Forecasting Tourist Arrivals in Bali: A Grid Search-Tuned Comparative Study of Random Forest, XGBoost, and a Hybrid RF-XGBoost Model

Abstract

Tourism planning, infrastructure growth, and economic stability. This study presents an extensive comparative evaluation of Random Forest (RF), Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and a novel Hybrid RF-XGBoost model in the prediction of monthly international tourist arrivals. A full time series dataset of a ten-year period (2014–2024) from the Central Bureau of Statistics of Bali was used for training and testing the models. Hyperparameter optimization using Grid Search with cross-validation (Grid Search CV) was used for all the machine learning models to obtain best predictive performance. Two robust metrics, Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE), were used to assess forecasting accuracy. Results show that the Random Forest model outperforms all competitors with lowest RMSE (41,772.68) and MAPE (6.30%), indicating high forecasting precision and robustness, especially during structural breaks such as the COVID-19 pandemic. The hybrid model also performs well, with LSTM indicating higher error rates, illustrating its shortcomings on small-to-medium-scale tourism time series. Besides, the study provides six-month ahead predictions (January–June 2025) with 95% prediction intervals, showing an ongoing trend of recovery. The findings affirm the superiority of bagging-based ensemble methods over polynomial-based methods in capturing nonlinearity, seasonality, and exogenous shocks in tourist demand. The study plugs the growing amount of data-driven tourism analytics by offering a reproducible, high-precision forecasting model for developing countries and seasonally driven destinations.