Estudo de caso: aplicação de machine learning para a previsão de tendências das ações das bolsas de valores brasileira e norte americana
Carregando...
Data
2023-01-31
Autores
Rodella, Vinícius Gomes
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A predição da tendência no mercado acionário é um trabalho difícil, uma vez que os bens
negociados tendem a sofrer diversas influências, como a política internacional, o movimento
do câmbio e a política interna do país em que o ativo está sendo negociado. Apesar disso, o
profissional trader aplica técnicas gráficas e de indicadores para a detecção de padrões e assim
efetuar sua análise buscando operar com alta frequência na bolsa na tentativa de obter ganhos
significativos, contrariando a hipótese do mercado eficiente. Esta detecção de padrões e a
possibilidade da análise de séries temporais, também permitem a realização de diversos estudos
que buscam modelar possíveis tendências dos ativos, sendo impulsionado pelo avanço da
inteligência artificial, possibilitando que um hardware mais robusto execute modelos robustos
e complexos. Neste contexto, este trabalho busca uma nova abordagem para este campo de
pesquisa apresentando a aplicação dos métodos de machine learning: Random Forest e
Adaboost, capazes de seguirem uma determinada estratégia de investimento utilizando análise
gráfica das ações, buscando prever a tendência de alta ou de baixa dos ativos nas bolsas de
valores brasileira e norte americana em diferentes períodos. Desta maneira, foi desenvolvido
uma metodologia utilizando 6 ações de cada região em estudo, onde foi escolhida a seguinte
combinação dos indicadores para os modelos: média móvel, índice de força relativa,
divergência e convergência da média móvel e taxa de variação. Durante o desenvolvimento,
métodos diferentes para o cálculo dos indicadores e a captura dos dados foram testados,
verificando também, a necessidade do balanceamento dos dados e uma melhor escolha dos
períodos de análise para cada ação com a finalidade de atingir uma predição satisfatória, ou
seja, onde a inteligência artificial (IA) tenha aprendizado real, evitando o overffiting. Assim, os
resultados indicam que as ações norte americanas variaram menos do que as brasileiras e isso
pode ser justificado pela maior volatilidade da bolsa brasileira. Por fim, o uso desses modelos
de aprendizado de máquina aplicados para previsão das tendências das ações nas bolsas de
valores mostrou-se inviável na prática devido a necessidade de balancear os dados, mas com
um potencial muito promissor.
Prediction of trend in the stock market is a difficult task, as the traded goods tend to suffer various influences, such as international politics, exchange rate movement, and the domestic policy of the country in which the asset is being traded. Despite, the professional trader applies graphic and indicator techniques for pattern detection and thus performs their analysis seeking to operate with high frequency in the stock exchange in an attempt to obtain significant gains, contradicting the efficient market hypothesis. This pattern detection and the possibility of time series analysis also allow for the conduct of various studies that aim to model possible asset trends, being propelled by the advancement of artificial intelligence, allowing for more robust hardware to execute robust and complex models. In this context, this work seeks a new approach to this research field by presenting the application of the machine learning methods: Random Forest and Adaboost, capable of following a certain investment strategy using graphic analysis of stocks, seeking to predict the trend of assets in the Brazilian and North American stock exchanges at different periods. In this way, a methodology was developed using 6 actions from each region under study, where the following combination of indicators was chosen for the models: moving average, relative strength index, divergence and convergence of the moving average, and rate of variation. During development, different methods for calculating indicators and data capture were tested, also verifying the need for data balancing and a better choice of analysis periods for each action in order to achieve a satisfactory prediction, that is, where artificial intelligence (AI) has real learning, avoiding overfitting. Thus, the results indicate that North American actions varied less than Brazilian actions and this can be justified by the greater volatility of the Brazilian stock exchange. Finally, the use of these machine learning models applied to predict the trends of assets in stock exchanges proved unfeasible in practice due to the need to balance data, but with a very promising potential.
Prediction of trend in the stock market is a difficult task, as the traded goods tend to suffer various influences, such as international politics, exchange rate movement, and the domestic policy of the country in which the asset is being traded. Despite, the professional trader applies graphic and indicator techniques for pattern detection and thus performs their analysis seeking to operate with high frequency in the stock exchange in an attempt to obtain significant gains, contradicting the efficient market hypothesis. This pattern detection and the possibility of time series analysis also allow for the conduct of various studies that aim to model possible asset trends, being propelled by the advancement of artificial intelligence, allowing for more robust hardware to execute robust and complex models. In this context, this work seeks a new approach to this research field by presenting the application of the machine learning methods: Random Forest and Adaboost, capable of following a certain investment strategy using graphic analysis of stocks, seeking to predict the trend of assets in the Brazilian and North American stock exchanges at different periods. In this way, a methodology was developed using 6 actions from each region under study, where the following combination of indicators was chosen for the models: moving average, relative strength index, divergence and convergence of the moving average, and rate of variation. During development, different methods for calculating indicators and data capture were tested, also verifying the need for data balancing and a better choice of analysis periods for each action in order to achieve a satisfactory prediction, that is, where artificial intelligence (AI) has real learning, avoiding overfitting. Thus, the results indicate that North American actions varied less than Brazilian actions and this can be justified by the greater volatility of the Brazilian stock exchange. Finally, the use of these machine learning models applied to predict the trends of assets in stock exchanges proved unfeasible in practice due to the need to balance data, but with a very promising potential.
Descrição
Palavras-chave
Inteligência Artificial, Artificial Intelligence, Mercado de capitais, Capital market