Zurück zu den Modellen
Forecaster

FlowStateForecaster

API-ReferenzQuellcode
Categorical in XPred int

Zero-shot forecaster wrapping IBM FlowState via granite-tsfm.

FlowState, developed by IBM Research, is an encoder-decoder architecture, employing an S5-based encoder and a functional basis decoder.

Univariate only. Implementation adapted from [1].

Schnellstart

python
from sktime.forecasting.flowstate import FlowStateForecaster

estimator = FlowStateForecaster(model_path: str='ibm-research/flowstate', revision: str='r1.1', scale_factor: float=1.0, config: dict | None=None, batch_first: bool=True, prediction_type: str='mean')

Parameter(6)

model_pathstr, default=”ibm-research/flowstate”
Hugging Face model id or local path.
revisionstr, default=”r1.1”

Model revision on the Hugging Face Hub. Always forwarded to from_pretrained; do not duplicate in config.

scale_factorfloat, default=1.0
Temporal scaling passed to the model at predict time.
configdict, optional, default=None

Extra kwargs for FlowStateForPrediction.from_pretrained.

batch_firstbool, default=True

past_values layout for the model.

prediction_type{“mean”, “median”}, default=”mean”
Point forecast type passed to the model.

Beispiele

>>> from sktime.datasets import load_airline
>>> from sktime.forecasting.flowstate import FlowStateForecaster
>>> from sktime.split import temporal_train_test_split
>>> y = load_airline ()
>>> y_train, _ = temporal_train_test_split (y)
>>> f = FlowStateForecaster ()
>>> f. fit (y_train)
>>> y_pred = f. predict (fh = [1, 2, 3 ])

Referenzen

[2]

Graf et al., FlowState: Sampling Rate Invariant Time Series Forecasting, arXiv:2508.05287