Previsão de preços de soja no Brasil por meio de técnicas de machine learning

Abstract

Agricultural commodities play an important role in the global economy, with these products accounting for approximately 40% of Brazil's agricultural exports. Price forecasting in this segment is a major challenge, as it has enabled a positive and significant influence on the results of the activity, such as producer profits and the competitiveness of the sector, as well as the strategies of entrepreneurs and investors. This academic experiment investigates and develops commodity price forecasting models in Brazil, with the application of advanced machine learning methods, specifically, soybean prices in Brazil. It aims to contribute to the understanding of how the application of improved machine learning techniques based on commodity prices, assuming the hypothesis that such models prove predictive performance in relation to more traditional time series models. The justification is to provide stakeholders with critical information for informed decision-making and strengthening the country's economy. To certify this objective, time series models, such as SARIMA, and machine learning standards, such as those based on decision trees (Random Forest) and regularization methods (LASSO), were explored. The results obtained with the application of different models indicated that, although SARIMA captures the seasonality of prices, machine learning, especially Lasso, will reach near accuracy and will be efficient in predicting the behavior of soybean prices in Brazil.