Revisiting the Efficacy of Signal Decomposition in AI-based Time Series Prediction
This work exposes a widespread error in time series modeling that could invalidate progress in relevant areas, requiring recalibration to avoid scientific detours and practical losses.
The paper challenges the effectiveness of signal decomposition in AI-based time series prediction by identifying that improper dataset processing with future label leakage leads to misleadingly superior results, and shows that when data is processed causally, the benefits of decomposed signals diminish.
Time series prediction is a fundamental problem in scientific exploration and artificial intelligence (AI) technologies have substantially bolstered its efficiency and accuracy. A well-established paradigm in AI-driven time series prediction is injecting physical knowledge into neural networks through signal decomposition methods, and sustaining progress in numerous scenarios has been reported. However, we uncover non-negligible evidence that challenges the effectiveness of signal decomposition in AI-based time series prediction. We confirm that improper dataset processing with subtle future label leakage is unfortunately widely adopted, possibly yielding abnormally superior but misleading results. By processing data in a strictly causal way without any future information, the effectiveness of additional decomposed signals diminishes. Our work probably identifies an ingrained and universal error in time series modeling, and the de facto progress in relevant areas is expected to be revisited and calibrated to prevent future scientific detours and minimize practical losses.