A consistent framework for decomposition-based time series forecasting

doi:10.1016/j.compeleceng.2025.110773

A consistent framework for decomposition-based time series forecasting

2025-10-28

0

PRE

AI

DOI：10.1016/j.compeleceng.2025.110773

Original

Original request for help

Save

Abstract

En 中文

Time-series forecasting plays a pivotal role in a wide range of real-world applications, particularly in domains such as energy management and transportation systems, where accurate predictions are essential for planning, safety, and efficiency. Although decomposition-based methods have been extensively explored to capture temporal patterns, such as trends and seasonality, many of these approaches suffer from false success due to data leakage and boundary effects. Such issues lead to inconsistencies between the training and testing phases, ultimately undermining the reliability of the forecasting models. To overcome these limitations, we propose a novel decomposition-agnostic framework designed to guarantee consistency while enhancing forecasting accuracy. The core of our approach is a recursive decomposition mechanism that effectively mitigates boundary effects and resolves task mismatch between training and inference. Unlike conventional decomposition-based techniques that either overlook boundary issues or incorrectly apply decomposition to test data, our framework eliminates these pitfalls, providing a more principled and robust solution. We validate the effectiveness of the proposed framework through extensive experiments on multiple real-world benchmark datasets. The empirical results consistently demonstrate that our method achieves substantial improvements over state-of-the-art baselines, not only in terms of forecasting accuracy but also in stability across different data conditions.

Keywords:

Time series

Decomposition

Machine learning

Deep learning

A consistent framework for decomposition-based time series forecasting

Abstract

Journal

Organization