Research map / Time series / Long-term forecasting
Efficiently Enhancing Long-term Series Forecasting via Ultra-long Lookback Windows
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes IRPA, a lightweight four-linear-layer plug-in framework that decomposes, patches, and refines ultra-long lookback windows (480-3600) into a short effective input plus similarity-based auxiliary predictions for long-term time series forecasting.
Problem
Long-term time series forecasting struggles to exploit long lookback windows due to overfitting, computational/memory costs, and difficulty extracting relevant information.
Approach
An Input Refinement Module applies seasonal-trend decomposition and patching, then uses Hellinger-distance similarity to select historical patches to refine a short window; a Prediction Auxiliary Module adds similar-history and seasonal-pattern predictions to the output, so models can use ultra-long history with few parameters.
Key results
- IRPA achieves best results in 54 of 64 long-term forecasting settings and reduces average MSE by 9.4% versus iTransformer.
- IRPA reduces MSE by 9.7% on iTransformer, 15.0% on PatchTST, 14.4% on TimesNet, and 25.2% on DLinear when used as a plug-in.
- On Weather (k=720), IRPA with 0.94M parameters achieves MSE 0.289-0.292 range comparisons; PatchTST+lookback goes OOM while PatchTST+IRPA runs at 3723MB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add IRPA | iTransformer · Weather | MSE: 12.4% | Helped |
| Add IRPA | PatchTST · Solar | MSE: 22.6% | Helped |
| Add IRPA | TimesNet · Traffic | MSE: 26.0% | Helped |
| Add IRPA | DLinear · Solar | MSE: 33.0% | Helped |
| Add Seasonal Refinement Module (SRM) (ablation) | IRPA · ETTm2, Weather, ECL | MSE: 0.242 -> 0.239; 0.221 -> 0.219; 0.169 -> 0.169 | Helped |
| Add Trend Refinement Module (TRM) (ablation) | IRPA · ETTm2, Weather, ECL | MSE: 0.243 -> 0.239; 0.229 -> 0.219; 0.176 -> 0.169 | Helped |
| Add Historical Similarity Auxiliary Module (HSAM) (ablation) | IRPA · ETTm2, Weather, ECL | MSE: 0.242 -> 0.239; 0.220 -> 0.219; 0.170 -> 0.169 | Helped |
| Add Historical Seasonal Pattern Prediction Auxiliary Module (HSPPAM) (ablation) | IRPA · ETTm2, Weather, ECL | MSE: 0.240 -> 0.239; 0.224 -> 0.219; 0.176 -> 0.169 | Helped |
Surprising findings
- Simply extending the lookback window of baselines to 3600 often hurts accuracy (e.g., TimesNet MSE rises from 0.365 to 0.538 on Weather), indicating overfitting.
Limitations
- Gains are reported on standard benchmarks with fixed refined length r=96 and patch length 96; sensitivity to these choices is only partly explored.
- Optimal lookback depends on sampling frequency, requiring per-dataset selection of window length.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Solar ECL Traffic
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Solar ECL Traffic
Metrics: MSE MAE
Baselines: iTransformer PatchTST Autoformer TimesNet SCINet DLinear
Builds on: DLinear PatchTST RevIN
Closest papers by meaning
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.85)
- Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.85)
- ARM: Refining Multivariate Forecasting with Adaptive Temporal-Contextual Learning (ICLR 2024 · similarity 0.84)
- Introducing Spectral Attention for Long-Range Dependency in Time Series Forecasting (NeurIPS 2024 · similarity 0.84)
- Adaptive Meta-Learning Probabilistic Inference Framework for Long Sequence Prediction (AAAI 2024 · similarity 0.84)
- Affirm: Interactive Mamba with Adaptive Fourier Filters for Long-term Time Series Forecasting (AAAI 2025 · similarity 0.84)
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.