Abstract
Continuous monitoring stations are the sensing layer of urban air-quality management, yet their records are often used to infer pollution “drivers” without controlling for weather. We ask what predicts fine particulate matter (PM2.5) in Navi Mumbai, a coastal Indian city, using screened daily data from five stations for 2021–2025, hourly data from two, and ERA5 reanalysis weather checked against Mumbai surface and radiosonde observations. Vector autoregressions with exogenous weather and calendar terms were estimated without imputation, together with hourly models, a stagnation index and out-of-sample forecasts. Lagged co-pollutants add little to daily prediction: equivalence tests bound six of seven effects within ±5% per standard deviation, while lagged ozone, an oxidant signal, raises PM2.5 by 4%. Omitting weather, or using PM10, manufactures co-pollutant “drivers”. Rainfall and boundary-layer height lower same-day PM2.5 by 10–24% per standard deviation, and date effects shared by all stations explain 78% of the variance, three-quarters of it weather. A stagnation index raises the frequency of days above the national standard two- to fourfold and, combined with persistence, warns as skillfully as an extreme gradient boosting (XGBoost) forecaster, which itself does not beat persistence. Monitoring data support citywide, meteorology-triggered warnings better than single-pollutant controls inferred from comovement.