Análise comparativa entre modelos ARIMA e modelos de Machine Learning para previsão de séries temporais de commodities

Abstract

This study addresses the comparative analysis between Arima models and machine learning models for forecasting commodity time series applied to soybeans, a Brazilian commodity. The methodological approach of this study can be categorized as quantitative and descriptive. This research focuses on the description and analysis of data related to the time series of soybean prices. Furthermore, the approach is quantitative, as it makes use of statistical and Machine Learning (ML) methods for predictive data analysis. As for the econometric models (ARIMA), the methodology of Box, Jenkins and Reinsel (2008) was used and for the ML models (Random Forest and SVR), the methodology of Parmezan (2016) was used. The data were obtained from the CEPEA/ESALQ website, using the series of prices in dollars differentiated in 1 day. Both econometric models and ML models stood out with predictive accuracy, obtaining satisfactory results, showing that simple linear models such as ARIMA are still useful and also the potential of ML to deal with financial ST, having the ability to anticipate changes with the time. When investigating the scarcity of work applying these methods to the Brazilian financial market and the scarcity of work in Brazil on commodities in the last decade in Brazil, it is noted that there are several promising directions for future research. Finally, it was possible to achieve all the objectives of this work and also contribute to the advancement of knowledge in the field of time series forecasting applied to finance and commodities.