TimesFM: A Decoder-Only Foundation Model for Time-Series Forecasting
TimesFM (Time-series Foundation Model) · Also known as: Time-series Foundation Model, Google TimesFM, TimesFM forecaster, Zaman Serisi Temel Modeli
TimesFM is a pre-trained foundation model for univariate time-series forecasting introduced by Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou from Google in 2024. The model adopts a decoder-only transformer architecture, similar in spirit to large language models, and is trained on a large corpus of real-world and synthetic time-series data. Its central innovation is the ability to perform accurate zero-shot forecasting across diverse domains without task-specific 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
TimesFM is well-suited for zero-shot or few-shot univariate time-series forecasting when labeled training data are scarce or when rapid deployment across many heterogeneous series is needed. It assumes relatively stationary or slowly-drifting patterns and works best with regularly-sampled data. It is less appropriate for multivariate forecasting with strong cross-series dependencies, very irregular or sparse observations, or domains requiring interpretable mechanistic models. Alternatives include Chronos, Moirai, and PatchTST for related pre-trained or patch-based approaches.
Strengths & limitations
- Zero-shot generalization: delivers competitive forecasts on unseen time series without fine-tuning
- Scalable patch tokenization reduces sequence length and computational cost compared to point-level transformers
- Decoder-only architecture enables flexible horizon extension through autoregressive generation
- Pre-trained on diverse large-scale corpora, capturing broad temporal patterns across domains
- Primarily designed for univariate forecasting; cross-variable dependencies are not explicitly modeled
- Performance may degrade on highly irregular, sparse, or non-stationary series outside training distribution
- Large pre-trained model size can be resource-intensive for edge or embedded deployment
- Zero-shot capability may still be outperformed by well-tuned domain-specific models on in-distribution tasks
Frequently asked
Does TimesFM require fine-tuning to work on a new dataset?
No. TimesFM is designed for zero-shot forecasting, meaning it can produce predictions on entirely new time series without any additional training. Fine-tuning is possible and may improve performance on a specific domain, but the model is explicitly built and evaluated in the zero-shot regime where no in-domain labeled data are available.
How does patch tokenization differ from standard point-level time-series transformers?
Standard transformers treat each individual time step as a token, leading to very long sequences and quadratic attention cost. Patch tokenization groups consecutive time steps into fixed-size blocks before embedding them, shortening the effective sequence length substantially. This reduces memory and compute requirements and also encourages the model to learn local temporal structure within each patch.
Can TimesFM handle multivariate time series?
TimesFM in its original formulation targets univariate forecasting: each series is processed independently. While it can be applied channel-by-channel to multivariate data, it does not explicitly model cross-variable interactions. For tasks where inter-series correlations are important, multivariate-native models or additional adaptation strategies would be more appropriate.
Sources
- Das, A., Kong, W., Sen, R., & Zhou, Y. (2024). A decoder-only foundation model for time-series forecasting. ICML. link ↗
How to cite this page
ScholarGate. (2026, June 2). TimesFM (Time-series Foundation Model). ScholarGate. https://scholargate.app/en/deep-learning/timesfm
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 →