A Method for Predicting Rare Events by Multidimensional Time Series with the Use of Collective Methods

Abstract

A method for predicting rare events by the preceding dynamics of features is considered. The method is analyzed on the example of the problem of predicting revocation of the license of a credit institution on the basis of the reporting indicators published at least six months before the regulator made the appropriate decision. The technology developed is based on the calculation of collective solutions by sets of recognition algorithms. Investigations have shown that the most effective prediction is obtained with the use of collective algorithms involving various types of decision forests and combinatorial and logical methods. The method developed also involves the procedure of ranking the indicators according to their information value, in which the collective ranking is calculated on the basis of information estimates obtained with the use of built-in procedures within individual recognition methods.