<![CDATA[Quarterly tourism-related output is seasonal and highly vulnerable to structural disruptions, so models fitted to stable historical patterns may fail after shocks. This study detects changepoints and compares forecasting models for Indonesia's real GDP in Accommodation and Food Service Activities, used as a proxy for tourism-related economic output. The dataset comprised 50 quarterly observations at constant 2010 prices from 2013Q1 to 2025Q2. PELT was applied to detrended, seasonally adjusted log-training residuals. Ordinary SARIMA, SARIMA models with pulse and step interventions, six exponential-smoothing specifications, and a seasonal-naive benchmark were screened using training AICc, 16 rolling one-step validation origins, and an untouched ten-quarter fixed test. PELT identified 2017Q2, 2020Q2, and 2022Q1 as changepoints, with 2020Q2 remaining the most robust under stronger penalties. SES achieved the smallest rolling-validation RMSE (6,283.9), while the pulse-plus-temporary-step SARIMA intervention was best within its family (RMSE 6,917.4). On the fixed 2023Q1-2025Q2 test, SARIMA (0,1,1) × (0,0,0)₄ performed best overall, with MAE of Rp5,301.9 billion, RMSE of Rp6,131.6 billion, sMAPE of 5.23%, and MASE of 1.016. For the selected baseline SARIMA, residual autocorrelation and ARCH effects were not significant; the intervention model, however, retained significant residual autocorrelation. Residuals were non-normal across the retained models, and the baseline SARIMA underpredicted the later recovery. PELT provided structural diagnosis, but explicit intervention effects did not guarantee superior holdout forecasts. Parsimonious SARIMA provides the most defensible baseline, supplemented by changepoint monitoring and scenario-based intervention analysis.]]>
Copyrights © 2026
