<![CDATA[Global economic uncertainty, trade-regime shifts, and supply-chain disruptions have made export-import trade-value forecasting increasingly complex. This study compares the performance of Random Forest, Extra Trees Regression, and SARIMA in predicting monthly trade values of HS commodities in the apparel and footwear sector, covering HS 61, HS 62, HS 63, and HS 64. The dataset was obtained from Indonesia?s Central Bureau of Statistics (BPS) and organized as a monthly time series using a leakage-safe workflow through a time-based train-validation-test split. The modeling stage employed 19 predictive features consisting of historical, local statistical, calendar-seasonal, and exogenous variables, and each model was tuned on the validation set before being evaluated on the holdout test. Performance was assessed using MAE, RMSE, MAPE, sMAPE, and wMAPE, with MAPE as the primary ranking metric. The main contribution of this study lies in providing a fair and replicable comparison of three forecasting models under a time-based evaluation protocol for HS commodity trade data, making the model selection results more representative of real implementation settings. The results show that Random Forest achieved the best MAPE at 22.7746%, slightly outperforming Extra Trees Regression at 22.9469%, while SARIMA recorded 28.9794%. These findings indicate that tree-based ensemble models are more adaptive to volatile trade data, whereas SARIMA remains relevant as a statistical baseline for structured seasonal patterns.]]>
