<![CDATA[Indonesia's international tourism sector, one of the country's foremost foreign exchange contributors, suffered a near-total collapse when the COVID-19 pandemic took hold in early 2020, with arrivals plunging by approximately 75% relative to 2019's baseline. The recovery that followed was highly non-linear shaped by uneven border reopening timelines, shifting travel restrictions, and residual traveler hesitancy introducing volatility patterns that stretch far beyond the bounds of normal seasonal fluctuation. This study examines how two methodologically distinct forecasting approaches, the Autoregressive Integrated Moving Average (ARIMA) and Random Forest (RF), perform under these conditions, using 72 months of official BPS Indonesia data spanning January 2019 to December 2024, with 2019–2023 as the training window and 2024 reserved as the holdout test period. ARIMA modeling followed the Box-Jenkins procedure, yielding an optimal SARIMA(1,1,1)(1,1,1)12 specification. The Random Forest model was built upon 18 engineered autoregressive lags, rolling window statistics, and cyclical month encodings with hyperparameters tuned through GridSearchCV. Empirical validation on the 2024 test set revealed that the parametric ARIMA model significantly outperformed the machine learning approach, achieving a MAPE of 5.49% compared to 8.66% for Random Forest, representing a 36.6% reduction in relative percentage error, alongside substantial improvements in MAE (42.7%) and RMSE (21.9%). Feature importance analysis indicated that short-term temporal dependencies (lag_1 and lag_2) heavily dominate the tree ensemble, which fundamentally limited its capability to extrapolate the expanding post-pandemic growth trend. These results suggest that classical parametric frameworks with proper regular and seasonal differencing offer more reliable mid-term demand estimates for tourism planners and policymakers navigating structural recovery trajectories.]]>
