Research map / Time series / Long-term forecasting
Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift
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 TFPS, a time-frequency dual-domain forecaster that uses subspace-clustering pattern identification and a mixture of pattern-specific MLP experts to handle patch-level distribution shift.
Problem
Long-term time series forecasting degrades when patches within a series follow heterogeneous, drifting distributions, since single uniform models cannot capture all patterns.
Approach
A dual-domain (time and Fourier) patch encoder feeds a subspace-clustering Pattern Identifier that assigns patches to clusters, and a Mixture of Pattern Experts routes each patch to specialized MLP experts, rather than normalizing away non-stationarity.
Key results
- TFPS achieves top-1 MSE in 57 of 72 settings in the main long-term forecasting table.
- On ETTh2 and ETTm2 versus foundation models, TFPS reports MSE improvements of 10.9% and 7.3% respectively.
- Ablation on ETTh1 and ETTh2 shows the full dual-branch model with PI and MoPE gives the lowest MSE at every horizon, and removing PI (replacing with linear) generally raises MSE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TFPS instead of time-domain baselines | Time-series forecasting · Nine real-world multivariate datasets | MSE: 9.5% | Helped |
| Use TFPS instead of frequency-domain baselines | Time-series forecasting · Nine real-world multivariate datasets | MSE: 16.9% | Helped |
| Use TFPS instead of time-frequency baselines | Time-series forecasting · Nine real-world multivariate datasets | MSE: 5.2% | Helped |
| Use TFPS instead of foundation models | Time-series forecasting · ETTh2 | MSE: 10.9% | Helped |
| Use TFPS instead of foundation models | Time-series forecasting · ETTm2 | MSE: 7.3% | Helped |
| Use TFPS instead of foundation models | Time-series forecasting · Traffic | MSE: -3.6% | Hurt |
| Add the frequency branch to the time branch (ablation) | TFPS · ETTh1 H=192 | MSE: 0.459 -> 0.423 | Helped |
| Use PI instead of a linear layer in the time branch (ablation) | TFPS · ETTh1 H=192 | MSE: 0.451 -> 0.423 | Helped |
| Add PI and MoPE to the time-branch encoder (ablation) | TFPS · ETTh1 H=192 | MSE: 0.460 -> 0.423 | Helped |
| Add the time branch to the frequency branch (ablation) | TFPS · ETTh1 H=192 | MSE: 0.507 -> 0.423 | Helped |
| Use PI instead of a linear layer in the frequency branch (ablation) | TFPS · ETTh1 H=192 | MSE: 0.535 -> 0.507 | Helped |
| Add PI and MoPE to the frequency-branch encoder (ablation) | TFPS · ETTh1 H=192 | MSE: 0.583 -> 0.507 | Helped |
Surprising findings
- TFPS is slightly worse than foundation models on the Traffic dataset (e.g., -3.6% IMP), which the authors attribute to milder distribution shift.
- On Exchange-Rate with H=720, TFPS has MSE 1.011 versus about 0.84 for baselines (IMP -13.3%).
Limitations
- Assumes a fixed patch length, which the authors flag as a limitation.
- Does not handle evolving distribution shifts such as newly emerging patterns (stated as future work).
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Exchange-Rate Weather Electricity Traffic ILI
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Exchange-Rate Weather Electricity Traffic ILI
Metrics: MSE MAE IMP (relative MSE improvement)
Baselines: PatchTST DLinear TimesNet iTransformer FEDformer FITS TFDNet-IK TSLANet AutoTimes Moment Timer SIN
Builds on: PatchTST FEDformer Mixture of Experts
Closest papers by meaning
- Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.89)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.88)
- ARM: Refining Multivariate Forecasting with Adaptive Temporal-Contextual Learning (ICLR 2024 · similarity 0.88)
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.88)
- Kairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models (NeurIPS 2026 · similarity 0.88)
- Multi-Scale Finetuning for Encoder-based Time Series Foundation Models (NeurIPS 2025 · similarity 0.88)
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.