Decompose-learn-revise: A reliability-aware framework for robust time series forecasting

Univariate time series forecasting is particularly sensitive to anomalies, distribution shifts, and long-tail prediction errors. Existing methods mainly improve robustness through backbone architectures or training strategies, while ignoring which part of the intrinsic knowledge can be learned and which part to be discarded. To this end, this paper proposes a framework that integrates padding-free robust decomposition, selective learning, and retrieval-based prediction revision. The proposed decomposition mechanism performs series-level adaptive decomposition while restricting supervision to boundary-free regions without distortion caused by padding. Selective learning dynamically filters uninformative time steps, whereas the revision module compensates residual errors at inference through similarity-guided memory retrieval. Extensive experiments on fourteen datasets demonstrate consistent improvements in both accuracy and robustness across heterogeneous domains. Ablation studies further confirm the complementary contributions of each component.