The Impact of Model Retraining Frequency on Predictive Performance in Air Pollution Forecasting

This dataset contains the full numerical results supporting the paper "The Impact of Model Retraining Frequency on Predictive Performance in Air Pollution Forecasting". It provides model evaluation metrics R², Root Mean Square Error (RMSE) for a LightGBM bias-correction model (LGBM BCM) trained on NASA GEOS-CF atmospheric reanalysis data to predict surface PM2.5 concentrations at 10 urban monitoring sites across Latin America, Africa, and Southeast Asia. Sites covered: Bogotá (Colombia), Jakarta (Indonesia), Kampala (Uganda), Hanoi (Vietnam), Guatemala City (Guatemala), Buenos Aires (Argentina), São Paulo (Brazil), San José (Costa Rica), Lagos (Nigeria), Tijuana (Mexico), and Lima (Peru). Dataset structure: The Excel workbook contains four sheets: Hourly PM2.5 — R², RMSE, and NSE for all combinations of baseline training duration (6–24 months) and retraining window (6–24 months) across all sites, evaluated on hourly-averaged PM2.5 predictions. Daily PM2.5 — R² and RMSE for daily PM2.5 predictions across sites with sufficient temporal coverage. % Improvement (LGBM BCM) — Percentage improvement in R² and RMSE of the bias-corrected model relative to the raw GEOS-CF output, across all baseline and retraining window combinations. Summary — Mean and standard deviation of R² and RMSE improvement aggregated by site and by retraining interval, providing a cross-location overview of retraining frequency effects. Variables: R²: Coefficient of determination between predicted and observed PM2.5 RMSE:Root Mean Square Error (µg/m³) Baseline duration: Length of initial training period (months) Retraining window: Length of data used in each subsequent model update (months) Data sources: Ground-level PM2.5 observations were sourced from AirNow and regional monitoring networks. Atmospheric predictor variables were derived from the NASA GEOS Composition Forecast (GEOS-CF) model, accessed via NASA OSS. Intended use: These results are intended to allow readers to reproduce the figures and statistical comparisons in the associated paper, and to support future benchmarking of adaptive retraining strategies for air quality forecasting models in data-sparse urban environments.