<![CDATA[Accurate trade-value forecasts are increasingly required to be not only precise but also interpretable, so that model outputs can be audite d before informing trade and logistics decisions. This study extends a previously validated leakage-safe forecasting framework for Indonesia's apparel and footwear commodities (HS 61, HS 62, HS 63, and HS 64) by adding formal feature attribution to the best-performing model. Monthly trade values from Statistics Indonesia (BPS) covering January 2017 to March 2026 were modeled using nineteen predictors spanning historical lags, rolling statistics, calendar-seasonal signals, and exogenous variables, under a strict chronological train-validation-test protocol. Random Forest, the strongest model in the prior comparison with a holdout MAPE of 22.7746 percent, ahead of Extra Trees Regression at 22.9469 percent and SARIMA at 28.9794 percent, is interpreted using permutation importance and SHapley Additive exPlanations (SHAP). Attribution results show that contribute most to out-of-sample predictions, while exogenous variables. The framework demonstrates that accuracy and transparency can be obtained jointly on official trade statistics.]]>
Copyrights © 2026
