AI-Based Forecasting of National Tourism Revenues: Integrating Economic, Fiscal, Political, and Environmental...

This study proposes an advanced framework for forecasting national tourism revenues by systematically comparing machine learning (ML), deep learning (DL), and hybrid architectures on a country-year panel. Baseline models using only trade and economic indicators have limited explanatory power, whereas adding fiscal, political, and environmental variables substantially improves accuracy. Among ML methods, LightGBM performs best; among DL models, the Transformer excels by capturing nonlinear interactions and temporal dependencies. Building on these results, we introduce a hybrid residual boosting model that integrates the Transformer's predictive strength with LightGBM's structural interpretability. The hybrid model outperforms single models across mean absolute error, root mean square error, mean absolute percentage error, and R², simultaneously minimizing errors and maximizing explanatory power. Methodologically and theoretically, the framework advances tourism economics while offering policymakers actionable guidance on fiscal planning, political stability, and environmental sustainability. Importantly, the empirical results are correlational and reflect predictive associations; they should not be interpreted as causal effects of policy interventions. Methodological novelty lies in a regression-oriented, two-stage residual-boosting design that (1) learns a Transformer as the primary forecaster on the country-year panel, (2) fits LightGBM to the Transformer residuals to correct systematic errors under distributional heterogeneity, and (3) yields a decomposed forecast (base + residual correction) that facilitates transparent error attribution beyond prior DL-DL stacking hybrids. Importantly, the reported relationships are associational and derived from predictive modeling; they should not be interpreted as causal effects of policy levers.