Implementing LSTM-Based Deep Learning for Forecasting Food Commodity Prices with High Volatility: A Case Study in East Java Province

Abstract

Accurate food price forecasting is essential for maintaining market stability and food security. East Java Province was selected as the study area because it is one of Indonesia’s main food production centers and a major contributor to national inflation. This study compares three deep learning architectures LSTM, Bi-LSTM, and hybrid CNN-LSTM to forecast the prices of four key food commodities (red chili, shallots, medium-grade rice, and beef) in East Java. Hyperparameter tuning was performed using grid search, and performance was evaluated using MAPE, MAE, and RMSE. The results show that the Bi-LSTM model consistently provides the best performance compared to LSTM and CNN-LSTM across the four analyzed commodities. Based on MAPE, MAE, and RMSE values, Bi-LSTM achieved the lowest forecasting errors for all commodities. The MAPE values of Bi-LSTM were 1.73% for red chili, 0.60% for shallots, 0.23% for medium-grade rice, and 0.08% for beef, all of which were lower than those of LSTM and CNN-LSTM models. These findings highlight Bi-LSTM’s bidirectional architecture, which leverages contextual information from both past and future data sequences, making it the most robust and effective model for forecasting food prices under varying volatility. The study provides practical insights for policymakers and supply chain stakeholders in supporting price stability and food security.