Information technology of forecasting non-stationary time-series data using singular spectrum analysis

Abstract

The information technology of forecasting non-stationary time series data, which cannot be reduced to stationary is proposed in the paper. Today, this time series class is often found in various fields, including economics, sociology, and is characterized by nonlinear trend, presence of several periodic components with variable frequency and amplitude, high noise level. Identification of non-stationary time series components is achieved using the method of singular spectrum analysis (SSA), which does not require a priori information about the time series structure. It is proposed to use several phase spaces, which can be constructed using different parameter of time window length in the SSA method, for building the models of predicting and identifying the most stable time series components. It is assumed that the time series is described by linear recurrence formulas, the coefficients of which are calculated in various phase spaces. Forecasting results are characterized by stability and efficiency as the non-stationary time series data analysis is performed in various states of the system and the most significant components are considered. The proposed information technology allows to select the amount of considered phase spaces in the forecasting model and their dimensions, as well as to make an effective short-term forecast of non-stationary time series data.