Aplicação de métodos baseados em lógica fuzzy e de aprendizado de máquina para a previsão de preços de commodities

Abstract

Corn and soybeans are two of the main agricultural commodities in Brazil, playing a strategic role in the Brazilian economy. Forecasting their prices is of utmost importance to support decision-making in the agro-industrial sector. This study aims to evaluate the use of Fuzzy Time Series (FTS) methods for forecasting the prices of these commodities, considering univariate and multivariate approaches. These models were applied and compared to Machine Learning (ML) algorithms, such as Random Forest, XGBoost, LightGBM, among others. The dataset used in the first article comprises daily corn price data for the period from 2007 to 2025, in the state of Minas Gerais, with three subsets, beginning in 2007, 2014, and 2021. The data used in the second article refer to the global price of soybeans, other commodities, and their derivatives, collected monthly from 1960 to 2025. After partitioning the data into training and test sets, the FTS models and machine learning algorithms were trained and evaluated on the test set. Performance evaluation metrics were used, such as Mean Absolute Percentage Error (MAPE), Mean Absolute Scaled Error (MASE), Mean Directional Accuracy (MDA), among other measures widely used in the evaluation of forecasting models. In the univariate case, the results indicate that the Probabilistic Weighted FTS (PWFTS) model presented the best performance, with lower forecasting errors and higher directional accuracy when trained with data beginning in 2014, and with a one-day forecasting horizon. In contrast, the high order FTS (HOFTS) and Weighted HOFTS (WHOFTS) models presented slightly lower performance than that observed for PWFTS. Nevertheless, HOFTS and WHOFTS maintained consistent performance, although with variations among the different training strategies used. For the multivariate case, the results indicated that the models based on fuzzy logic, especially the Weighted Multivariate FTS (WMVFTS) with 20% of test data and 60 partitions, obtained the best overall performance, presenting the lowest forecasting errors and high directional accuracy. It is concluded that Fuzzy Time Series methods constitute a good alternative for forecasting agricultural commodity prices, proving to be suitable both in univariate and multivariate contexts and at different time scales.