Métodos clássicos e baseados em aprendizado de máquina para previsão de preço de tomate in natura

Abstract

Predicting agricultural prices is crucial for decision-making in the agribusiness sector, impacting farmers’ income, food prices, and the economy as a whole. Tomatoes are one of the main vegetables produced and traded in Brazil, with a dynamic market affected by factors such as climatic variations and seasonality. This study aims to compare the performance of different classical and machine learning-based methods in predicting tomato prices in the Ceagesp and Ceasa/PR wholesale markets. The ARIMA, SARIMA, ARIMAX, SVR, LSTM, and CNN forecasting methods were used, and time series data on tomato prices and other correlated variables were collected between 2010 and 2021. The ADF test was used to determine the order of difference required to make the series stationary. Principal component analysis was used to reduce data dimensionality by extracting a smaller number of components that represent most of the variation observed in the data. The Ljung-Box test was applied to analyze residuals and verify model adequacy to observed data. The performance of different models was analyzed using RMSE and MAPE metrics. Variable importance analysis is an important step in machine learning models and was applied to identify the most influential predictor variables on the response variable. The non-parametric Wilcoxon statistical test was used to evaluate the difference between candidate models. Results showed that some variables were highly correlated with each other. This study concluded that the SVR model had higher accuracy compared to other models. The choice of the most suitable method may vary according to the objective of the forecast, period, availability, and quality of data, among other factors.