possuía o algoritmo de otimização “trainlm”, função de ativação da 1ª camada “tansig” e função de saída “purelin”, havendo 35 neurônios em sua camada oculta. Essa configuração foi posteriormente utilizada para realizar as simulações do preço de VALE3 nos períodos entre 2 de janeiro de 2017 a 1 de janeiro de 2019, onde obteve-se um resultado excelente do erro em simulações do preço em d+1 de apenas 1,50% e, utilizando-se como modelo de gatilhos de compra e venda do ativo, foi atingida uma rentabilidade de 5,5% líquidos ao mês, durante o período estudado.
The work presented here aims to characterize a statistical-predictive model of artificial neural networks, which purpose is to obtain probabilistic future prices, using past data of the financial asset VALE3 on the Brazilian stock exchange. The choice for using these neural networks is based on the fact that they are able to classify and group a large amount and complexity of sample data. In order to analyze the probable future prices of the asset, a sample of the most important asset of the Brazilian Stock Exchange on 01/25/2021 was demarcated. The first impasse to establish the predictive method is to be able to refute, using the so-called “Hurst Coefficient”, the “Random Walk” financial theory, which states that prices evolve at random and cannot be predicted. Later, with the "Random Walk" theory refuted, it became probabilistically possible to forecast the future prices of assets with the help of neural networks, establishing a horizon of at most one week of operation, that is, 5 days. Using the MATLAB software and varying the inner layers of neural networks, the input variables and the activation functions, different prediction efficiency results can be obtained and, to verify the performance of the predicted prices with the real prices, performance meters will be used, such as the coefficient of determination (R²) and the Mean Squared Error (EMQ). After performing all tests, swapping between the activation functions, with the range of hidden neurons from 5 to 35, varying from 5 to 5, it was possible to identify that the best neural network configuration obtained, that is, the one that had the smallest error of predicted result versus actual result, was the one that had the optimization algorithm "trainlm", activation function of the first layer "tansig" and output function "purelin", with 35 neurons in its hidden layer. This configuration was later used to perform the price simulations of VALE3 in the periods between January 2, 2017 and January 1, 2019, in which an excellent error result in d+1 price simulations of only 1.50% was obtained and, using the asset's purchase and sale triggers as a model, a return of 5.5% net per month was achieved during the period studied.
The work presented here aims to characterize a statistical-predictive model of artificial neural networks, which purpose is to obtain probabilistic future prices, using past data of the financial asset VALE3 on the Brazilian stock exchange. The choice for using these neural networks is based on the fact that they are able to classify and group a large amount and complexity of sample data. In order to analyze the probable future prices of the asset, a sample of the most important asset of the Brazilian Stock Exchange on 01/25/2021 was demarcated. The first impasse to establish the predictive method is to be able to refute, using the so-called “Hurst Coefficient”, the “Random Walk” financial theory, which states that prices evolve at random and cannot be predicted. Later, with the "Random Walk" theory refuted, it became probabilistically possible to forecast the future prices of assets with the help of neural networks, establishing a horizon of at most one week of operation, that is, 5 days. Using the MATLAB software and varying the inner layers of neural networks, the input variables and the activation functions, different prediction efficiency results can be obtained and, to verify the performance of the predicted prices with the real prices, performance meters will be used, such as the coefficient of determination (R²) and the Mean Squared Error (EMQ). After performing all tests, swapping between the activation functions, with the range of hidden neurons from 5 to 35, varying from 5 to 5, it was possible to identify that the best neural network configuration obtained, that is, the one that had the smallest error of predicted result versus actual result, was the one that had the optimization algorithm "trainlm", activation function of the first layer "tansig" and output function "purelin", with 35 neurons in its hidden layer. This configuration was later used to perform the price simulations of VALE3 in the periods between January 2, 2017 and January 1, 2019, in which an excellent error result in d+1 price simulations of only 1.50% was obtained and, using the asset's purchase and sale triggers as a model, a return of 5.5% net per month was achieved during the period studied.
Descrição
Citação
Trabalho de conclusão de curso de graduação