Prediksi Harga Beras Tingkat Perdagangan Besar Menggunakan Algoritma Extreme Gradient Boosting Berbasis Indikator Ekonomi Makro

Abstract

Wholesale rice prices are an important indicator for food-price monitoring because their movements affect supply chains, distribution actors, and household purchasing power. This study develops an Indonesian wholesale rice-price forecasting model using Extreme Gradient Boosting (XGBoost) with macroeconomic indicators and temporal features. Official secondary data from Statistics Indonesia were collected from January 2018 to May 2026, comprising 101 monthly observations. The predictor variables include general inflation, rice production, harvested area, Farmers’ Terms of Trade for food crops, population, and the food Consumer Price Index. In addition, lag-1, lag-3, and a three-month moving-average feature were constructed from historical rice-price observations. The dataset was divided chronologically into 84 training observations and 17 testing observations. Hyperparameter optimization was performed using Grid Search with TimeSeriesSplit on the training data. XGBoost performance was compared with ARIMA(1,1,1) and Random Forest Regression using RMSE, MAE, MAPE, and R². The test results show that XGBoost achieved an RMSE of IDR 234.17/kg, an MAE of IDR 189.43/kg, a MAPE of 1.63%, and an R² of 0.9621. These results outperform ARIMA and Random Forest on the same test set. Feature-importance analysis indicates that the lag-1 price, three-month moving average, and Farmers’ Terms of Trade provide the largest predictive contributions. The findings indicate that multivariate XGBoost has potential as a prototype for data-driven wholesale rice-price monitoring.