Understanding PM₂.₅ predictability across diverse urban environments: A comparative evaluation of statistical forecasting models in four Indian cities

Air pollution is a major global environmental challenge with serious implications for public health and environmental sustainability. Reliable air quality forecasting systems are essential for providing early warnings, supporting pollution mitigation strategies, and informing policy decisions. This study evaluates the performance of three statistical time-series forecasting models—Auto Regression (AR), Moving Average (MA), and Auto Regressive Integrated Moving Average (ARIMA)—for forecasting PM₂.₅ concentrations in four Indian cities (Pune, Delhi, Ahmedabad, and Mumbai) representing diverse climatic conditions, pollution characteristics, and geographical settings. Forecasting performance was evaluated for 1-day, 3-day, and 7-day prediction horizons using historical datasets of varying lengths (1-year, 2-year, and 4-year). Among the models investigated, ARIMA consistently produced the lowest Root Mean Square Error (RMSE) across all cities and forecasting horizons. The results further indicate that using longer historical datasets generally improves forecasting performance, although the extent of improvement varies among cities. For the 3-day forecasts, RMSE values decreased from 11.32 to 11.06 µg m⁻³ for Pune, from 33.76 to 26.17 µg m⁻³ for Delhi, from 22.64 to 22.54 µg m⁻³ for Ahmedabad, and from 12.42 to 9.21 µg m⁻³ for Mumbai when the training dataset was increased from one year to two years. Forecast accuracy decreased with increasing prediction horizon, while cities exhibiting relatively stable PM₂.₅ concentrations, such as Mumbai and Pune, achieved lower prediction errors than Delhi and Ahmedabad, where PM₂.₅ concentrations showed greater temporal variability. Overall, the findings demonstrate that statistical time-series models, particularly ARIMA, provide a computationally efficient, cost-effective, and scalable approach for short-term PM₂.₅ forecasting and offer a practical alternative for operational air quality forecasting in cities with limited computational and monitoring infrastructure.