# Practical Adversarial Multivalid Conformal Prediction

**Authors:** Osbert Bastani, Varun Gupta, Christopher Jung, Georgy Noarov, Ramya Ramalingam, Aaron Roth  
**arXiv:** [2206.01067](https://arxiv.org/abs/2206.01067)  
**Date:** 2022-06-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.01067v1/

## Key points

- The paper introduces a method called MVP for reliable predictions in sequential tasks against .
- MVP does not require a held-out validation set and uses all available data for training models.
- MVP offers stronger coverage guarantees through threshold calibrated prediction sets and conditional coverage guarantees.
- The authors provide theoretical analysis and empirical evaluations comparing MVP to split conformal prediction.
- MVP showcases benefits in real-world scenarios, including improved group-wise coverage and flexibility in analysis within specific subsets of interest.
- Overall, MVP addresses limitations of existing approaches and provides strong evidence for its effectiveness in practical applications.

## Summary

The paper introduces the  method called , which aims to provide reliable predictions for sequential tasks against . Unlike other methods, MVP does not require a held-out validation set and allows for the use of all available data for training models. It offers stronger coverage guarantees in two ways: through threshold calibrated prediction sets that maintain correct empirical coverage regardless of the chosen threshold, and by allowing users to specify arbitrary subsets of the feature space with conditional coverage guarantees. The authors provide theoretical analysis and extensive empirical evaluations to support the effectiveness of MVP, comparing it to split conformal prediction and showcasing its benefits in real-world scenarios. These include improved group-wise coverage and flexibility in analysis within specific subsets of interest. Overall, MVP addresses limitations of existing approaches and provides strong evidence for its effectiveness in practical applications.