Skip to content
Edit this page View source of this page

User Guide

This guide is written for practitioners who need calibrated anomaly decisions without working through the full conformal-inference literature first.

Start with the task you have, then read only the theory needed to understand the assumptions behind that task.

Getting Started

If you are new to conformal prediction, start here:

Page Description
Statistical Concepts Plain-language reference for p-values, FDR, exchangeability, power, and shift
Conformal Inference Practitioner-level explanation of what the guarantees mean and when they apply

Core Concepts

Page Description
Conformalization Strategies Split, Cross-Validation, Jackknife+, and Bootstrap strategies explained
Choosing Strategies Decision framework: which strategy to use for your dataset and requirements
Detector Compatibility How to use PyOD, scikit-learn, or your own custom detectors

Advanced Topics

Page Description
FDR Control Control false discovery rates when testing many observations
Weighted Conformal Handle distribution shift between training and test data

Practical Usage

Page Description
Input Validation Parameter constraints and what error messages mean
Batch Evaluation Evaluate performance on labeled test sets
Streaming Evaluation Online evaluation for real-time detection
Exchangeability Martingales Sequential evidence monitoring on streaming conformal p-values
Best Practices Production patterns, data preparation, and model selection
Logging Configure progress bars and debug output
Troubleshooting Solutions to common issues
  1. New to conformal prediction? Start with Statistical Concepts, then Conformal Inference
  2. Choosing a strategy? Read Choosing Strategies
  3. Going to production? Review Best Practices and Troubleshooting
  4. Dealing with distribution shift? Study Weighted Conformal

Validity Checklist

Before relying on a statistical guarantee, confirm:

  • The calibration data is representative of normal data for the target workflow.
  • The detector score direction is configured correctly.
  • Multiple anomaly decisions go through select(...) or another documented FDR procedure.
  • Distribution shift is handled explicitly instead of ignored.
  • Any reported empirical FDR or power uses labels that were not used for training, calibration, or threshold selection.