Functional Time Series Models for Dynamic Updating Prediction of Sugar Production

Functional data analysis has been widely applied across various fields, yet its use in predicting crop production outcomes, particularly sugar production in Sub-Saharan Africa, remains limited. Traditional crop forecasting relies heavily on conventional time series models, with minimal integration of dynamic updating techniques for functional time series. This study proposes three dynamic updating methods —Ridge Regression, Penalized Least Squares, and Local Polynomial Regression —to enhance sugar production forecasting in African countries. Monthly sugar production data were structured into N-yearly curves, allowing for real-time prediction updates during the current season. Model performance was evaluated against the auto-ARIMA algorithm, revealing that at least one proposed functional time series method outperformed ARIMA across all datasets. Specifically, Ridge Regression provided the best estimates for Eswatini. While the Penalized Least Square and Local Polynomial Regression were the most effective for Kenya and South Africa, respectively. Overall, dynamic updating proved beneficial in refining initial predictions and enabling timely adjustments to industry strategies. The study advocates integrating dynamic updating with traditional forecasting approaches to enhance prediction reliability, particularly for volatile agricultural datasets.