Figure 1.

Relative importance of predictor variables in machine learning model forecasts of COVID-19 growth rates. The colour saturation of weekly blocks reflects the relative importance of predictor variables in forecast models trained on the preceding two months of data. Importance is defined as the decrease in model prediction accuracy when the variable in question is omitted from the model. An elastic net model was fitted to select the best predictors for a given time window. The selected variables were then used to build a random forest model, to assess the relative importance of each variable while accounting for non-linear effects. The response variable, COVID-19 growth rate aggregated at the global level, is plotted in the lower panel for reference.