Forecasting Multivariate Time Series with a Dynamic-System-Based Hybrid Model

This paper proposes a time series forecasting approach that integrates nonlinear dynamic systems reconstruction with statistical forecasting methods. The suggested technique is as follows: the time series is decomposed into dynamic and stochastic components, each of which is used to build independent predictive models. In the end, the forecasting results are combined together. The research compares the prediction results and errors values for with the suggested technique, ARIMA, SVR, and a sparse identification method SINDy on synthetic and real-world data sets. Forecasting accuracy utilizing the suggested technique outperforms that of competing methods. Thus, the proposed approach may be utilized to solve practical issues involving the prediction of the behavior of diverse processes for which the mathematical model is unknown but data are accessible at discrete periods.