The Use of Long Short-Term Memory Models to Estimate Soybean Pricing: A Regional Climate Data Evaluation From Brazil

This work uses machine learning methods to analyze the influence of the Brazilian climate on international soybean price variability. For this purpose, climatic data, historical series of the extended national consumer price index, dollar price historical series, soybean harvest historical series, and international soybean price historical series are collected from different data sources. The data is then preprocessed and grouped by state. Afterward, data from the five most productive Brazilian states, Mato Grosso (MT), Mato Grosso do Sul (MS), Goiás (GO), Paraná (PR), and Rio Grande do Sul (RS), which together correspond to 74.9% of total production, are chosen. Then, predictive models were built using the long short-term memory (LSTM) method, which is applied to each state dataset considering two different settings: with and without climatic data. Thus, two models were built for each Brazilian state, totaling ten predictors. These models are evaluated using 10-fold sequential split cross-validation, whose predictive results are used for performance comparison through statistical hypothesis testing. In the experimental evaluation, when trained with climatic data, the PR model achieved the lowest average values of root mean squared error (RMSE) and mean absolute error (MAE), which are 20.90 and 16.72, respectively. Regarding predictors trained without climate data, the RS and PR obtained the lowest averaged values of RMSE (20.84) and MAE (16.81), respectively. After performing statistical hypothesis tests, no significant differences were found between the models, stating, with certainty 95%, that the Brazilian climatic data did not influence soybean prices in our study.