The impact of model retraining frequency on predictive performance in air pollution forecasting

High-resolution PM 2.5 forecasts are increasingly produced with machine-learning models, yet practical guidance on how often these models should be retrained and validated remains limited. This study quantifies the impact of retraining frequency and bias correction on air-quality prediction skill across multiple cities with contrasting emission sources and meteorological regimes, using The Goddard Earth Observing System composition forecast (GEOS-CF) fused with in-situ observations. Site-specific models are trained on at least two years of hourly data, with bias correction and alternative update schedules (6–18-month baselines and 6–12-month retraining cycles) evaluated using RMSE, R 2 , and SHAP-based feature importance. Bias-corrected models consistently improves GEOS-CF forecasts by more than 107% in R 2 and reduces RMSE by over 75%, with annual retraining providing the largest gains (13% increase in R 2 , 12% reduction in RMSE) relative to more frequent updates. SHAP analysis shows that the dominant predictors and their relative importance vary by city, with combinations of boundary-layer height, aerosol optical depth, humidity, wind, and nitrogen oxides driving PM 2.5 levels, demonstrating that a single global pre-trained model is inadequate and that locally tuned models are required. Together, these results define minimum data requirements, preferred retraining intervals, and the need for site-specific bias-corrected models, offering concrete design rules for operational PM 2.5 forecasting systems.