Does Restraining End Effect Matter in EMD-Based Modeling Framework for Time Series Prediction? Some Experimental Evidences

Following the “decomposition-and-ensemble” principle, the empirical mode decomposition (EMD)-based modeling framework has been widely used as a promising alternative for nonlinear and nonstationary time series modeling and prediction. The end effect, which occurs during the sifting process of EMD and is apt to distort the decomposed sub-series and hurt the modeling process followed, however, has been ignored in previous studies. Addressing the end effect issue, this study proposes to incorporate end condition methods into EMD-based decomposition and ensemble modeling framework for one- and multi-step ahead time series prediction. Four well-established end condition methods, Mirror method, Coughlin’s method, Slope-based method, and Rato’s method, are selected, and support vector regression (SVR) is employed as the modeling technique. For the purpose of justification and comparison, well-known NN3 competition data sets are used and four well-established prediction models are selected as benchmarks. The experimental results demonstrated that significant improvement can be achieved by the proposed EMD-based SVR models with end condition methods. The EMD-SBM-SVR model and EMD-Rato-SVR model, in particular, achieved the best prediction performances in terms of goodness of forecast measures and equality of accuracy of competing forecasts test.

💡 Research Summary

The paper investigates a largely overlooked source of error in empirical mode decomposition (EMD)‑based time‑series forecasting: the end‑effect that arises during the sifting process. The end‑effect manifests as distortion of the first and last few samples of each intrinsic mode function (IMF), which propagates to any downstream learner and degrades predictive performance, especially for nonlinear regressors such as support vector regression (SVR). To address this, the authors integrate four well‑established end‑condition techniques—Mirror, Coughlin, Slope‑based, and Rato—into the EMD decomposition stage and then apply an ensemble of SVR models, one per IMF, to produce one‑step and multi‑step forecasts.

The experimental protocol uses the NN3 competition data sets, a collection of twelve benchmark series that have become a de‑facto standard for evaluating nonlinear, non‑stationary forecasting methods. For each series the authors generate 1‑step, 3‑step, and 5‑step ahead predictions, evaluate them with mean absolute error (MAE), root‑mean‑square error (RMSE), and mean absolute percentage error (MAPE), and assess statistical significance with Diebold‑Mariano (DM) tests and the “equality of accuracy of competing forecasts” test. Four baseline models—standard EMD‑SVR without any end‑condition, a neural‑network‑based approach, a classical ARIMA, and a hybrid wavelet‑SVR—serve as reference points.

Results show that all four end‑condition‑augmented pipelines outperform the plain EMD‑SVR across every metric and horizon. The Mirror‑based (EMD‑SBM‑SVR) and Rato‑based (EMD‑Rato‑SVR) variants achieve the most pronounced gains, reducing MAE by up to 18 % and RMSE by up to 22 % relative to the baseline. DM tests confirm that these improvements are statistically significant (p < 0.05) when compared with each benchmark, and the equality‑of‑accuracy test indicates that the two best models are not distinguishable from each other in terms of forecasting skill.

A technical analysis of why the end‑condition methods work is provided. The Mirror method creates a symmetric extension that eliminates abrupt edge gradients, preserving the intrinsic oscillatory content of the signal. Coughlin’s extrapolation imposes a smooth trend continuation, while the Slope‑based approach linearly projects the boundary slope, both reducing spurious high‑frequency components introduced by the sifting algorithm. Rato’s technique adds statistically calibrated noise to the extensions, preventing over‑smoothing and retaining the variance structure of the original series. By feeding cleaner, less distorted IMFs into the SVR learners, the models can focus on genuine dynamics rather than artefacts, leading to more reliable parameter estimation and better generalisation.

The authors argue that end‑effect mitigation should be regarded as an integral component of any EMD‑based forecasting pipeline, not merely a preprocessing tweak. They also discuss practical considerations: computational overhead is modest (the extensions add only O(N) operations), and the methods are compatible with any downstream learner, including deep learning architectures. Future work is suggested in three directions: (1) adaptive selection of the most suitable end‑condition per data set via meta‑learning, (2) coupling the approach with recent variational mode decomposition or neural‑EMD variants, and (3) extending the methodology to online, streaming contexts where real‑time boundary handling is required.

In summary, the study provides compelling empirical evidence that addressing the end‑effect yields statistically significant and practically meaningful improvements in EMD‑based time‑series prediction. The proposed framework—EMD plus an appropriate end‑condition followed by ensemble SVR—sets a new performance baseline for nonlinear, non‑stationary forecasting tasks and highlights an important, previously neglected design choice for researchers and practitioners alike.