Online Conformal Prediction

Online conformal prediction provides a distribution-free method for generating prediction sets with guaranteed coverage probabilities, even when dealing with non-stationary or adversarial data streams. Current research focuses on improving efficiency and validity under various feedback scenarios (e.g., full, semi-bandit), incorporating Bayesian methods for robustness, and adapting to time-correlated data through techniques like blocking and decoupling. This approach is significant for its rigorous uncertainty quantification, enabling reliable predictions in diverse online learning settings, including those with limited feedback or evolving data distributions, and finds applications in areas like time series forecasting and online decision-making.