Split Conformal Prediction

Split conformal prediction is a non-parametric method for constructing prediction intervals with guaranteed coverage probability, regardless of the underlying data distribution. Current research focuses on improving its performance under various conditions, including non-exchangeable data (e.g., time series), data contamination, and out-of-distribution samples, often employing techniques like kernel density estimation to enhance conditional coverage. This robust and computationally efficient approach is finding applications in diverse fields, from material classification using deep learning models to insurance risk modeling, by providing reliable uncertainty quantification for improved decision-making.