Back to models
Forecaster

T0Forecaster

API referenceSource
Categorical in XPred intMissing valuesExogenousMultivariate

Interface to the T0 zero-shot forecaster by The Forecasting Company.

T0 is a pretrained time series foundation model for zero-shot forecasting, released by The Forecasting Company [1], [2]. The model is loaded from a pretrained checkpoint and applied without task-specific training: the context is stored in fit and forecasts are produced zero-shot in predict.

T0 natively supports known-future covariates (exogenous data), passed to the underlying model via its future_covariates argument. These covariates must span the context window and the forecast horizon, so exogenous data must be provided both in fit (past values) and predict (future values).

Capabilities and behaviour:

  • Multivariate - a multi-column y is forecast column-by-column, each as an independent series (capability:multivariate=True).

  • Probabilistic - T0 returns quantiles natively, so predict_quantiles, predict_interval and (by default) predict_var are available (capability:pred_int=True); predict returns the median (0.5 quantile).

  • Exogenous - known-future covariates are conditioned on, but must cover every step 1..max(fh), so a non-contiguous fh requires contiguous X (capability:non_contiguous_X=False).

  • Missing values - NaN entries in the context are treated as missing (capability:missing_values=True).

  • In-sample - not supported; fh must be strictly in the future (capability:insample=False).

Quickstart

python
from sktime.forecasting.t0 import T0Forecaster

estimator = T0Forecaster(model_path: str | None='theforecastingcompany/t0-alpha', device: str | None=None, context_length: int | None=None, random_state=None, license_accepted: bool=False, ignore_deps: bool=False)

Parameters(6)

model_pathstr or None, default=”theforecastingcompany/t0-alpha”

Path to the T0 HuggingFace model checkpoint. The default checkpoint is a gated model on the HuggingFace Hub, so downloading it requires accepting the vendor’s terms on the model page and authenticating with a HuggingFace token. If None, a small randomly-initialized T0 model is built locally (from the tfc-t0 model code) instead of downloading any checkpoint - this produces untrained (meaningless) forecasts and is intended only for testing or offline use.

devicestr or None, default=None
Device for inference, e.g., “cpu”, “cuda”, or “mps”. If None, uses “cuda” when a CUDA device is available, otherwise “cpu”.
context_lengthint or None, default=None
Maximum context length for inference. If None, the full context is used.
random_stateint, RandomState instance or None, optional, default=None

Random seed for reproducibility, sklearn-compatible. If None, no seeding is applied and the ambient random state is used. If set, inference runs in a forked RNG seeded from random_state, leaving the global RNG untouched. T0’s inference is deterministic, so this does not change the forecast; it is accepted for interface consistency.

license_acceptedbool, optional, default=False

Whether the user has read and accepted the license terms of the tfc-t0 package and the T0 models, licensed by The Forecasting Company. Must be set to True to use T0Forecaster; otherwise fit raises. To view the license, call T0Forecaster.print_license(); the model card and gated-access terms are at https://huggingface.co/theforecastingcompany/t0-alpha.

ignore_depsbool, optional, default=False
If True, dependency checks are skipped.

Examples

Univariate point forecast (zero-shot):
>>> from sktime.datasets import load_airline
>>> from sktime.forecasting.t0 import T0Forecaster
>>> from sktime.split import temporal_train_test_split
>>> y = load_airline ()
>>> y_train, y_test = temporal_train_test_split (y, test_size = 12)
>>> forecaster = T0Forecaster (license_accepted = True)
>>> forecaster. fit (y_train)
>>> y_pred = forecaster. predict (fh = [1, 2, 3 ]) Probabilistic forecast (T0 returns quantiles natively):
>>> y_quantiles = forecaster. predict_quantiles (
... fh = [1, 2, 3 ], alpha = [0.1, 0.5, 0.9 ]
... )
>>> y_interval = forecaster. predict_interval (
... fh = [1, 2, 3 ], coverage = 0.9
... ) Forecast with known-future exogenous data (X in both fit and predict):
>>> from sktime.datasets import load_longley
>>> y, X = load_longley ()
>>> y_tr, y_te, X_tr, X_te = temporal_train_test_split (y, X, test_size = 3)
>>> forecaster = T0Forecaster (license_accepted = True)
>>> forecaster. fit (y_tr, X = X_tr)
>>> y_pred = forecaster. predict (fh = [1, 2, 3 ], X = X_te) Multivariate forecast on a real dataset (each column forecast independently):
>>> from sktime.datasets import load_longley
>>> _, y_multi = load_longley () # 5-column economic indicators frame
>>> y_multi_train = y_multi. iloc [: - 3 ]
>>> forecaster = T0Forecaster (license_accepted = True)
>>> forecaster. fit (y_multi_train)
>>> y_pred = forecaster. predict (fh = [1, 2, 3 ])

References