Research map / Time series / Foundation models for time series
TRACE: Grounding Time Series in Context for Multimodal Embedding and Retrieval
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 TRACE, a multimodal time-series retriever that aligns channel- and sample-level time-series embeddings with text via dual-level hard-negative contrastive learning, used for retrieval, RAG-augmented forecasting, and as a standalone encoder.
Problem
Cross-modal retrieval between multivariate time series and text is underexplored; existing time-series retrievers lack semantic grounding, align poorly across modalities, and handle multi-channel signals weakly, while RAG for time-series foundation models is rarely studied.
Approach
A masked-reconstruction time-series encoder with channel identity tokens and channel-biased attention (RoPE on true time offsets), followed by contrastive alignment to frozen Sentence-Transformer text embeddings at both sample and channel levels with hierarchical hard-negative mining; the resulting embedding serves as a retriever feeding soft prompts into frozen forecasters.
Key results
- Text-to-TS and TS-to-Text retrieval on 2,000 query pairs: TRACE reaches about 90% top-1 label matching and 44% top-1 modality matching, far above Moment and TS2Vec.
- Weather event classification: TRACE with RAG reaches 89.76% accuracy and 72.36 F1, versus 86.09% for the best train-from-scratch baseline (TimesNet).
- RAG forecasting on Weather (H=24) reduces TRACE MSE from 0.718 without RAG to 0.696 with TS+Text retrieval; TS-to-TS retrieval P@1 is 0.900 versus 0.682 for CTSR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TRACE instead of Moment as the retrieval encoder | Text–Timeseries retrieval, TS-to-Text · TS-to-Text retrieval | Label matching P@1: 55.73 -> 90.08 | Helped |
| Use TRACE instead of Moment as the retrieval encoder | Text–Timeseries retrieval, Text-to-TS · Text-to-TS retrieval | Label matching P@1: 64.67 -> 89.63 | Helped |
| Use TRACE instead of CTSR | Weather TS-to-TS retrieval · Weather TS-to-TS retrieval | P@1: 0.682 -> 0.900 | Helped |
| Add retrieved context to forecasting | Timer-XL · Weather forecasting, H=24 | MAE: 0.729 -> 0.712 | Helped |
| Add retrieved context to forecasting | Time-MoE · Weather forecasting, H=24 | MAE: 0.635 -> 0.611 | Helped |
| Add retrieved context to forecasting | Moment · Weather forecasting, H=24 | MAE: 0.645 -> 0.631 | Helped |
| Add retrieved TS+Text context | TRACE forecasting · Weather forecasting, H=24 | MAE: 0.576 -> 0.555 | Helped |
| Add retrieved text to the TS-only prompt (ablation) | TRACE forecasting · Weather forecasting, H=24 | MAE: 0.556 -> 0.555 | No clear effect |
| Add retrieval augmentation (ablation) | TRACE · Weather event classification | Accuracy: 85.20 -> 89.76 | Helped |
| Remove channel identity tokens (CIT) (ablation) | TRACE · Forecasting and reconstruction | Average MSE | Hurt |
| Replace channel-biased attention with full attention (ablation) | TRACE · Forecasting and reconstruction | Average MSE | Hurt |
| Replace channel-biased attention with causal attention (ablation) | TRACE · Forecasting and reconstruction | Average MSE | Hurt |
Surprising findings
- Fine-tuned time-series foundation models (Time-MoE, Moment, Timer-XL, Chronos) underperform train-from-scratch models on weather event classification.
- Retrieval precision of TRACE exceeds the classification accuracy of the train-from-scratch models.
Limitations
- Relies on supervised textual alignment, which may not be available in all domains.
- RAG uses simple embedding concatenation and the augmentation architecture is not optimized.
- Evaluation of RAG limited to three open-source foundation models (Time-MoE, Timer-XL, Moment).
Experimental setup
Datasets: NOAA Storm Events Database NOAA GHCN-Hourly TimeMMD MTBench
Benchmarks: curated weather dataset TimeMMD Health TimeMMD Energy TimeMMD Environment
Metrics: P@1 P@5 MRR ROUGE MAE MSE accuracy F1 query time
Baselines: DLinear iTransformer PatchTST TimesNet TimeMixer FSCA Chronos TimesFM Timer-XL Time-MoE Moirai Moment
Builds on: TimeMMD MTBench RoPE reversible instance normalization Sentence-Transformer
Closest papers by meaning
- Empowering Time Series Analysis with Large-Scale Multimodal Pretraining (NeurIPS 2026 · similarity 0.88)
- TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster (NeurIPS 2025 · similarity 0.87)
- TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment (AAAI 2025 · similarity 0.87)
- Align-RAG: Alignment Is All You Need for TSFM In-Context Learning (NeurIPS 2026 · similarity 0.87)
- STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · similarity 0.87)
- $S^2$IP-LLM: Semantic Space Informed Prompt Learning with LLM for Time Series Forecasting (ICML 2024 · 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.