Tourism Demand Forecasting Using Machine Learning with Cross-Model Feature Ranking

This study proposes a machine learning (ML) framework for forecasting tourist visits to a destination, leveraging feature ranking consistency across heterogeneous models. The proposed framework uses a unified stepwise feature selection pipeline to systematically determine a subset of given features that reduces prediction error of an ML model. The transferability of feature rankings is demonstrated through experiments using ML models including Random Forest, XGBoost, LightGBM, and Support Vector Regression (SVR), and a dataset related to tourist visits to West Sumatra, Indonesia. Experimental results show that i) SVR using top four features in the feature ranking based on SHAP achieves the highest accuracy and ii) application of the feature ranking based on XGBoost improves performance stability across other models, reducing error fluctuations. These findings reveal that an optimized feature subset combined with cross-model ranking transfer can outperform more complex ensembles, highlighting the importance of ranking robustness in tourism forecasting. The proposed framework provides a methodological foundation for building interpretable and dataefficient decision support systems for tourism management and policy planning.