deup

Direct Epistemic Uncertainty Prediction (DEUP) for any scikit-learn model — with first-class, leakage-correct time-series support.

DEUP estimates epistemic uncertainty by training a secondary error predictor on your model's out-of-sample errors — no ensembles, no Bayesian retraining, works with the model you already use.

Method credit: DEUP is due to Lahlou et al., 2023 (TMLR). This package is a maintained, benchmarked, scikit-learn-compatible implementation with time-series / cross-sectional finance support and aggregation-reliability diagnostics.

Repository: https://github.com/ursinasanderink/deup · Docs: https://ursinasanderink.github.io/deup/

Quickstart

from sklearn.ensemble import RandomForestRegressor
from deup import DEUPRegressor

model = DEUPRegressor(base_model=RandomForestRegressor())
model.fit(X_train, y_train)

pred, unc = model.predict(X_test, return_uncertainty=True)

Tabular gradient boosting (LightGBM / XGBoost / CatBoost):

from deup.domains.tabular import TabularDEUP

model = TabularDEUP(backend="lgbm", cv=5).fit(X_train, y_train)
unc = model.predict_epistemic(X_test)

Time-series / cross-sectional finance (flagship preset):

from deup.domains.finance import CrossSectionalDEUP

model = CrossSectionalDEUP(horizon=20).fit(panel_df)
pred, unc = model.predict(test_df, return_uncertainty=True)
health = model.health_report(test_df)

Install

pip install deup                 # core (numpy + scikit-learn)
pip install "deup[gbm]"          # + LightGBM (TabularDEUP backend)
pip install "deup[xgb]"          # + XGBoost
pip install "deup[catboost]"     # + CatBoost
pip install "deup[gbm-all]"      # all gradient-boosting backends
pip install "deup[finance]"      # + pandas (CrossSectionalDEUP)
pip install "deup[docs]"         # + MkDocs site locally

Why this package?

The only public DEUP code was a stale research repo of notebooks — no maintained pip-installable package. Major UQ libraries (torch-uncertainty, uncertainty-toolbox, MAPIE) don't implement DEUP. deup fills that gap with leakage-correct OOF error construction and walk-forward / purged CV for time-series and finance.

Comparison (California housing, Spearman ρ vs realized squared error)

Method	ρ	Notes
DEUP	0.509	OOF error predictor (RF base)
DEUP + LightGBM	0.444	TabularDEUP(backend="lgbm")
DEUP + XGBoost	0.400	TabularDEUP(backend="xgb")
DEUP + CatBoost	0.407	TabularDEUP(backend="catboost")
Ensemble disagreement	0.460	Bootstrap variance
Conformal residual	0.447	|y − ŷ| on cal split
Laplace (last-layer)	0.015	Not applicable to trees

Full results: Benchmarks.

Documentation

Topic	Link
Getting started	docs/getting-started
Five-axis guide	docs/concepts
Domain presets	docs/domains
Tutorials	tabular · finance · conformal · active learning
When is agg-g reliable?	reliability
PyTorch / TorchUncertainty	pytorch-integration
Contributing	CONTRIBUTING.md

Status

v0.4.0 — complete library: core DEUP, conformal calibration, reliability diagnostics, domain presets (tabular GBM backends, finance, vision), benchmarks, tutorials, TorchUncertainty integration.

Citing

Cite this software (CITATION.cff), Lahlou et al. (2023) for the DEUP method, and Sanderink (2026) for cross-sectional ranking and aggregation- reliability extensions:

Sanderink, U. (2026) 'When Alpha Breaks: Two-Level Uncertainty for Safe Deployment of Cross-Sectional Stock Rankers', arXiv preprint arXiv:2603.13252. Available at: https://arxiv.org/pdf/2603.13252

License

Apache-2.0. See LICENSE.