Research map / Time series / Long-term forecasting
TimeBase: The Power of Minimalism in Efficient Long-term 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 TimeBase, an ultra-lightweight (0.39K parameters) long-term forecaster that extracts low-rank basis components from segmented series and forecasts at segment level, also usable as a complexity reducer for patch-based models.
Problem
Long-term time series forecasting typically relies on models with thousands to millions of parameters, incurring high computational and memory cost despite time series often having temporal pattern similarity and low-rank structure.
Approach
Segment the series into patches of basis length P, learn a small set of R basis components via a linear layer with an orthogonality regularizer, and forecast future segments by linearly combining these basis components, so parameters scale with T/P and L/P rather than TL.
Key results
- TimeBase uses 0.39K parameters and 2.77M MACs on Electricity (720 horizon), a 2.56x parameter and 4.58x MAC reduction versus SparseTSF's 1.00K parameters.
- TimeBase ranks Top-2 on 16 of 17 normal-scale datasets against ten baselines, with the exception being ETTh2.
- As a plug-and-play component, TimeBase cuts PatchTST MACs by 77.74% to 93.03% while reducing MSE by 1.27% on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TimeBase instead of DLinear | LTSF forecasting across 17 normal-scale datasets · 17 normal-scale datasets | MSE: 8.82% | Helped |
| Use TimeBase instead of SparseTSF or FITS | LTSF forecasting across 17 normal-scale datasets · 17 normal-scale datasets | MSE: 6.15∼6.34% | Helped |
| Add TimeBase basis extraction before patching | PatchTST · 7 datasets, 56 forecasting metrics | MSE: 1.27% | Mixed |
| Add TimeBase basis extraction before patching | PatchTST · 7 datasets | MACs: 77.74% to 93.03% | Helped |
| Add TimeBase basis extraction before patching | PatchTST · 7 datasets | parameters: 56.95% to 91.43% | Helped |
| Vary orthogonal loss weight λorth (ablation) | TimeBase · ETTh1 and ETTh2 | MSE | Mixed |
| Vary orthogonal loss weight λorth (ablation) | TimeBase · Traffic and Electricity | MSE | No clear effect |
| Increase basis count from too few components (e.g., R = 2) (ablation) | TimeBase · Traffic and Electricity | MSE | Helped |
Surprising findings
- A model with only 0.39K parameters matches or beats much larger Transformer and linear baselines on most normal-scale benchmarks.
Limitations
- Performance on ETTh2 is not top-2, and results are only described as competitive relative to larger models.
- Basis length P must be chosen from prior periodicity or FFT analysis, and the orthogonal weight and R need tuning per dataset.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic Solar Wind METR-LA Exchange ZafNoo
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic Solar Wind METR-LA Exchange ZafNoo
Metrics: MSE MAE parameters MACs max memory epoch time CPU inference time
Baselines: TimeMixer iTransformer PatchTST DLinear TimesNet FEDformer Autoformer Informer FITS SparseTSF
Builds on: DLinear SparseTSF PatchTST Channel Independence
Closest papers by meaning
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.90)
- Parsimony or Capability? Decomposition Delivers Both in Long-term Time Series Forecasting (NeurIPS 2024 · similarity 0.89)
- STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · similarity 0.89)
- Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.88)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.88)
- TStruct: Learning Shared Temporal Structures for Long-Term Time-Series Forecasting (NeurIPS 2026 · 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.