Accurate prediction of pollutant concentration is very important part in any air quality management program (AQMP). The conventional time series modelling techniques like ARIMA has showing poor prediction and forecasting, as air quality is non-linear and complex phenomenon. Although neural networks have been applied for prediction of air quality data in the previous studies, the model performance was very poor as they don’t consider the data as time series in their algorithms. Combining the aspects of both neural networks and time series analysis, Nonlinear Autoregressive models with exogenous input (NARX) based neural networks were found to predict chaotic time series better because of better learning and faster convergence than the conventional neural network algorithms. In the present work, meteorological and traffic parameters near busy urban corridors were used to train NARX based neural network model for the prediction of ambient air quality. Diagnostic analysis between different model variables was done to understand the relationship between one other. The developed model predicted NOx and SO2 concentrations with a very good performance over the entire dataset. © 2018, Springer International Publishing AG, part of Springer Nature.