Research map / Time series / Long-term forecasting
Rethinking Time Series Tokenization from a Frequency Perspective
Zelong Tian, Ming Jin, Bo Du, Shirui Pan
Read the paperCheck your idea against 137,217 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.
Argues that common subsequence-based tokenization of time series introduces spectral distortions that push learned representations away from true trend, periodicity, and seasonality, and studies this from a frequency perspective.
Problem
Tokenizing time series by partitioning them into subsequences (patches) is standard in time series modeling, but the authors claim current tokenization strategies inherently distort the spectrum, hurting representations of trend, periodicity, and seasonality.
Approach
Analyzes subsequence tokenization theoretically through a frequency lens, showing it induces spectral distortion; the abstract does not specify the proposed fix.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace patch tokenizer with TimeFT in PatchTST | PatchTST · ETTm1 (H=96) | MSE: 0.321 -> 0.290 | Improved |
| replace phase tokenizer with TimeFT in PhaseFormer | PhaseFormer · ETTm1 (H=96) | MSE: 0.312 -> 0.293 | Improved |
| replace patch tokenizer with TimeFT in PatchTST | PatchTST · Traffic (H=96) | MSE: 0.403 -> 0.395 | Improved |
| add Stage 1 frequency-domain decomposition (ablation) | TimeFT · ETTm1 (H=96) | MSE: 0.328 -> 0.290 | Improved |
| add Stage 2 spectral shifting (ablation) | TimeFT · ETTm2 (H=96) | MSE: 0.194 -> 0.162 | Improved |
| add Stage 3 information-preserving token extraction (ablation) | TimeFT · ETTm1 (H=96) | MSE: 0.375 -> 0.290 | Improved |
| add Stage 3 token extraction (ablation) | TimeFT · Traffic (H=96) | MSE: 0.475 -> 0.395 | Improved |
| use TimeFT tokens instead of patch tokens in PatchTST backbone | PatchTST · ETTh1 (avg over H) | MSE: 0.417 vs TimeMosaic 0.412 | Mixed |
| use TimeFT instead of baselines on Weather | PatchTST · Weather (avg over H) | MSE: 0.232 vs xPatch 0.224 | No clear effect |
| increase tokenization stride for longer inputs (ablation) | TimeFT · ETTm1 (N=96 vs N=720) | MSE/MAE | Mixed |
| use TimeFT tokens for classification | PatchTST · UEA (10 datasets) | accuracy (%): 69.13 -> 71.31 | Improved |
| use TimeFT tokens for anomaly detection | PatchTST · 5 anomaly datasets | F1 (%): 82.79 -> 86.12 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 4 on most shared results.
- TimeMosaic — TimeMosaic: Temporal Heterogeneity Guided Time Series Forecasting via Adaptive Granularity Patch and Segment-wise Decoding (AAAI 2026) · better on 2 of 3 shared results
- xPatch — xPatch: Dual-Stream Time Series Forecasting with Exponential Seasonal-Trend Decomposition (AAAI 2025) · better on 0 of 2 shared results
- PhaseFormer — PhaseFormer: From Patches to Phases for Efficient and Effective Time Series Forecasting (ICLR 2026) · better on 1 of 1 shared result
- TimeFilter — TimeFilter: Patch-Specific Spatial-Temporal Graph Filtration for Time Series Forecasting (ICML 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization (ICML 2025 · similarity 0.89)
- VarDrop: Enhancing Training Efficiency by Reducing Variate Redundancy in Periodic Time Series Forecasting (AAAI 2025 · similarity 0.88)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.87)
- FilterNet: Harnessing Frequency Filters for Time Series Forecasting (NeurIPS 2024 · similarity 0.87)
- Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.86)
- TimeTok: Granularity-Controllable Time-Series Generation via Hierarchical Tokenization (NeurIPS 2026 · 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.
