Forecasting Non-Stationary Time Series Using Kernel Regression for Control Problems

Abstract

A combined algorithm for a time series analysis is considered based on two basic methods: the empirical mode decomposition and kernel regression. The essence of the presented algorithm is the sequential calculation of nuclear regressions and residues, which results in the decomposition of the original series into an additive mixture of the number of regressions and residual series. The illustrative examples for the application of the proposed algorithm (immunology, economics, and other fields of studies) are provided along with their statistical results of numerical simulation. The results obtained would be useful for a smart control system design and real-time decision making support as it concerns the problems of stochastic control over a wide range of poorly formalized objects from various applied areas.