Research map / Time series / Foundation models for time series
LightGTS: A Lightweight General Time Series Forecasting Model
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 LightGTS, a ~4M-parameter general time series forecasting foundation model using period-aligned adaptive patching with a flex projection layer and non-autoregressive periodical parallel decoding, matching or beating much larger TSFMs in zero-shot forecasting.
Problem
Existing time series foundation models rely on huge parameter counts and multi-source pretraining, which is computationally expensive; fixed tokenization also fails to handle varying scales (sampling rates) and intrinsic periods across pretraining datasets.
Approach
Segment series into patches whose length equals the intrinsic cycle length (periodical patching) so each token covers one full period; a flex projection layer resizes patch embedding weights via a pseudoinverse-based resize to handle varying patch sizes; and periodical parallel decoding feeds the replicated last encoder token (with exponential decay weights) into a non-autoregressive decoder.
Key results
- LightGTS-mini (4M params) zero-shot achieves best average MSE on most datasets (e.g., ETTm1 0.327 vs Time-MoE 0.376, Chronos 0.435), with 1st place in 31 of 36 metric entries in the zero-shot table.
- In zero-shot, LightGTS-tiny (1.3M) reports ~27% average MSE reduction over the best baselines; LightGTS-mini reports 27% and 28% MSE reductions vs Chronos and MOIRAI.
- In full-shot setting, LightGTS-mini reports about 7% average MSE reduction over six small SOTA baselines; Table of efficiency shows 4M params and 213M MACs vs 453M/300M/700M for Time-MoE/MOIRAI/Chronos.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use periodical rather than fixed patching with PPD (ablation) | LightGTS · ETT-avg zero-shot | MSE: 0.436 -> 0.328 | Helped |
| use periodical rather than fixed patching with AR decoding (ablation) | LightGTS · ETT-avg zero-shot | MSE: 0.442 -> 0.341 | Helped |
| use periodical rather than fixed patching with MAE decoding (ablation) | LightGTS · ETT-avg zero-shot | MSE: 0.537 -> 0.388 | Helped |
| use PPD instead of AR decoding with periodical patching (ablation) | LightGTS · ETT-avg zero-shot | MSE: 0.341 -> 0.328 | Helped |
| use PPD instead of MAE decoding with periodical patching (ablation) | LightGTS · ETT-avg zero-shot | MSE: 0.388 -> 0.328 | Helped |
| use the last encoder token rather than a learnable embedding (ablation) | LightGTS · Weather zero-shot | MSE: 0.278 -> 0.208 | Helped |
| use the last encoder token rather than the CLS token (ablation) | LightGTS · Weather zero-shot | MSE: 0.341 -> 0.208 | Helped |
| use the last encoder token rather than the mean encoder token (ablation) | LightGTS · Weather zero-shot | MSE: 0.328 -> 0.208 | Helped |
| set reference patch size to 24 rather than 48 (ablation) | LightGTS · Weather zero-shot | MSE: 0.208 -> 0.213 | No clear effect |
| set reference patch size to 96 rather than 48 (ablation) | LightGTS · Weather zero-shot | MSE: 0.208 -> 0.210 | No clear effect |
| set reference patch size to 192 rather than 48 (ablation) | LightGTS · Weather zero-shot | MSE: 0.208 -> 0.212 | No clear effect |
| apply periodical tokenization to Timer (ablation) | Timer · Weather and Solar zero-shot | MSE: Timer: 0.190, 0.261, 0.332, 0.385; + Flex-resize: 0.178, 0.238, 0.297, 0.357 | Helped |
Surprising findings
- Ablation shows periodical patching helps even Timer as a plug-in without retraining, with flex-resize giving up to ~28-33% gains on Solar.
- Replacing the decoder input with the last encoder token beats learnable, CLS, and mean tokens, and MAE decoding performs worst.
Limitations
- FFT-based period finding can misalign with true intrinsic period on weakly periodic data (ETT, Weather), hurting performance.
- Relies on known or estimated cycle length; the method's gains depend on periodicity of the data (little gain for Exchange).
Experimental setup
Datasets: Aus. Electricity Demand Wind Wind Farms Solar Power London Smart Meters Phoneme EigenWorms PRSA Temperature Rain StarLightCurves Worms Saugeen River Flow
Benchmarks: ETTm1 ETTm2 ETTh1 ETTh2 Traffic Weather Exchange Solar Electricity
Metrics: MSE MAE MACs parameters max memory inference time
Baselines: Timer MOIRAI Chronos TimesFM Time-MoE PDF iTransformer Pathformer FITS TimeMixer PatchTST
Builds on: Transformer RoPE RevIN PatchTST
Closest papers by meaning
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.87)
- SEMPO: Lightweight Foundation Models for Time Series Forecasting (NeurIPS 2025 · similarity 0.87)
- Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization (ICML 2025 · similarity 0.86)
- Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.86)
- TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.86)
- Apollo-Forecast: Overcoming Aliasing and Inference Speed Challenges in Language Models for Time Series Forecasting (AAAI 2025 · 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.