Vlasenko O. Soft Computing methods and models of risk analysis processes intellectualization

This research is devoted to the soft computing methods and models of data mining applications in the tasks of operational risk analysis. The existing methods and approaches in the tasks of forecasting, identification and analysis applied to operational risk analysis were analyzed in order to identify their weaknesses and points for improvement through the use of soft computing. A novel five-layer hybrid neuro-fuzzy model with multidimensional Gaussians in the consequent layer and a fast learning method with variants based on standard quadratic error and a specialized criterion are proposed. It is distinguished by high accuracy, processing speed and computational effectiveness. An ensemble of hybrid neuro-fuzzy models with multidimensional Gaussians in a consequent layer and a method of its synthesis are proposed, characterized by reduced learning error, improved generalization capabilities and a simplified procedure for selecting hyperparameters. The neuro-fuzzy models learning methods are improved by using empirical mode decomposition as a tool for preliminary data decomposition and noise reduction, which differs from competitors by the advanced noise reduction in the case of highly dynamic data. Were improved the method of constructing hierarchical knowledge-oriented fuzzy systems by incorporating probability estimates into the process of logical inference and using neuro-fuzzy models as components of a hierarchical structure have been further developed, which, unlike existing methods, allows to handle critical values in operational risk analysis. A number of simulation experiments based on synthetic and real data were performed, the results of which confirm the feasibility of applying the proposed techniques to solve the operational risk analysis processes intellectualization problem. The practical task of financial portfolio operative risk management is solved on the basis of the developed methods and models.