Research map / Time series / Foundation models for time series
VerbalTS: Generating Time Series from Texts
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.
Introduces VerbalTS, a diffusion-based framework that generates multivariate and univariate time series directly from unstructured text using a multi-focal text processor and multi-view noise estimator.
Problem
Time series synthesis is usually conditioned on structured metadata or class labels, which lose sample-specific details and sequential event information and require costly expert labeling; the paper targets generating time series from free-form text.
Approach
Reprograms text embeddings into learnable anchor-based representations for temporal resolutions, spatial variables, and diffusion stages, and injects them via adaptive layer-norm-style gating into a multi-view (temporal, spatial, diffusion) diffusion noise estimator.
Key results
- On Synth-M, VerbalTS reaches FID 29.05 and J-FTSD 32.91 versus 41.94 and 58.08 for the best prior baseline TEdit, with CTTP 64.13 vs 52.55.
- On Weather, VerbalTS achieves FID 6.13 vs 14.86 for TEdit, a large fidelity gain on the real-world dataset.
- Ablation removing all anchor vectors (spatial, temporal, diffusion) degrades Synth-M FID from 29.05 to 34.18 and CTTP from 64.13 to 60.07.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use text conditions with VERBALTS instead of the strongest structured-condition baseline | Time series generation · Synth-M | FID: 41.94±1.92 -> 29.05±0.62 | Helped |
| Use text conditions with VERBALTS instead of the strongest structured-condition baseline | Time series generation · BlindWays | CTTP: 7.96±0.41 -> 9.18±0.26 | Helped |
| Use text conditions with VERBALTS instead of the strongest structured-condition baseline | Time series generation · Weather | JFTSD: 18.33±0.30 -> 8.56±0.21 | Helped |
| Use text conditions with VERBALTS instead of the strongest structured-condition baseline | Time series generation · Synth-U | CTTP: 28.09±0.06 -> 37.56±0.07 | Helped |
| Use text conditions with VERBALTS instead of the strongest structured-condition baseline | Time series generation · ETTm1 | FID: 26.84±0.91 -> 24.19±0.53 | Helped |
| Use text conditions with VERBALTS instead of the strongest structured-condition baseline | Time series generation · Traffic | CTTP: 8.52±0.08 -> 8.43±0.03 | No clear effect |
| Add spatial semantic anchors (ablation) | VERBALTS · Synth-M | FID: 29.98±0.68 -> 29.05±0.62 | Helped |
| Add spatial semantic anchors (ablation) | VERBALTS · Weather | FID: 7.01±0.33 -> 6.13±0.21 | Helped |
| Add temporal semantic anchors after spatial anchors (ablation) | VERBALTS · Synth-M | FID: 34.17±0.56 -> 29.98±0.68 | Helped |
| Add temporal semantic anchors after spatial anchors (ablation) | VERBALTS · Weather | FID: 7.30±0.25 -> 7.01±0.33 | Helped |
| Add diffusion semantic anchors after spatial and temporal anchors (ablation) | VERBALTS · Synth-M | FID: 34.18±0.39 -> 34.17±0.56 | No clear effect |
| Add diffusion semantic anchors after spatial and temporal anchors (ablation) | VERBALTS · Weather | FID: 7.60±0.41 -> 7.30±0.25 | Helped |
Surprising findings
- Structured-condition baseline TimeWeaver learns faster early in training, but VerbalTS overtakes it later, which the authors attribute to the richer information in unstructured text.
Limitations
- Slow generation efficiency due to long reverse diffusion process
- Text-time series pairs are scarce, so several datasets rely on manually constructed or externally annotated text
- Evaluation relies on a self-trained CTTP proxy model for semantic alignment
Experimental setup
Datasets: Synth-U Synth-M Weather BlindWays ETTm1 Traffic
Benchmarks: Synth-M Synth-U BlindWays Weather ETTm1 Traffic
Metrics: FID J-FTSD CTTP score
Baselines: TimeVQVAE DiffTime TimeWeaver TEdit
Builds on: DDIM conditional diffusion model (CSDI-style conditional diffusion) TEdit TimeWeaver adaptive layer normalization
Closest papers by meaning
- Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.87)
- STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · similarity 0.86)
- TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts (AAAI 2025 · similarity 0.86)
- TimeTok: Granularity-Controllable Time-Series Generation via Hierarchical Tokenization (NeurIPS 2026 · similarity 0.86)
- BRIDGE: Bootstrapping Text to Control Time-Series Generation via Multi-Agent Iterative Optimization and Diffusion Modeling (ICML 2025 · similarity 0.86)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 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.