Regime-Gated Explainable AI for Philippine Tourism Source-Market Decisions: Evidence from a 2008–2023 Multi-Source Panel

Abstract

International-arrival forecasts are often evaluated as if one model should remain optimal across normal demand, border closure, and recovery. That assumption is consequential for the Philippines, where source markets recovered at sharply different rates after COVID-19. This study develops a regime-gated explainable artificial intelligence (XAI) framework for source-market planning. Monthly visitor-arrival reports from the Philippine Department of Tourism were harmonized with World Development Indicators for 14 markets from January 2008 to October 2023, producing 2,646 market-month observations and covering 84.46% of 2019 arrivals. Persistence, seasonal-naive, ridge, random-forest, extra-trees, and gradient-boosting variants were evaluated with strictly temporal holdouts for 2019, 2022, and 2023. Exact grouped baseline Shapley values translated the best stable-regime model into auditable market-level explanations, while a recovery-volume matrix converted the outputs into portfolio actions. In 2019, residual Extra Trees achieved 7.25% weighted absolute percentage error (WAPE), versus 13.13% for persistence; a month-cluster bootstrap placed the improvement at 5.88 percentage points (95% CI: 3.56–8.47). The ranking reversed in transition: persistence outperformed the same model in 2022 (23.87% versus 32.30%) and 2023 (15.28% versus 28.25%). A transparent gate that used machine learning only in stable demand and persistence during transition reduced mean fold WAPE from 17.43% to 15.47%. Shapley analysis showed that prior demand level dominated stable forecasts, followed by recent momentum, seasonality, and market identity. The novelty is not another claim that AI universally improves tourism forecasting; it is an empirically validated, explanation-linked rule for deciding when AI should and should not be trusted, combined with a Philippine source-market decision portfolio.