Previsão de indicadores diários de preços no mercado futuro de commodities agrícolas utilizando aprendizagem de máquina

• Ano de defesa: 2020
• Autor(a) principal: Ludovico, Sérgio Nunes
• Orientador(a): Salgado, Ricardo Menezes
• Banca de defesa: Miguel, Eliseu César, Rezende, Marcelo Lacerda
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Alfenas
• Programa de Pós-Graduação: Programa de Pós-Graduação em Estatística Aplicada e Biometria
• Departamento: Instituto de Ciências Exatas
• País: Brasil
• Palavras-chave em Português: Bolsa de valores; Mercado de ações-Previsão; Investimentos; Matemática financeira.
• Área do conhecimento CNPq: CIENCIAS AGRARIAS
• Link de acesso: https://repositorio.unifal-mg.edu.br/handle/123456789/1762

The prediction of values in a time series is the object of study in several fields of knowledge. In the future market for agricultural commodities, this type of information can be used to minimize investment risks and contribute to the increase in the volume of negotiations for various commodities. As the prices of these assets are influenced by many external variables, forecasts are generally made through fundamental or technical analysis and this work is carried out by specialists in the field. This restricts the access of individuals who could invest, but does not do so because they do not have the knowledge that is necessary for the survival of this business. This study proposes a computational model, using machine learning techniques and algorithms, to predict future values in historical data series. When executing it several times, in a randomized way, seven different types of forecasts are obtained for each commodity series analyzed. The series are records of price quotations maintained by CEPEA, in US$, for sugar, live cattle, coffee, ethanol, corn and soybeans. The performance and stability of the predictions of the algorithms: k-nearest neighbors; random forest; artificial neural network; support vector machine; and extreme gradient boosting and joint learning methods: ensemble by average and stacking, are measured using statistics from the MAE, RMSE and MAPE error metrics. This constituted the computational experiment and demonstrated that the support vector machine is the algorithm with the best performance for this group of series. With the techniques applied, the results show that the forecasts have high performance during the validation of the model, suggesting that they are useful in the horizon of one step ahead. The results of this research indicate that this approach has the potential to be used as an alternative for automation of technical analysis, contributing to the reduction and quantification of forecast errors in the short term. Through the routine and frequent application of this technique, speculators and hedgers can benefit from using this approach, as support to decision making, to reduce the risks of negotiations.