Assessing the quality of forecasting quantitatively using model confidence set: from naive models to neural networks

In this work, we evaluated a collection of models under the lens of Model Confidence Set, an approach that allows to test whether differences in forecasting model performance are statistically significant. The data used for the analysis comes from sensors that detect pollutants in one of the most polluted areas of Italy and Europe: the Po Valley and were collected from 2018 to 2023. We focused on forecasting PM10 levels by training the models on its previous values and other pollutants. Hence, different versions of the data set were created in order to evaluate the effect of missing data handling strategies. Depending on the data set used, the core group of best models differed. However, a common group of models existed across all versions. The results showed that, despite the use of more complex models, those consistently appearing in the set with the best performance were mostly simple models such as elastic net or random forest.