Decomposition-based Ensemble Model for Non-stationary Time Series Forecast

Time series forecasting is a traditional but still an essential topic nowadays in various application fields of finance, healthcare, and manufacturing. One of the major issues falls in the limited prediction accuracy resulting from the nonstationary and nonlinear behavior of time series. Previous studies have discussed the possibility of combining machine learning techniques and decomposition methods to model the nonstationary and nonlinear patterns. However, those hybrid models did not consider the end effect problem introduced by the decomposition methods, which could introduce significant forecasting errors in practice due to the lack of information beyond the time series boundary. Therefore, a novel decomposition-guided time series forecasting framework is proposed in this work to learn the complex temporal pattern while mitigating or even eliminating the end effect in the forecasting. To better demonstrate the key idea of this unique integration, the intrinsic time-scale decomposition (ITD) and Gaussian Process (GP) are considered as examples to show how the machine learning models can "learn" from the decomposition state space with similar recurrent patterns to minimize the temporal variation. A case study based on the time series data collected from an automotive assembly line is conducted to show the effectiveness of the proposed framework.