Moirai: Universal Time-Series Forecasting Transformer
Moirai (Universal Time-Series Forecasting Transformer) · Also known as: Unified Time-Series Transformer, Universal Forecasting Transformer, MOIRAI, Evrensel Zaman Serisi Tahmin Transformatörü
Moirai is a foundation model for universal time-series forecasting introduced by Gerald Woo and colleagues at Salesforce Research in 2024 and presented at ICML. The core idea is to pre-train a single large Transformer on an exceptionally diverse corpus of time-series data (LOTSA) spanning many domains and frequencies, enabling zero-shot and few-shot forecasting on unseen datasets without task-specific retraining. Moirai employs patch-based tokenization, any-variate attention, and a mixture-of-distributions output head to handle variable frequencies, multiple variates, and probabilistic prediction in a unified architecture.
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
Moirai is most appropriate when labeled training data for the target time series is scarce or unavailable, making zero-shot or few-shot forecasting attractive. It suits diverse enterprise settings where many different series with varying frequencies must be forecasted with a single deployable model. Key assumptions include that the target domain is not fundamentally unlike the pre-training corpus. It is less suited when the target series has highly domain-specific structure far outside general temporal patterns, or when interpretability and strict statistical guarantees are required. Alternatives include Chronos, TimesFM, and classical statistical models such as ETS or ARIMA for simpler settings.
Strengths & limitations
- Zero-shot generalization across diverse domains and frequencies without retraining
- Probabilistic output via mixture-of-distributions head provides calibrated uncertainty quantification
- Any-variate attention handles both univariate and multivariate series in a single architecture
- Pre-trained on the large LOTSA corpus, reducing the need for large target-domain datasets
- Inference cost is higher than lightweight statistical models such as ETS or ARIMA
- Performance may degrade on series with highly domain-specific dynamics underrepresented in pre-training data
- The foundation model paradigm offers limited interpretability compared to classical decomposition methods
- Requires careful context window selection; very long historical contexts may challenge memory and latency budgets
Frequently asked
Does Moirai require fine-tuning on target data to be useful?
No. Moirai is designed for zero-shot forecasting, meaning it can produce forecasts on a completely new time series without any task-specific training. However, optional few-shot fine-tuning on a small sample of target data can further improve accuracy when such data is available, without requiring full retraining from scratch.
How does Moirai handle time series with different sampling frequencies?
Moirai supports multiple patch sizes that correspond to different temporal granularities. At inference time, the appropriate patch size is selected based on the frequency metadata of the input series, allowing the same model to process hourly, daily, weekly, and monthly data under a unified architecture without separate model variants.
What kind of uncertainty estimates does Moirai produce?
Moirai's output head parameterizes a mixture of probability distributions — such as Normal, Student-t, and Negative Binomial components — over the forecast horizon. This yields full predictive distributions from which quantile forecasts, prediction intervals, and expected values can be derived, supporting calibrated probabilistic decision-making rather than point forecasts alone.
Sources
- Woo, G., Liu, C., Kumar, A., Xiong, C., Savarese, S., & Sahoo, D. (2024). Unified training of universal time series forecasting transformers. ICML. link ↗
How to cite this page
ScholarGate. (2026, June 2). Moirai (Universal Time-Series Forecasting Transformer). ScholarGate. https://scholargate.app/en/deep-learning/moirai
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 →