Análise e previsão do resultado financeiro da produção de commodity agrícola

Abstract

Agriculture is one of the main pillars of the global economy. The activities in the agricultural sector carry risks associated with factors beyond the control of the chain agents. In view of this, mathematical models have been applied that use important variables to predict the financial outcome of the agribusiness sector. In this context, the first part of this dissertation combines bibliometric (systematic) analysis with the Latent Dirichlet Allocation (LDA) method, culminating in a semi-automatic approach for the literary review, while the second part consists of the application of some mapped methodologies. The main objective was to automate the identification of relevant topics and to obtain a final bibliographic portfolio (BP) that covered the period from 2015 to 2022, for the extraction of methodologies used in the articles, as well as other bibliometric analyses. The 30 articles in the BP address issues about the methodologies applied in the temporal analysis of agricultural commodities. The articles that made up the BP were allocated into categories based on the nature of the forecasting models used, classified as: (i) machine learning (ML), (ii) machine learning and artificial neural networks (ML-ANN); (iii) machine learning and ensemble methods (ML-Ensemble); (iv) machine learning and hybrid methods (ML-Hybrid); and (v) statistical methods. Regarding the results, the most prominent topic was called "Forecasting Methods Applied to Agribusiness Time Series." The most used classes were the ML-Hybrid (41.95%) and statistical methods (29.31%), followed by ML-ANN methodologies (14.94%), ML (9.20%), and ML-Ensemble (4.60%). Therefore, there is an opportunity to further explore ensemble methodologies and ML. The second part of this study aimed to investigate the behavior of the variables price, production, production cost, and financial outcome of soybeans. To achieve this goal, time series methods (integrated moving average model, Naïve method, and Holt-Winters method) and machine learning models (artificial neural networks and support vector regression) are used to generate short-term forecasts for the quantity produced, production cost, prices, and financial outcome. The performance of the models is evaluated using mean absolute percentage error, mean absolute error, root mean square error, and standard deviation of errors. This research uses secondary data obtained from government databases (SEAB/DERAL and CONAB). The results indicated that, in general, both support vector regression (SVR) and extreme learning machine (ELM) are capable approaches for generating adequate forecasts for the studied variables. The first is effective in the context of production cost and production forecasts, while the second is more suitable for price and financial outcome forecasts. The theoretical contribution of this study was to identify literature gaps on forecasting methods applied to agribusiness. The practical implication was to identify forecasting methodologies to support decision-making. Finally, the consulted literature showed that the financial outcome has not been the focus of study, representing a good research opportunity.