M-EDEM: A MNN-based Empirical Decomposition Ensemble Method for improved time series forecasting

Various fields of temporal prediction have benefited from decomposition ensemble methods (DEMs) that incorporate decomposition techniques. Historically, researchers decompose training and test data together in the pre-processing stage, which results in information leakage and hindcasting experiments. RDEMs, or rolling decomposition ensemble methods, decompose training and test sets separately to prevent information leakage during continuous forward prediction. However, boundary effects that exist in most decomposition methods still adversely affect RDEM prediction performance. Furthermore, RDEMs also suffer from heavy computation overhead during rolling decomposition, further limiting their use. A new decomposition mechanism based on learned decomposition mapping is presented in this paper. The relationship between the original time series and the decomposed results is learned by a simple neural network using a supervised approach. The neural network is then used during the rolling decomposition process to relieve the repetitive computation overhead. Furthermore, extended mapping and partial decomposition are proposed to largely alleviate the boundary effects on prediction performance. Comparative studies on public datasets demonstrate that our proposed method outperforms existing RDEMs in terms of operation speed and prediction accuracy.

Automated summary

This paper presents a new decomposition mechanism based on learned decomposition mapping. By using a neural network to learn the relationship between original time series and decomposed results, the repetitive computation overhead during rolling decomposition is relieved. Additionally, extended mapping and partial decomposition methods are proposed to alleviate boundary effects on prediction performance. Comparative studies demonstrate that the proposed method outperforms existing RDEMs in terms of operation speed and prediction accuracy.