Closed ktindiana closed 9 months ago
Another subset could be the "max" (e.g. highest probability, highest peak flux) that matches an event. When monitoring the scoreboard I tend to pick those out.
Other subsets for exploring the correct rejections and false alarm performance might tell us what is going on for each model, the source and explanation for the metrics.
This is not clear. What kind of subset would give more insight?
This feature request will be considered the implementation of metrics for:
Implemented feature in the code to calculate First, Last, Max, Mean forecast assessment when appropriate for different quantities. Decided not to include Min.
The "First" feature applied to All Clear results in a contingency table with one forecast for each observed SEP event that fell inside a model's prediction window. This indicates whether a model hit or missed a given SEP event. This contingency table cannot assess the Correct Negatives.
Metrics calculated for First, Last, Max, or Mean are only relevant for times when an SEP event was observed. All forecasts outside of an observed SEP event period are not included. They should be interpreted as - "When an observed SEP event occurred, the model performed..."
From Leila: Metrics for one forecast per event (earliest and/or best and/or latest) which gives different information and more readily understandable regarding performance for hits and misses - and I think this is most relevant to what the operator/analyst sees in real-time. Other interesting metrics could be found for different types of forecast subsets, such as dropping the "not clear" subset for UMASEP and other ongoing forecasts for events. Other subsets for exploring the correct rejections and false alarm performance might tell us what is going on for each model, the source and explanation for the metrics.