A FORECASTING PROTOCOL FOR URBAN AIR POLLUTION IN A TRANSPORT INFLUENCED ENVIRONMENT: A COMPARISON OF UNIVARIATE, SEASONAL, MULTIVARIATE AND DYNAMIC REGRESSION MODELS FOR PM2.5, PM10, O3, AND NO2

Abstract

This study proposes a structured and empirically validated forecasting protocol for the short-term prediction of weekly urban air pollution time series, focusing on 𝑃𝑀. , � �𝑀 , 𝑂, and 𝑁𝑂 in a transport-influenced urban environment. Rather than assuming the superiority of a specific modeling approach, the framework applies a diagnostic-driven and out-of-sample evaluation strategy for systematic model selection. It combines statistical diagnostics—including stationarity testing, autocorrelation analysis, and seasonality detection—with benchmark comparisons and rolling one-step-ahead validation under realistic operational conditions. The modeling framework includes univariate ARIMA models, seasonal SARIMA extensions, multivariate VAR systems, and selective dynamic regression models (SARX), allowing the assessment of linear, seasonal, and interdependent pollutant dynamics. The empirical analysis used 200 consecutive weekly observations, with the final 10 reserved for out-of-sample testing. Forecast accuracy was evaluated using MAE, RMSE, MAPE, and sMAPE, complemented by Diebold–Mariano tests. The results show that no single model dominated across all pollutants: SARIMA achieved the most favorable comparative performance for 𝑃𝑀. and 𝑃𝑀 , while VAR and SARX occupied the leading positions for � �𝑂, and 𝑂 shows weak and unstable predictability. These findings indicate that model performance depends on the statistical structure of the individual pollutant series rather than a universally superior modeling approach. Although the analysis indicates temporal relationships among pollutants that may be relevant to transport-related air-quality dynamics, it does not establish direct causal effects of transport activity. Overall, the proposed protocol provides a transparent and reproducible framework for selecting and comparing forecasting models and for supporting air-quality and urban transport management.