Research map / Time series / Long-term forecasting
Tackling Time-Series Forecasting Generalization via Mitigating Concept Drift
Zhiyuan Zhao, Haoxin Liu, B. Aditya Prakash
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ShifTS (SAM plus RevIN) to forecasting backbone | Crossformer · ILI (avg over horizons) | MSE improvement: 81.9% MSE improvement | Improved |
| add ShifTS to forecasting backbone | iTransformer · ILI (avg over horizons) | MSE improvement: 13.8% MSE improvement | Improved |
| add ShifTS to forecasting backbone | iTransformer · ETTm1 (avg over horizons) | MSE improvement: 5.1% MSE improvement | Improved |
| add ShifTS to older backbones | Informer · ETTh1 (avg over horizons) | MSE improvement: 90.7% MSE improvement | Improved |
| add ShifTS to TimeMixer | TimeMixer · ETTm1 (avg over horizons) | MSE improvement: 2.3% MSE improvement | Improved |
| use ShifTS instead of concept drift baselines (GroupDRO, IRM, VREx, EIIL) | Crossformer · ILI | MSE: 2.036 (best concept drift baseline, EIIL) -> 0.668 | Improved |
| use invariant-learning concept drift methods alone without temporal shift handling | Crossformer · Exchange | MSE: 0.819 -> 0.821-0.846 | No clear effect |
| replace RevIN with SAN inside ShifTS (ablation) | ShifTS on Crossformer · Exchange (avg) | MSE: 0.470 -> 0.407 | Improved |
| mitigate both temporal shift and concept drift versus only one (ablation) | ShifTS · Exchange horizon 96 | MSE | Improved |
| add concept drift mitigation (SAM) to models with built-in norm/denorm (ablation) | PatchTST, iTransformer · Exchange horizon 96 | MSE | Improved |
| add temporal shift mitigation (RevIN) to a model without normalization (ablation) | Crossformer · Exchange horizon 96 | MSE | Improved |
| higher mutual information between horizon exogenous features and target | ShifTS · six datasets | performance gain: p=0.012 | Improved |
Closest papers by meaning
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.87)
- Improving Time Series Forecasting via Instance-aware Post-hoc Revision (NeurIPS 2025 · similarity 0.87)
- Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift (NeurIPS 2025 · similarity 0.87)
- Time-Series Forecasting for Out-of-Distribution Generalization Using Invariant Learning (ICML 2024 · similarity 0.87)
- Introducing Spectral Attention for Long-Range Dependency in Time Series Forecasting (NeurIPS 2024 · similarity 0.86)
- Sequence Complementor: Complementing Transformers for Time Series Forecasting with Learnable Sequences (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
