In this study, we developed a price prediction model for the CEPEA/ESALQ white crystalline sugar and hydrous ethanol fuel index that incorporates variables related to sugarcane supply and climate conditions. Machine learning models such as Long Short-Term Memory (LSTM), Transformer, and Multilayer Perceptron (MLP) are used along with statistical autoregressive integrated moving average (ARIMA) models, all incorporating exogenous variables. These variables include estimates from the Sugarcane Modular Agriculture Simulator (SAMUCA), data from the National Supply Company (Conab), actual production, and climate indicators constructed from temperature, precipitation, and solar radiation data provided by NASA POWER. The results show that the model using SAMUCA production estimates had comparable accuracy to the model based on historical production data (used as a benchmark), albeit with some variation in the error metrics. The MLP model achieved the best performance for sugar (MAPE 3.84%), while LSTM was the most effective for ethanol (MAPE 1.87%). Machine learning techniques have outperformed traditional methods in capturing seasonal, climate, and nonlinear patterns. The proposed approach allows for pre-harvest price forecasting and monthly updates, providing a strategic tool to market players. The model can also be applied to other crops and regions where limited production data are available.
