Figure 5. Prediction performance with training data from the first six seasons and validation on the last three.

Plate A shows the correlation between the observed incidence and the model-predicted response. We show significant positive correlation, particularly within the trigger periods, between the model predictions and the actual held-out data. This gives us confidence to construct ROC curves for each week. Plates B-D show the ROC curves for the last three weeks of each of the three seasons in the out-of-sample period (potentially, these computations can be repeated for all possible partitions of study weeks into training and test samples). Plates E-G illustrate that the normalized decision variable, which is the normalized response from the model, identifies the South and Southeastern counties as the trigger zones.