Research map / Time series / Long-term forecasting
Understanding Transformers for Time Series: Rank Structure, Flow-of-ranks, and Compressibility
Annan Yu, Danielle C. Maddix, Boran Han, Xiyuan Zhang, Abdul Fatir Ansari, Oleksandr Shchur, Christos Faloutsos, Andrew Gordon Wilson, Michael W. Mahoney, Bernie Wang
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 |
|---|---|---|---|
| apply truncated SVD to Chronos attention matrices down to about 23.7% of original size without finetuning (ablation) | pretrained Chronos · In-domain / zero-shot forecasting | WQL, MASE relative to original: WQL 1.053, MASE 1.005 at ratio 0.237 | No clear effect |
| compress Chronos attention below about 20% of original size (ablation) | pretrained Chronos · In-domain forecasting | WQL: 1.053 -> 1.991 -> 4.409 (ratio 0.237, 0.151, 0.073) | Hurt |
| apply same truncated SVD compression to T5 LLM instead of Chronos (ablation) | T5 · language modeling perplexity | LPPL relative to original: 1.652 at ratio 0.237; 1.028 at ratio 0.889 | Hurt |
| pretrain a compressed model with layer-dependent rank rather than compress a pretrained one | Chronos · time-series forecasting Pareto frontier | accuracy vs time | Improved |
| compress Chronos with low-rank attention and layer-dependent rank | Chronos · inference | inference time / memory: -65.4% time, -81.4% memory | Improved |
| raise the rank of the fixed input embedding from 1 to 32 (ablation) | Chronos pretraining · attention matrix numerical rank | numerical rank of W_Q | Improved |
| increase patch size k in Chronos-Bolt (ablation) | Chronos-Bolt embedding · embedding spectrum | numerical rank of embedding | Improved |
Closest papers by meaning
- Utilizing Image Transforms and Diffusion Models for Generative Modeling of Short and Long Time Series (NeurIPS 2024 · similarity 0.89)
- Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.88)
- Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers (ICML 2025 · similarity 0.87)
- Rough Transformers: Lightweight and Continuous Time Series Modelling through Signature Patching (NeurIPS 2024 · similarity 0.87)
- On the Role of Attention Masks and LayerNorm in Transformers (NeurIPS 2024 · similarity 0.87)
- Time Series Representations with Hard-Coded Invariances (ICML 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.
