Forecaster

TimerForecaster

Timer foundation model forecaster.

Wraps the Timer generative pre-trained Transformer for zero-shot time series forecasting via the HuggingFace transformers library.

Timer uses autoregressive generation on continuous time series tokens. The model is pre-trained on the Unified Time Series Dataset (UTSD) covering diverse domains and temporal patterns.

The model is cached using the multiton pattern to avoid reloading weights when multiple forecaster instances share the same model.

Quickstart

python

from sktime.forecasting.timer import TimerForecaster

estimator = TimerForecaster(model_name='thuml/timer-base-84m', context_length=2880, device='cpu')

Parameters(3)

model_namestr, default=”thuml/timer-base-84m”: Name or path of the pre-trained Timer model on HuggingFace. Options include:
“thuml/timer-base-84m” (84M parameters)
“thuml/timer-xl-84m” (Timer-XL variant)
context_lengthint, default=2880: Number of historical observations to use as input context. Timer supports variable context lengths. If the series is shorter, the full series is used.
devicestr, default=”cpu”: Device to run the model on. Options: “cpu”, “cuda”, “cuda:0”, etc.

Examples

>>> from sktime.forecasting.timer import TimerForecaster
>>> from sktime.datasets import load_airline
>>> y = load_airline ()
>>> forecaster = TimerForecaster (
... model_name = "thuml/timer-base-84m",
... )
>>> forecaster. fit (y)
>>> y_pred = forecaster. predict (fh = [1, 2, 3 ])

References

[1]

Liu et al., “Timer: Generative Pre-trained Transformers Are Large Time Series Models”, ICML 2024. https://arxiv.org/abs/2402.02368

[2]

Liu et al., “Timer-XL: Long-Context Transformers for Unified Time Series Forecasting”, ICLR 2025.