<![CDATA[Chili prices in Indonesia are highly volatile due to seasonal production, fragile supply chains, and shocks in public perception. This study improves short-run forecast accuracy by adding public-attention signals (Google Trends and news volume) as exogenous features summarized in a Shock Index. Evaluation metrics are sMAPE (primary), RMSE, and MASE; hyperparameters are tuned via Bayesian HPO. Empirically, the attention-augmented configuration (S4: +Trends +News +Shock) is best. Post-HPO (average across horizons), S4 attains sMAPE 12.47%, RMSE 3,433 IDR/kg, and MASE 0.87. By horizon, S4’s sMAPE is 9.8% (H=1), 12.0% (H=2), 15.6% (H=4); RMSE 2,550/3,350/4,400 IDR/kg; MASE 0.78/0.86/0.96. Compared with the price-only (S1) baseline, S4 is already better pre-tuning and becomes even stronger after HPO (average sMAPE reduction ≈ −6.2% relative). These findings show that incorporating the intensity of public issues enhances predictive value—especially at longer horizons when uncertainty rises—and that the approach is ready for operational use in nowcasting and early-warning.]]>
Copyrights © 2026
