Research map / Time series / Long-term forecasting
TimeStacker: A Novel Framework with Multilevel Observation for Capturing Nonstationary Patterns in Time Series Forecasting
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 TimeStacker, a time series forecaster that stacks patches of decreasing size and aggregates them with a frequency-similarity (FreqAttention) module, reporting state-of-the-art accuracy with fewer parameters on ETT, Weather, Traffic, Electricity and Exchange.
Problem
Forecasting non-stationary time series, where a single STFT patch size cannot capture both global frequency trends and local details.
Approach
Sequentially stack patches of decreasing size across layers (large to small), with a Smooth Layer and a frequency-domain similarity attention (Hadamard-filtered Fourier features for Q/K, time-domain aggregation for V) inside each Stacker Block.
Key results
- Average MSE/MAE on ETTh1 of 0.433/0.423, best among listed baselines in Table 1 (SOFTS 0.449/0.442, iTransformer 0.454/0.447).
- Versus iTransformer, MAE reduced 7.07% on ETTm1, 5.04% on Weather, and MSE reduced 6.67% on Exchange.
- Uses 24.9 MB memory and 115 ms iteration time on ETTh1 versus RLinear's 59 ms, while achieving the lowest MSE in the efficiency comparison.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TimeStacker instead of the strongest listed baseline | ETTh1 forecasting · ETTh1 | MSE: 0.441 -> 0.433 | Helped |
| use TimeStacker instead of the strongest listed baseline | ETTh2 forecasting · ETTh2 | MAE: 0.395 -> 0.390 | Helped |
| use TimeStacker instead of the strongest listed baseline | ETTm1 forecasting · ETTm1 | MSE: 0.381 -> 0.381 | No clear effect |
| use TimeStacker instead of the strongest listed baseline | ETTm2 forecasting · ETTm2 | MAE: 0.323 -> 0.316 | Helped |
| use TimeStacker instead of the strongest listed baseline | Traffic forecasting · Traffic | MSE: 0.409 -> 0.508 | Hurt |
| use TimeStacker instead of the strongest listed baseline | Electricity forecasting · Electricity | MAE: 0.264 -> 0.275 | Hurt |
| use TimeStacker instead of the strongest listed baseline | Weather forecasting · Weather | MAE: 0.271 -> 0.264 | Helped |
| use TimeStacker instead of the strongest listed baseline | Exchange forecasting · Exchange | MSE: 0.346 -> 0.336 | Helped |
| add FreqAttention (ablation) | TimeStacker · ETTh1 (prediction length 96) | MSE: 0.388 -> 0.379 | Helped |
| add FreqAttention instead of traditional self-attention (ablation) | TimeStacker · ETTh1 (prediction length 96) | MSE: 0.386 -> 0.379 | Helped |
| add Hadamard product in FreqAttention (ablation) | TimeStacker · ETTh1 (prediction length 96) | MSE: 0.384 -> 0.379 | Helped |
| use Hadamard product instead of a linear transformation (ablation) | TimeStacker · ETTh1 (prediction length 96) | MSE: 0.381 -> 0.379 | Helped |
Surprising findings
- On Traffic, TimeStacker (MSE 0.508, MAE 0.335) is worse than SOFTS, iTransformer and others, and the authors attribute a slight decline to growing numbers of variables.
Limitations
- Performance declines as the number of variables increases; the model may hit bottlenecks on multivariate series (Traffic, Electricity).
- Efficiency is not uniformly best: slower than RLinear and uses more memory than SparseTSF/DLinear.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Traffic Electricity Exchange
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Traffic Electricity Weather Exchange
Metrics: MSE MAE
Baselines: SOFTS SparseTSF iTransformer TimeMixer SAMformer PatchTST Crossformer DLinear RLinear Pathformer
Builds on: PatchTST iTransformer RevIN DLinear
Closest papers by meaning
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.88)
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.86)
- Rethinking the Power of Timestamps for Robust Time Series Forecasting: A Global-Local Fusion Perspective (NeurIPS 2024 · similarity 0.86)
- FilterNet: Harnessing Frequency Filters for Time Series Forecasting (NeurIPS 2024 · similarity 0.86)
- Time-MMD: Multi-Domain Multimodal Dataset for Time Series Analysis (NeurIPS 2024 · similarity 0.85)
- TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data (AAAI 2025 · 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.