Comparative Evaluation of Forecasting Methods for Tourist Arrival Prediction: A Sliding Window-Based Analysis

The tourism sector in West Nusa Tenggara (NTB) plays a strategic role in driving regional economic growth. However, its management still faces challenges in data-driven planning, particularly in accurately forecasting tourist arrivals. This issue is further complicated by the seasonal and volatile nature of tourist visit patterns, which are highly susceptible to external disruptions such as pandemics. This study initiates the development of a deep learning-based forecasting system to support the implementation of smart tourism in NTB. This study evaluates and compares the performance of three time series forecasting methods—SARIMA, Prophet, and XGBoost—using a sliding window approach to assess the temporal stability of their predictive performance. The analysis uses monthly international tourist arrival data from 2010 to 2024. The experimental results reveal that the SARIMA(1,0,2)(0,1,1,12) model provides the most stable accuracy, with an average MAPE of 35.22%, making it suitable for macro-level planning. The XGBoost model achieved the lowest MAPE of 29.84%, although it exhibited greater variability across windows. In contrast, the Prophet model demonstrated high sensitivity to data anomalies, particularly during the pandemic period. These findings suggest that classical statistical models like SARIMA remain relevant in handling periodic and limited datasets but have limitations in capturing complex patterns that may be better modeled through deep learning approaches.