Title: Parameter-Free and Group Conditional Online Conformal Prediction

Abstract

Ensuring robust uncertainty quantification (UQ) is vital for deploying machine learning models in dynamic real-world settings, particularly when data distributions evolve over time and violate the assumption of exchangeability. Existing Online Conformal Prediction (OCP) techniques typically resolve this challenge by prioritizing either group-wise error control or the ability to function without specific learning-rate dependencies. However, achieving group-conditional coverage is paramount for maintaining fairness across diverse data subsets and delivering more granular UQ assurances. Similarly, optimizing algorithms without fixed parameters is essential for resilience against unknown data shifts and adversarial conditions.

In this work, we introduce a novel parameter-free algorithm designed for group-conditional OCP, which we show delivers superior guarantees for group-conditional coverage. Our empirical evaluations, conducted on both synthetic datasets and real-world scenarios, reveal that our approach enhances the reliability of current parameter-free OCP methods. Furthermore, the prediction intervals generated by our method are comparable in width to those produced by well-calibrated group-conditional approaches. By integrating group-conditional coverage with parameter-free online learning, this study establishes a new baseline for fair and robust uncertainty quantification in non-stationary environments.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC