API: Estimators

Bases: _DEUPBase, RegressorMixin

Direct Epistemic Uncertainty Prediction for regression.

Parameters:

Name	Type	Description	Default
`base_model`	`Any`	The regressor whose uncertainty we estimate. Defaults to :class:`~sklearn.ensemble.HistGradientBoostingRegressor`.	`None`
`error_model`	`Any`	Secondary error predictor `g`. Defaults to HGB regressor.	`None`
`features`	`Any`	Optional :class:`~deup.core.features.FeaturePipeline` for stationarizing features (default: raw `X`).	`None`
`aleatoric`	`Any`	Optional aleatoric estimator `a(x)` (e.g. :class:`~deup.core.aleatoric.Heteroscedastic`). When set, `predict_epistemic` returns `max(0, g - a)`.	`None`
`cv`	`Any`	An int (`KFold` folds) or any splitter with `split(X, y, groups)`.	`5`
`loss`	`Any \| None`	Error-target loss (`"squared"` by default).	`None`
`target_transform`	`TargetTransform \| None`	Stabilization for `g`'s target: `"log"` (default), `"asinh"`, `"none"`.	`None`
`decompose`	`bool`	If `True`, always apply `max(0, g - a)` even when `aleatoric` is `None` (with `a=0` this is a no-op).	`False`
`random_state`	`int \| None`	Seed when `cv` is an int.	`None`

Attributes:

Name	Type	Description
`base_model_, error_estimator_, oof_, aleatoric_`		Fitted components.

`predict(X, return_uncertainty=False, groups=None)`

Predict, optionally returning (prediction, epistemic_uncertainty).

Bases: _DEUPBase, ClassifierMixin

Direct Epistemic Uncertainty Prediction for classification.

Uses predict_proba for OOF error collection and defaults to logloss as the error target.

Parameters:

Name	Type	Description	Default
`base_model`	`Any`	Classifier whose uncertainty we estimate. Defaults to HGB classifier.	`None`
`loss`	`Any \| None`	`"logloss"` (default) or `"brier"`.	`None`
`cv`	`Any`	An int (`StratifiedKFold`) or custom splitter.	`5`
`features`	`Any`	Same as :class:`DEUPRegressor`.	`None`
`aleatoric`	`Any`	Same as :class:`DEUPRegressor`.	`None`
`target_transform`	`Any`	Same as :class:`DEUPRegressor`.	`None`
`error_eps`	`Any`	Same as :class:`DEUPRegressor`.	`None`
`decompose`	`Any`	Same as :class:`DEUPRegressor`.	`None`
`random_state`	`Any`	Same as :class:`DEUPRegressor`.	`None`

Bases: _DEUPBase, RegressorMixin

Direct Epistemic Uncertainty Prediction for cross-sectional ranking.

Defaults to loss="rank" and residualize_rank=True so the reported signal is decoupled from mechanical rank geometry (Finding 3). Requires groups at fit (e.g. dates). Defaults to :class:~deup.splitters.PurgedWalkForward when cv is an int.

Parameters:

Name	Type	Description	Default
`base_model`	`Any`	The ranker / regressor whose ordering uncertainty we estimate.	`None`
`residualize_rank`	`bool`	If `True` (default), apply isotonic rank-geometry residualization to `g`.	`True`
`features`	`Any`	Same as :class:`DEUPRegressor`.	`None`
`aleatoric`	`Any`	Same as :class:`DEUPRegressor`.	`None`
`target_transform`	`Any`	Same as :class:`DEUPRegressor`.	`None`
`error_eps`	`Any`	Same as :class:`DEUPRegressor`.	`None`
`decompose`	`Any`	Same as :class:`DEUPRegressor`.	`None`
`random_state`	`Any`	Same as :class:`DEUPRegressor`.	`None`

Notes

groups is required at fit(X, y, groups=...) and should be passed to predict(..., groups=...) / predict_epistemic(..., groups=...) at inference for correct within-date residualization.