mapie

Mapie, short for Model-Agnostic Prediction Interval Estimation, is an open-source Python library that enables the generation of predictive intervals for any regression model. It provides a practical means of quantifying uncertainty in point predictions by constructing model-agnostic prediction intervals.

The library implements conformal prediction methods that offer valid coverage guarantees under the assumption of exchangeable data. It supports a variety of procedures, including split conformal, cross-conformal, and jackknife-based approaches such as jackknife+, as well as cross-validated variants. Mapie is designed to be model-agnostic and integrates with common machine learning workflows, particularly those that use scikit-learn-compatible estimators.

Key features include compatibility with any regressor that adheres to the scikit-learn API, automatic handling of calibration or hold-out sets, control over the miscoverage level, and the ability to produce both single-shot and cross-validated prediction intervals. The library also offers utilities for evaluating interval quality, visualization, and easy incorporation into existing pipelines.

Usage typically involves training a base regression model and wrapping it with Mapie to produce prediction intervals for new samples. The result is a lower and upper bound that form the predictive interval, reflecting the uncertainty of the model’s prediction.

Mapie serves researchers and practitioners seeking principled, model-agnostic uncertainty quantification in regression tasks. Related topics include conformal prediction, uncertainty quantification, and Python-based machine learning workflows.