The present study introduces the novel Bayesian approach to solve a difficult problem that modellers face when doing linear regression for air quality prediction; i.e. the non-uniqueness of the input variables and the functional forms to be selected. Using the historical information of ozone, nitrogen dioxide, respirable particulates and eight meteorological elements recorded during the high ozone seasons (May-October) in Macau between 2006 and 2007 as the training data, the Bayesian approach was applied for model selection from 16383 candidates. The model with the best efficiency-robustness balance selected is not the most complicated one. It was then examined against the most complicated model with the data between 2008 and 2009. Results show that the selected model yields better performance with the root-mean-squared error (RMSE) and the coefficient of determination (r2) equal to 19.18 µg/m3 and 0.81, respectively. In addition, it is capable to capture 87% of the ozone episodes with tolerable false alarm percentage of 35%. Therefore applying the Bayesian approach to improve the development process of an operational air quality forecasting model, particularly on model selection, is successful.