Research map / Time series / Foundation models for time series
Byte Pair Encoding for Efficient Time Series Forecasting
Leon Götz, Marcel Kollovieh, Stephan Günnemann, Leo Schwinn
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace sample-wise tokens with motif-based tokenization | Chronos (T5) · 7 zero-shot datasets | MSE / speedup: 40% MSE, 2314% efficiency on average | Improved |
| add conditional decoding to motif tokenization (ablation) | motif-based tokenizer · 7 zero-shot datasets | MSE: 39.8% -> 48.0% MSE improvement; 23.14x -> 28.22x acceleration | Improved |
| add conditional decoding (ablation) | motif tokenizer, high compression, small model · ETTm1 | MSE: -44.3% MSE | Improved |
| replace fixed-length patches with motif tokenization | patch-based Chronos (len 4 and 8) · 7 zero-shot datasets | MSE: 12.7% average gain with cd | Mixed |
| use higher compression tokenizer (fewer bins, larger quantization error) (ablation) | motif tokenizer · 7 datasets x 5 model sizes | MSE: 25 of 35 settings best | Improved |
| vary minimum motif occurrence p_min (ablation) | motif tokenizer, small models · Electricity and Traffic | MSE | Mixed |
| increase tokenizer training set from 1k to 1M series (ablation) | motif tokenizer · 5 datasets avg | MSE: 0.569 -> 0.533 | Improved |
| use longer motifs for generation (ablation) | motif tokenizer medium, small model · 5 datasets | token-wise MSE correlation: rho -0.26 on ETTh1 | Improved |
| use motif tokens instead of per-sample tokens | Chronos · Traffic zero-shot | MSE: 2.717 -> 0.574 | Improved |
Closest papers by meaning
- Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization (ICML 2025 · similarity 0.89)
- Kairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models (NeurIPS 2026 · similarity 0.88)
- TimeTok: Granularity-Controllable Time-Series Generation via Hierarchical Tokenization (NeurIPS 2026 · similarity 0.87)
- Apollo-Forecast: Overcoming Aliasing and Inference Speed Challenges in Language Models for Time Series Forecasting (AAAI 2025 · similarity 0.87)
- TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.86)
- Timer-XL: Long-Context Transformers for Unified Time Series Forecasting (ICLR 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.
