TimeGPT
A Time Series Foundation Model · Also known as: TimeGPT-1, Time series GPT
TimeGPT is a time series foundation model introduced by Garza and White in 2023 that unifies forecasting, anomaly detection, and classification in a single pre-trained model. Inspired by large language models, TimeGPT is pre-trained on diverse time series and transfers well to downstream tasks with minimal fine-tuning.
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
TimeGPT is ideal for forecasting when labeled data for the specific task is limited, as pre-training provides strong initialization. Use it for multi-step ahead forecasting with uncertainty quantification. For large task-specific datasets, domain-specific models may outperform transfer learning. TimeGPT excels in cross-domain applications where knowledge transfer is valuable.
Strengths & limitations
- Pre-trained on massive diverse time series corpus, learning universal patterns applicable to multiple domains
- Enables few-shot learning and strong transfer learning to downstream tasks with minimal fine-tuning
- Provides uncertainty estimates alongside point forecasts, useful for risk-aware decision making
- Handles diverse time series characteristics (different scales, frequencies, lengths) through learned representations
- Proprietary model with API access only; not available for on-premise deployment or modification
- Fine-tuning requires careful data preparation and validation to avoid negative transfer
- Performance may underperform specialized models trained extensively on specific domains with large datasets
Frequently asked
How does TimeGPT differ from task-specific forecasting models?
Task-specific models are trained from scratch on historical data for each application, requiring substantial data collection and training. TimeGPT is pre-trained on massive diverse data, learning general patterns across domains. Fine-tuning on task-specific data typically requires less data and training time than training from scratch, enabling faster deployment and better few-shot performance.
What pre-training data was used for TimeGPT?
TimeGPT was pre-trained on a diverse corpus of time series from multiple industries including energy, finance, transportation, and weather. The exact composition is not fully disclosed, but the goal was coverage across domains and frequencies to learn universal time series patterns. This diversity enables good transfer to new domains.
Can TimeGPT models be fine-tuned for specific applications?
Yes, TimeGPT supports fine-tuning on task-specific data, which typically improves performance. Fine-tuning adapts learned representations to specific domain characteristics. However, fine-tuning requires careful validation to ensure the model learns domain patterns without overfitting. Cross-validation and regularization are important during fine-tuning.
Sources
- Garza, F., & White, C. W. (2023). TimeGPT-1: A Time Series Foundation Model. In ICML 2024 Time Series Workshop. link ↗
How to cite this page
ScholarGate. (2026, June 3). A Time Series Foundation Model. ScholarGate. https://scholargate.app/en/deep-learning/timegpt
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.
- Latent Diffusion ModelsDeep learning↔ compare
- Mamba (State Space Model)Deep learning↔ compare
- N-BEATSxDeep learning↔ compare
- Vision TransformerDeep learning↔ compare