API: Domain presets

Finance

Bases: BaseEstimator

Panel-data preset for cross-sectional stock rankers.

Defaults: PurgedWalkForward + rank residualization + finance g-features + HealthIndex for per-date context gating.

Parameters:

Name	Type	Description	Default
base_model	Any	Primary ranker (defaults to HGB regressor inside :class:DEUPRanker).	None
date_col	str	Column holding the cross-section date / group label.	'date'
target_col	str | None	Column holding the ranking target. If horizon is set, defaults to f"target_{horizon}d" when that column exists.	None
horizon	int | None	Optional return horizon in days; selects target_{horizon}d when present.	None
feature_cols	list[str] | None	Explicit g-feature columns; otherwise uses :data:FINANCE_G_FEATURES present in the panel.	None
cv	Any	Walk-forward splitter settings (PurgedWalkForward when cv is int).	5
embargo	Any	Walk-forward splitter settings (PurgedWalkForward when cv is int).	5
health_index	HealthIndex | None	Context health scorer (default: three-component :class:HealthIndex).	None

fit(panel_df, y=None)

Fit on a long-format panel DataFrame.

predict_epistemic(panel_df)

Rank-residualized epistemic uncertainty per row.

calibrate(panel_df, *, alpha=0.1)

Conformal-calibrate on a held-out panel split (separate from fit).

health_report(panel_df)

Per-date composite context health (Finding 2 remedy).

rank_coupling_report(panel_df)

Diagnostic: rank-geometry coupling before/after residualization.

Add derived rank/regime columns expected by the finance g-feature preset.

Thesis-equivalent walk-forward g(x) on enriched residual panels.

Fits :class:~deup.core.error_estimator.ErrorEstimator with LightGBM on each expanding fold window — the direct migration of train_g_walk_forward.

Parameters:

Name	Type	Description	Default
enriched	Any	Panel with fold_id, horizon, g-features, and target_col.	required
fold_sort	FoldSort	"numeric" (recommended) sorts fold_02 < fold_10 < fold_100; "string" reproduces legacy thesis lexicographic order for frozen parity.	'numeric'

Subset of :data:FINANCE_G_FEATURES present with sufficient non-null rate.

Tabular

Ergonomic tabular preset — delegates to core DEUP estimators.

Parameters:

Name	Type	Description	Default
base_model	Any	Primary predictor f. When None, chosen from backend.	None
backend	BackendKind	"sklearn" (default), "lgbm", "xgb", or "catboost". Sets default base + error models when those are None.	'sklearn'
error_model	Any	Secondary error predictor g. When None and backend is a GBM, uses the same family (e.g. LightGBM for backend="lgbm").	None
task	TaskKind	"regression" (default) or "classification".	'regression'
cv	Any	Forwarded / used as in the core estimator.	5
random_state	Any	Forwarded / used as in the core estimator.	5
include_raw	Any	Forwarded / used as in the core estimator.	5

backend property

Configured gradient-boosting / sklearn backend.

estimator property

Underlying core estimator (for advanced composition).

Default f model for a tabular backend.

Default g-features for i.i.d. tabular data: raw X + log-density.

Vision

Classification preset: embedding → density + variance → g(x).

Parameters:

Name	Type	Description	Default
embedding	BaseEstimator | Callable[[ArrayLike], ArrayLike] | None	Optional sklearn transformer or callable mapping raw inputs to embeddings. Defaults to :class:IdentityEmbedding (flatten tensors). Inputs are embedded once at the API boundary so the base classifier always sees 2-D arrays.	None
cv	Any	KFold folds when int (i.i.d. vision batches).	5

Bases: BaseEstimator, TransformerMixin

Embed inputs, then append density + variance features (vision preset glue).

Bases: BaseEstimator, TransformerMixin

Flatten tensors to 2-D embeddings (test / baseline without a CNN).