Machine Learning and Air Pollution: an Empirical Application for Northern Italy

Air pollution is a serious threat for almost all European countries. Recently in Northern Italy PM10, PM2.5 and NO2 limits set by the EU were often overtaken. In order to fight this trend, it is crucial to analyse air pollutants. This work uses two machine learning techniques (Natural Cubic Splines and Bagging) to predict PM10 and NO2 daily concentration for the years 2016 to 2019. We assess the methods accuracy by means of the Root Mean Square Error, the Weighted Mean Absolute Percentage Error and the Pearson’s correlation between predictions and observed values. Bagging slightly outperforms Natural Cubic Splines and per-formances are better for NO2 than for PM10. We are able to provide estimates even when actual on field measurements are not available. Moreover, we could use these models to forecast pollutant levels for the next few days by plugging in estimates of our predictors. This approach provides national and local govern-ments with tools to study the air pollution problem, shifting from ex-post meas-urements and sanctions to ex-ante strategies and corrections.