A method of intelligent forecasting of random processes is described. This method is based on a more comprehensive use of statistical regularities inherent in the evolution of the process under consideration. Within the framework of the proposed approach, on the stage of training, the feature space is complemented by the parameters of mixed probability models that make it possible to reconstruct the coefficients of the stochastic differential equation describing the process under consideration. The additional statistical information imposes supplementary restrictions on the research area and thus, narrows the set of possible solutions and directs learning by preliminary rejection of improbable or highly unlikely decisions, and hence, makes the training more efficient and the forecasts more accurate.
Keywords:
time series, stochastic differential equation, mixture of probability distributions, statistical separation of mixtures, forecasting
