Skip to Main Content
We test the ability of a data-derived model of geomagnetic activity, originally optimized to have a high prediction efficiency (PE), for its ability to predict only large geomagnetic disturbances. Correlation-based metrics, such as prediction efficiency, are often used as a measure of model performance. This metric puts equal weight on prediction of both large and small measurements. However, for space weather purposes, one is often interested in knowing only if a large disturbance event will occur so less emphasis should be placed on small measurements. If only large events are of interest, then a correlation metric is not the best measure of model performance. In this work, we determine how well a data-derived model, originally optimized to have a high prediction efficiency, predicts large geomagnetic events. The ratio of the number of correct to false alarm forecasts, RF, is used as an event-predictor metric. It is shown that in the electrojet regions the data-derived model that predicts the north-south component of the ground magnetic field Bx has a spatial RF profile similar to that of the prediction efficiency. Maximal values of RF=4 are found at 0300 MLT when an event is defined as an excursion in the hourly-averaged north-south component of the ground magnetic field below -400 nT. Whereas the local time profile of PE(Bx) is similar to RF(Bx), the profile of PE(|dBx/dt|) differs substantially from RF(|dBx/dt|) in the noon sector. Epoch analysis shows that the poor performance in the noon sector is a result of pre-event levels of |dBx/dt| not being clearly separated from post-event levels.