Prediction Piecewise Stationary Processes

Abstract

Consider discrete stochastic processes that contain parameters changing abruptly at random times. For forecasting processes constructed assuming a constant filter parameters. Time of the change of parameters is fixed when the substantial difference predicted and observed values of the process. Upon detection of the jump change the filter options on the initial. The problem is solved under the assumption of normal approximation of random variables and the use of linear approximations of nonlinear modeling dependences. Simulation modeling of the offered algorithms revealed their working capacity. And the more interval of constancy of parameters, the better the time determined the jump. Conversely, if you change the parameters of the proposed method becomes inoperative (or any other).