Perbandingan Motode Long Short-Term Memory (LSTM) dan Hybrid LSTM-XGBoost dalam Prediksi Harga Komoditas Sektor Pertanian di Sulawesi Tenggara

Abstract

Agricultural commodity price instability remains a major challenge for farmers, fishers, and livestock breeders in Southeast Sulawesi Province, where 54,905 agricultural workers were still categorized as poor in 2023. Frequent price fluctuations complicate production and marketing decisions, creating the need for accurate price prediction systems. This study develops and compares two prediction models: a single Long Short-Term Memory (LSTM) model and a Hybrid LSTM-XGBoost model for 20 agricultural commodities across 17 regencies/cities in Southeast Sulawesi. Daily price data were collected from the Southeast Sulawesi Food Security Agency from January 2023 to December 2025. The LSTM model consisted of two LSTM layers (128 and 64 units), a Bahdanau Attention mechanism, and an Embedding layer to represent regions and commodities. The XGBoost component was used to correct LSTM residuals within a 30-day prediction horizon. Model performance was evaluated using Mean Absolute Percentage Error (MAPE). Results showed that the Hybrid LSTM-XGBoost model achieved better accuracy, with an average MAPE of 4.73% compared to 5.52% for the single LSTM model. Significant improvements were found in volatile commodities such as Bird’s Eye Chili and Curly Red Chili. The system can support production planning, food distribution, and SDG 1 (No Poverty).