Using deep learning models and exogenous production variables to forecast prices in the Brazilian sugarcane sector

Area of Science:

• Agricultural Economics
• Data Science
• Climate Science

Background:

• Accurate price forecasting for agricultural commodities like sugar and ethanol is crucial for market stability.
• Traditional forecasting methods often struggle with complex seasonal, climatic, and nonlinear patterns.

Purpose of the Study:

• To develop and compare machine learning and statistical models for forecasting CEPEA/ESALQ White Crystal Sugar and Hydrous Ethanol Fuel prices.
• To incorporate sugarcane supply (SAMUCA, Conab, actual production) and climate data (NASA POWER) into forecasting models.

Main Methods:

• Utilized Long Short-Term Memory (LSTM), Transformer, and Multilayer Perceptron (MLP) machine learning models.
• Employed the Autoregressive Integrated Moving Average (ARIMA) statistical model with exogenous variables.
• Constructed a climate indicator using temperature, precipitation, and solar radiation data.

Main Results:

• Machine learning models, particularly MLP for sugar (3.84% MAPE) and LSTM for ethanol (1.87% MAPE), demonstrated superior performance.
• Models using Modular Agronomic Simulator for Sugarcane (SAMUCA) production estimates showed accuracy comparable to historical production data.
• Machine learning techniques effectively captured complex market dynamics, outperforming traditional methods.

Conclusions:

• The developed models provide a strategic tool for market stakeholders, enabling advance price forecasting with monthly updates.
• The approach is adaptable to other crops and regions, especially those with limited production data.
• Integrating agronomic and climate data significantly enhances forecasting accuracy for agricultural commodities.