Research map / Time series / Long-term forecasting
Rethinking the Power of Timestamps for Robust Time Series Forecasting: A Global-Local Fusion Perspective
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 GLAFF, a model-agnostic plug-in that models timestamps with an attention mapper, robustly denormalizes them with median/IQR, and adaptively fuses the global mapping with local backbone forecasts.
Problem
Time series forecasting models underuse timestamps as global information and become unreliable when the history window contains anomalies or data drift.
Approach
Model timestamp features alone with a small Transformer encoder to produce a standard-distribution mapping, denormalize it with robust median/quantile statistics computed against history observations, and adaptively weight it against the backbone's local prediction via a softmax MLP over history residuals.
Key results
- GLAFF improves average MSE/MAE of four backbones by 12.5% across nine datasets (iTransformer improved by 5.5%, Informer by 23.8%).
- On LLM-based FPT (GPT2(3)/GPT2(6)), GLAFF gives consistent gains over the baseline, e.g., Electricity MSE 0.194 to 0.168.
- Ablation on iTransformer shows average MSE 0.2602 full vs 0.3000 without backbone and 0.2555 without quantile (Quantile variant slightly better on average than full model's disabled robust component only in average MSE comparison).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GLAFF to Informer | Informer · All nine datasets | average improvement: 23.8% | Helped |
| Add GLAFF to DLinear | DLinear · All nine datasets | average improvement: 13.1% | Helped |
| Add GLAFF to TimesNet | TimesNet · All nine datasets | average improvement: 7.5% | Helped |
| Add GLAFF to iTransformer | iTransformer · All nine datasets | average improvement: 5.5% | Helped |
| Add the backbone's local predictions alongside timestamp-only mapping (ablation) | GLAFF + iTransformer · Electricity, Traffic, Weather | average MSE: 0.3000 -> 0.2367 | Helped |
| Use attention blocks rather than equivalent-size MLP networks (ablation) | GLAFF + iTransformer · Electricity, Traffic, Weather | average MSE: 0.2646 -> 0.2367 | Helped |
| Use robust quantile denormalization rather than conventional inverse normalization (ablation) | GLAFF + iTransformer · Electricity, Traffic, Weather | average MSE: 0.2555 -> 0.2367 | Helped |
| Adaptively combine global mapping and local prediction rather than averaging them (ablation) | GLAFF + iTransformer · Electricity, Traffic, Weather | average MSE: 0.2591 -> 0.2367 | Helped |
Surprising findings
- Removing the backbone entirely (timestamps only) still gives competitive forecasts on several datasets.
- Removing timestamps from Informer, TimesNet, and iTransformer barely hurts their performance on Traffic.
Limitations
- Gains vary by dataset; common datasets such as ETTh1 and ILI show smaller improvements.
- Relies on timestamp features being available for history and future windows; holiday markers optional.
- Evaluated only on fixed history windows and benchmark datasets.
Experimental setup
Datasets: Electricity Exchange Traffic Weather ILI ETTh1 ETTh2 ETTm1 ETTm2
Benchmarks: Electricity Exchange Traffic Weather ILI ETTh1 ETTh2 ETTm1 ETTm2
Metrics: MSE MAE
Baselines: Informer DLinear TimesNet iTransformer FPT
Builds on: iTransformer Informer DLinear TimesNet FPT
Closest papers by meaning
- TimeStacker: A Novel Framework with Multilevel Observation for Capturing Nonstationary Patterns in Time Series Forecasting (ICML 2025 · similarity 0.86)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.86)
- Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.86)
- Improving Time Series Forecasting via Instance-aware Post-hoc Revision (NeurIPS 2025 · similarity 0.86)
- Breaking Silos: Adaptive Model Fusion Unlocks Better Time Series Forecasting (ICML 2025 · similarity 0.85)
- Time-MMD: Multi-Domain Multimodal Dataset for Time Series Analysis (NeurIPS 2024 · similarity 0.85)
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.