Predictive models for sugar and hydrous ethanol prices using supply, demand, and inventory indicators with machine learning techniques

Commodity price forecasting models are essential tools for investors, producers and other stakeholders in the sugar and biofuel value chain. However, many forecasting studies rely primarily on univariate approaches and do not adequately incorporate the economic fundamentals that shape commodity price dynamics. This study proposes a forecasting framework that incorporates supply, demand, and inventory-utilization indicators as separate fundamental specifications in price forecasting models for the CEPEA/ESALQ indicators of White Crystal Sugar and Hydrous Ethanol in São Paulo. This design allows each market-balance channel to be evaluated individually, reducing multicollinearity among closely related fundamentals while preserving the economic interpretation of supply, demand, and inventory conditions. To assess predictive performance, traditional time-series models (ARIMA and VAR/VEC) and machine-learning approaches, namely multilayer perceptron, long short-term memory (LSTM) and Transformer networks, were implemented. Forecast accuracy was evaluated using the Root Mean Square Error (RMSE) and the Mean Absolute Percentage Error (MAPE). The results indicate that incorporating economic fundamentals improves forecast performance, with indicator-specific multivariate specifications consistently outperforming univariate models. Among the approaches tested, LSTM-based models achieved the lowest forecast errors. These results suggest that forecasting models informed by market fundamentals can support decision-making and risk management in the sugar-energy sector.