Forecasting the Spread of the COVID-19 Epidemic in Lombardy: A Dynamic Model Averaging Approach

Gianmoena, L.; Rios, V.

Forecasting with accuracy the evolution of COVID-19 daily incidence curves is one of the most important exercises in the field of epidemic modeling. We examine the forecastability of daily COVID-19 cases in the Italian region of Lombardy using Dynamic Model Averaging and Dynamic Model Selection methods. To investigate the predictive accuracy of this approach, we compute forecast performance metrics of sequential out-of-sample real-time forecasts in a back-testing exercise ranging from March 1 to December 10 of 2020. We find that (i) Dynamic Model Averaging leads to a consistent and substantial predictive improvements over alternative epidemiological models and machine learning approaches when producing short-run forecasts. Using estimated posterior inclusion probabilities we also provide evidence on which set of predictors are relevant for forecasting in each period. Our findings also suggest that (ii) future incidences can be forecasted by exploiting information on the epidemic dynamics of neighboring regions, human mobility patterns, pollution and temperatures levels.