Abstract

Timely and reliable estimates of subnational economic activities are critical for effective policymaking, investment targeting, and close monitoring of local-level practices and development. In the Philippines, producing timely provincial Gross Domestic Product (GDP) or the Provincial Product Account (PPA) remains a challenge due to dependence on local government data provisions and complexity of consolidating inputs from data sources and province statistical offices. This creates an information gap and lag for decision-makers seeking to address provincial disparities or respond to shifting granular dynamics, as well as local-level budget proposals and monitoring. The rapid expansion of high-resolution geospatial and community-sourced data, characterized by unprecedented volume, velocity, and variety offers new opportunities to bridge this gap.

This study integrates monthly Visible Infrared Imaging Radiometer Suite (VIIRS) nighttime-lights radiance, OpenStreetMap points of interest (POI), and annual population projections with official provincial GDP data (2018-2024) to construct a comprehensive panel covering 117 provincial and highly urbanized cities. A streamlined big data pipeline focused on geospatial processing, feature engineering, and scalable integration is developed. Six regression models: Linear, Ridge, Lasso, Support Vector, RandomForest, and XGBoost, each tuned via Bayesian optimization under a rolling TimeSeriesSplit (2019-2022), are evaluated on the held out 2023 data using mean absolute percentage error (MAPE), with SHapley Additive exPlanations (SHAP) analysis enabling global and local interpretability.

The framework demonstrates that combining remote-sensing and community-sourced data with state-of-the-art machine learning yields precise and interpretable forecasts of subnational GDP – highlighting the operational potential of alternative data for timely and policy-relevant measurement.