An Analysis of Time-Series Forecasting Models for Optimizing School-Based Feeding Programs

Abstract

Accurately forecasting meal demand in School-Based Feeding Programs (SBFPs) is essential for enhancing program efficiency and minimizing food waste. This study assessed the performance of four predictive models—ARIMA, Long Short-Term Memory (LSTM), Prophet, and Random Forest Regression—in forecasting daily attendance for SBFPs. Attendance data from eight schools within the Tarlac City Schools Division, spanning August 19 to October 19, 2024, were used. The LSTM model demonstrated superior predictive accuracy, achieving the lowest RMSE (5.51), MAE (4.91), and sMAPE (0.27%), making it the most effective model. Prophet followed closely with an RMSE of 5.89, MAE of 4.96, and sMAPE of 0.27%. Random Forest Regression showed moderate performance with an RMSE of 6.56 and MAE of 5.19, while ARIMA underperformed significantly with an RMSE of 435.60 and MAE of 392.43. These findings highlight the potential of AI-driven forecasting models like LSTM to optimize resource allocation, reduce food waste, and improve the operational efficiency of SBFPs.