nonconform: Conformal Anomaly Detection in Python

Turn anomaly scores into calibrated decisions.

Traditional anomaly detectors output scores and require arbitrary thresholds. nonconform converts raw scores to conformal p-values and supports principled False Discovery Rate (FDR) control for rigorous decision-making.

The short version for practitioners:

• Use ConformalDetector.select(...) when you need decisions, not just scores.
• Treat every guarantee as conditional on its assumptions: exchangeability for standard conformal workflows, and covariate-shift assumptions plus reliable weights for weighted workflows.
• Use the docs to check whether your data collection process matches those assumptions before relying on the error-control claims.

The Problem

# Traditional approach: arbitrary threshold, no formal error control
scores = detector.decision_function(X_test)
anomalies = scores > 0.5

The Solution

from sklearn.ensemble import IsolationForest

from nonconform import ConformalDetector, Split

detector = ConformalDetector(
    detector=IsolationForest(random_state=42),
    strategy=Split(),
    score_polarity="auto",
)
detector.fit(X_train)

discoveries = detector.select(X_test, alpha=0.05)

When to Use nonconform

Use this library when you need:

• Statistical guarantees with clearly stated assumptions
• Principled thresholds instead of ad hoc cutoffs
• Multiple testing correction
• Calibrated uncertainty for downstream workflows

Citation

If you use nonconform in academic work, reports, or other published material, please cite the accompanying paper:

@misc{hennhöfer2026conformalanomalydetectionpython,
      title={Conformal Anomaly Detection in Python: Moving Beyond Heuristic Thresholds with 'nonconform'},
      author={Oliver Hennhöfer and Maximilian Kirsch and Christine Preisach},
      year={2026},
      eprint={2605.13642},
      archivePrefix={arXiv},
      primaryClass={stat.ML},
      url={https://arxiv.org/abs/2605.13642},
}

Guarantee Scope

nonconform does not make a weak detector "correct." It calibrates detector scores against reference data. The calibration can control false positives only when the reference data, test data, and selection procedure match the documented assumptions.

Workflow	Main assumption	Practical check
Standard conformal	Calibration and test points are exchangeable	Same population, same measurement process, no systematic time/order effect
FDR selection	Input p-values are valid and satisfy the method's dependence assumptions	Prefer select(...); avoid unsupported p-value post-processing
Weighted conformal	Feature distribution may shift, but the anomaly mechanism is stable and supports overlap	Inspect shift, weights, and domain plausibility before trusting WCS results

Quick Links

• Installation
• Quick Start
• User Guide
• Examples
• Common API Workflows
• API Reference

Key Features

• Split conformal inference with finite-sample marginal guarantees
• Detector agnostic design (PyOD, scikit-learn, custom detectors)
• Multiple calibration strategies, including data-efficient resampling variants
• FDR and weighted FDR workflows
• Covariate-shift handling via weighted conformal methods

Installation

pipuv

pip install nonconform

uv add nonconform

Ready to start? Continue with the Quick Start guide.