Sundial: Generative Time-Series Foundation Models
Sundial (Generative Time-Series Foundation Models) · Also known as: Sundial TSF, Time-Series Foundation Model (Generative), Sundial ICML 2025, Zaman Serisi Temel Modeli (Sundial)
Sundial is a family of generative time-series foundation models introduced by Yong Liu and colleagues at Tsinghua University (ICML 2025). Pre-trained on large and diverse time-series corpora, Sundial employs a decomposition-based architecture paired with a generative forecasting head to produce probabilistic multi-horizon forecasts. It represents a shift toward general-purpose, zero-shot-capable models for real-world temporal prediction tasks.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Sundial is well-suited for multi-variate or univariate forecasting tasks where labeled training data is scarce, domain transfer is needed, or probabilistic forecasts are required. It assumes the input can be meaningfully decomposed into trend and seasonal components. Limitations include high computational cost relative to lightweight baselines, potential miscalibration when the target domain differs substantially from pre-training data, and sensitivity to patch-length hyperparameters. Simpler alternatives such as N-BEATS or PatchTST may suffice for well-labeled, single-domain datasets.
Strengths & limitations
- Zero-shot and few-shot generalization across diverse time-series domains without task-specific retraining
- Probabilistic output via generative head enables native uncertainty quantification
- Decomposition inductive bias improves handling of non-stationary and multi-period signals
- Scalable model family allows balancing compute budget against forecasting capacity
- Large pre-training scale requires substantial compute and data infrastructure to reproduce
- Generative sampling at inference increases latency compared to deterministic point-forecast models
- May underperform lightweight specialized models on narrow, well-labeled single-domain tasks
- Patch length and decomposition hyperparameters require tuning and affect out-of-domain performance
Frequently asked
Does Sundial require fine-tuning on new data before deployment?
No. Sundial is designed for zero-shot forecasting, meaning it can be applied directly to new time-series domains without any task-specific fine-tuning. However, few-shot fine-tuning on a small labeled sample from the target domain can further improve calibration and accuracy when labeled data is available.
How does Sundial differ from Chronos and TimesFM?
All three are large pre-trained time-series foundation models, but Sundial emphasizes a generative forecasting head that produces full predictive distributions rather than point estimates or quantile outputs. Its architecture also incorporates an explicit trend-seasonal decomposition step, which distinguishes it from the tokenization-only approach in Chronos and the patching strategy in TimesFM.
Is Sundial appropriate for high-frequency financial tick data?
Sundial can be applied to high-frequency data, but its pre-training corpus composition may underrepresent tick-level financial signals, potentially limiting zero-shot calibration. In such settings, domain-adaptive fine-tuning and careful patch-length configuration are recommended to align the model's inductive biases with the statistical properties of financial microstructure data.
Sources
- Liu, Y., Qin, G., Shi, X., Hu, T., Wang, J., & Long, M. (2025). Sundial: A family of highly capable time series foundation models. ICML. link ↗
How to cite this page
ScholarGate. (2026, June 2). Sundial (Generative Time-Series Foundation Models). ScholarGate. https://scholargate.app/en/deep-learning/sundial
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
Compare side by side →