We calculated a derived variable called Epidemic Indicator, given by the sum of the number of patients in Intensive Care Unit (ICU) with the number of daily deaths. We noticed that this variable follows a regular descending trend, after the peak registered on April 1st. In fact, it is an easily measurable information which doesn't suffer from sampling methodology or weekly fluctuations. It is certainly not perfect since it can also be prone to underestimation, mostly during the acute peak of the crisis where the health system was stressed and hospitals saturated. But after that critical phase, the Epidemic Indicator should reflect pretty well the number of patients that are affected most severely by the virus

The main assumption of our analysis is that Epidemic Indicator can be a reliable variable for estimating the current epidemic progression and for modeling future trends. In fact, after April 1st this variable is decreasing at a constant rate. For this reason, we can model the future trend of the Epidemic Indicator using Linear Regression, one of the most popular classical Machine Learning algorithms for supervised learning. This algorithm is relatively easy to implement and works well when the relationship between covariates and response variable is known to be linear (in our case: Epidemic Indicator VS time). A known disadvantage is that Linear Regression over simplifies many real world problems. We separated the dataset in 20 regional subsets

Then, we created and trained a Linear Regression Model for each region. Trained models are then used for predicting the end of the epidemic, expressed by the number of days remaining before Epidemic Indicator is expected to be 0. Of course, not all regional models performed the same. Some regions have clear recovering trends (i.e. Lombardia), while others seem to have a less linear evolution (i.e. Lombardia). For this reason, every model also returns a performance score, the coefficient of determination R² of the prediction. An R² of 1 indicates that the regression predictions perfectly fit the data. therefore the closer the value gets to 1, the more we can trust our model and its predictions

More technical information about the model can be found at this link