A Mini Hybrid SARIMAX–LSTM Framework for Spatiotemporal Tourism Forecasting

A Mini Hybrid SARIMAX–LSTM Framework for Spatiotemporal Tourism Forecasting

This work proposes a lightweight, production-oriented hybrid pipeline that combines SARIMAX (to capture linear/seasonal structure and exogenous effects) with an LSTM + attention model (to learn non-linear residual dynamics) for tourism-demand forecasting. Using monthly arrivals (2010–2024), the hybrid approach outperforms Seasonal Naïve, standalone SARIMAX, and standalone LSTM baselines across multiple error metrics.

Why it matters. Accurate tourism forecasts support staffing, dynamic pricing, marketing allocation, and infrastructure planning. The proposed pipeline is both interpretable (via SARIMAX) and expressive (via LSTM-attention), and is deployable with modest compute.

Methods (brief).

• Data & preprocessing: monthly arrivals (2010–2024); seasonal Kalman imputation for missing values; winsorization for outliers; GDP/CPI as exogenous regressors.

• Stage 1 — SARIMAX: orders selected via AIC/BIC; residuals extracted.

• Stage 2 — LSTM + attention: 12-month lookback on residuals to model non-linear structure.

• Uncertainty: conditional heteroskedasticity modeled with GARCH to produce tighter, realistic forecast bands.

• Demo: interactive React/TypeScript + Leaflet choropleth to explore forecasts.

Results (summary).

Hybrid model achieved RMSE 120.5, MAE 98.3, sMAPE 12.1%, improving ~15% vs. the Seasonal Naïve baseline (sMAPE). Standalone SARIMAX and LSTM underperformed the hybrid on all metrics.

Resources.

• Live demo: https://noelframil.github.io/spain-tourism-map/