Forecasting performance metrics for the best-performing neural network (SRNN for γ = 7, IRNN for γ ≥ 14) compared with Dante.

The NNs are trained using search query frequencies generated only up to the last available ILI rate (the 2-week advantage of using Web search data is removed). We use leave-one flu season-out to train models, similarly to Dante. The best results for this comparison are shown in bold. The very last column (NNb) presents the average performance results of NNs where the temporal advantage of Web search activity information is maintained (see also S10 Fig that depicts IRNN’s forecasts when leave-one flu season-out is applied). The penultimate column (NNa) holds results for the same experiment as NNb with the addition of disabling leave-one flu season-out training.