Abstract
Accurate forecasting of agricultural exports is essential for supporting production planning, trade management, and evidence-based policymaking. However, forecasting cocoa exports remains challenging because export volumes are influenced by interconnected environmental, climatic, and economic factors. This study proposes an integrated forecasting framework for Indonesian cocoa export volumes by incorporating remote sensing indicators, climatic indices, and economic variables into statistical, machine learning, deep learning, and hybrid forecasting approaches. The exogenous variables consist of Enhanced Vegetation Index (EVI), Land Surface Temperature (LST), rainfall, cocoa price, Oceanic Niño Index (ONI), Dipole Mode Index (DMI), and exchange rate. The study evaluates benchmark models (Naïve, Seasonal Naïve, and ETS), statistical time-series models (ARIMAX and SARIMAX), machine learning and deep learning models (XGBoost and LSTM), and a hybrid SARIMAX–XGBoost model. Model performance is evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) over an independent 17-month test horizon from November 2023 to March 2025. XGBoost achieved the overall best forecasting performance, obtaining the lowest test RMSE of 2,609,038 kg, MAE of 2,096,098 kg, and MAPE of 7.56%. Its advantage over ARIMAX, SARIMAX, and the SARIMAX–XGBoost hybrid was supported by pairwise comparisons using the modified Diebold–Mariano (DM) and Kolmogorov–Smirnov Predictive Accuracy (KSPA) tests. The comparison with LSTM showed no statistically significant difference according to the DM test, although the KSPA test indicated a significant difference. These findings demonstrate that XGBoost provided the strongest out-of-sample forecasting performance among the evaluated models.