Statistical estimation of efficiency of earthquake prediction under uncertain identification of target data
A problem of estimation of the prediction efficiency parameters, namely: the success rate and the alarm rate, under uncertain identification of prediction targets in a test data set is considered. Due to the exponential distribution of earthquake magnitude, the problem is particularly significant in magnitude threshold based earthquake prediction. It has been shown that a proper approach to the problem must take into account the probability of incidental correlation of issued predictions with the "other" events that are not targets of prediction but that are included in the test data. A complete solution for the two-class and multiclass cases is given. Since in the multiclass case the final system of equations is ill-conditioned, an approximate way of estimation of the prediction rates is proposed. Numerical test have proved correctness of the solution and reasonability of the approximation. The same tests show that formerly used estimators can deliver unacceptable results.
Bibliogr. 3 poz.