Fig. 3.

Forecasting benchmark for different horizons separated by year. The model evaluation has been split by year, represented by six sub-panels as indicated by the subfigure titles. The dashed and dotted lines show the model error of the Naïve Seasonal Model and the Naïve Average Seasonal Model, respectively. Predictions are available for each of the 21 days on the forecasting horizon. For simplicity, the plots visualize the errors only for three different time points: the maximum (21 days) and 2 intermediate time points (7 and 14 days). DeepAR and the Temporal Fusion Transformer (TFT) have also been applied for multivariate forecasting experiments. These include covariates describing the age groups (AGE), laboratory component (LC), or both (AGE + LC)