Model-based ensembles: Lessons learned from retrospective analysis of COVID-19 infection forecasts across 10 countries.

Mathematical models of different types and data intensities are highly used by researchers, epidemiologists, and national authorities to explore the inherently unpredictable progression of COVID-19, including the effects of different non-pharmaceutical interventions. Regardless of model complexity, forecasts of future COVID-19 infections, deaths and hospitalization are associated with large uncertainties, and critically depend on the quality of the training data, and in particular how well the recorded national or regional numbers of infections, deaths and recoveries reflect the the actual situation. In turn, this depends on, e.g., local test and abatement strategies, treatment capacities and available technologies. Other influencing factors including temperature and humidity, which are suggested by several authors to affect the spread of COVID-19 in some countries, are generally only considered by the most complex models and further serve to inflate the uncertainty. Here we use comparative and retrospective analyses to illuminate the aggregated effect of these systematic biases on ensemble-based model forecasts. We compare the actual progression of active infections across ten of the most affected countries in the world until late November 2020 with "re-forecasts" produced by two of the most commonly used model types: (i) a compartment-type, susceptible-infected-removed (SIR) model; and (ii) a statistical (Holt-Winters) time series model. We specifically examine the sensitivity of the model parameters, estimated systematically from different subsets of the data and thereby different time windows, to illustrate the associated implications for short- to medium-term forecasting and for probabilistic projections based on (single) model ensembles as inspired by, e.g., weather forecasting and climate research. Our findings portray considerable variations in forecasting skill in between the ten countries and demonstrate that individual model predictions are highly sensitive to parameter assumptions. Significant skill is generally only confirmed for short-term forecasts (up to a few weeks) with some variation across locations and periods.