Research map / Time series / Foundation models for time series
TimeMRA: LLM-Empowered Time Series Forecasting via Multi-Scale Retrieval-Augmented Representations
Zongjiang Shang, Chengxi Jin, Binqing Wu, Dongliang Cui, Yue Yu, Haobang Sun, Chuanlin Xu, Ling Chen
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 |
|---|---|---|---|
| add cross-scale disentanglement constraint (CSDC) (ablation) | TimeMRA · ETTh1 (horizon 96) | MSE: 0.422 -> 0.369 | Improved |
| add temporal-to-semantic constraint (ablation) | TimeMRA · ETTh1 (horizon 96) | MSE: 0.417 -> 0.369 | Improved |
| add semantic-to-temporal constraint (ablation) | TimeMRA · ETTh1 (horizon 96) | MSE: 0.407 -> 0.369 | Improved |
| add scale-aware prompt generation (SAPG) (ablation) | TimeMRA · ETTh1 (horizon 96) | MSE: 0.372 -> 0.369 | Improved |
| use frequency-based multi-scale decomposition instead of predefined windows (ablation) | TimeMRA · ETTh1 (horizon 96) | MSE: 0.373 -> 0.369 | Improved |
| add LLM-augmented encoder (vs linear layers) (ablation) | TimeMRA · ETTh1 (horizon 96) | MSE: 0.400 -> 0.369 | Improved |
| add cross-modality retrieval module (ablation) | TimeMRA · ETTh1 (horizon 96) | MSE: 0.399 -> 0.369 | Improved |
| add time series encoder pathway (vs linear layers) (ablation) | TimeMRA · ETTh1 (horizon 96) | MSE: 0.377 -> 0.369 | Improved |
| use more advanced LLM backbone than GPT2-small (ablation) | TimeMRA · ETTh1 (horizon 96) | MSE: 0.369 -> 0.353 | Improved |
| use multi-scale retrieval-augmented representations vs TimeCMA | TimeCMA · ETTh1 few-shot (5% training data) | MSE: 0.679 -> 0.629 | Improved |
| vary number of dominant periods S away from 6 (ablation) | TimeMRA · ETTh1 | MSE | Mixed |
| vary router TopM away from 3 (ablation) | TimeMRA · ETTh1 | MSE | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TimeCMA — TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment (AAAI 2025) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment (AAAI 2025 · similarity 0.92)
- TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster (NeurIPS 2025 · similarity 0.90)
- $S^2$IP-LLM: Semantic Space Informed Prompt Learning with LLM for Time Series Forecasting (ICML 2024 · similarity 0.90)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.89)
- TimeXL: Explainable Multi-modal Time Series Prediction with LLM-in-the-Loop (NeurIPS 2025 · similarity 0.88)
- STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · 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.
