Forecaster#

class openstef_models.models.forecasting.forecaster.Forecaster(**data: Any) → None[source]

Bases: BaseConfig, BatchPredictor[ForecastInputDataset, ForecastDataset], ABC

Base for forecasters that handle multiple horizons simultaneously.

Designed for models that train and predict across multiple prediction horizons in a unified manner. These models handle the complexity of different lead times internally, providing a simpler interface for multi-horizon forecasting.

Ideal for models that can share parameters or features across horizons, avoiding the need to train separate models for each prediction distance.

Concrete forecasters subclass this directly and declare their fields (quantiles, horizons, hyperparams, etc.) as Pydantic model fields. Mutable runtime state (e.g. the underlying ML model) should be stored in PrivateAttr fields and initialised in model_post_init.

Invariants

Predictions must include all quantiles specified in the configuration
predict_batch() only called when supports_batching returns True

Example

Creating a forecaster with multiple horizons

>>> from openstef_core.types import LeadTime, Quantile
>>> from openstef_models.models.forecasting.flatliner_forecaster import FlatlinerForecaster
>>> forecaster = FlatlinerForecaster(
...     quantiles=[Quantile(0.1), Quantile(0.5), Quantile(0.9)],
...     horizons=[LeadTime.from_string("PT1H"), LeadTime.from_string("PT6H")],
... )
>>> len(forecaster.horizons)
2
>>> str(forecaster.max_horizon)
'PT6H'