Research map / Time series / Foundation models for time series
Revealing Scaling Paradox in Large-scale Time Series Models: Implications for More Efficient and Accurate Forecasting
Xin Qiu, Junlong Tong, Yirong Sun, Yunpu Ma, Xiaoyu Shen
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 |
|---|---|---|---|
| scale model from Small to Large | LLM4TS (single-dataset) · avg in-domain | MSE: 0.2226 -> 0.2846 | Hurt |
| scale model from Small to Large | TSFMs (single-dataset) · avg in-domain | MSE: 0.2286 -> 0.2540 | Hurt |
| add cross-dataset training with scale | LLM4TS (Small-C to Large-C) · avg in-domain | MSE: 0.2367 -> 0.2233 | Improved |
| scale TSFMs under cross-dataset learning from Small-C to Large-C | TSFMs · avg in-domain | MSE: 0.2257 -> 0.2543 | Hurt |
| increase training data ratio from 20% to 100% | LLM4TS and TSFMs · single-dataset MSE | MSE: MSE decreases but gap not narrowed | No clear effect |
| add critical-layer pruning and fine-tuning | LLM4TS-Large · single-dataset in-domain | MSE: 0.285 -> 0.219 | Improved |
| add critical-layer pruning and fine-tuning | LLM4TS-Base · single-dataset in-domain | MSE: 0.278 -> 0.226 | Improved |
| add critical-layer pruning and fine-tuning | LLM4TS-Large-C · cross-dataset in-domain | MSE: 0.223 -> 0.226 | Hurt |
| add critical-layer pruning and fine-tuning | TSFMs-Large-C · cross-dataset out-of-domain | MSE: 0.872 -> 0.821 | Improved |
| add critical-layer pruning and fine-tuning | All six single-dataset models · single-dataset | inference efficiency: 2.7 | Improved |
| use critical layers instead of full backbone (ablation) | FSCA, CALF, OFA · 5 datasets avg | MSE: 0.348 -> 0.329 | Improved |
| use critical layers instead of random layers (ablation) | FSCA, CALF, OFA · 5 datasets avg | MSE: 0.343 -> 0.329 | Improved |
Closest papers by meaning
- Less is More: Unlocking Specialization of Time Series Foundation Models via Structured Pruning (NeurIPS 2025 · similarity 0.91)
- Towards Neural Scaling Laws for Time Series Foundation Models (ICLR 2025 · similarity 0.90)
- Scaling Law for Time Series Forecasting (NeurIPS 2024 · similarity 0.90)
- Multi-Scale Finetuning for Encoder-based Time Series Foundation Models (NeurIPS 2025 · similarity 0.90)
- Exploring Representations and Interventions in Time Series Foundation Models (ICML 2025 · similarity 0.88)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · 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.
