Conformalized Quantile Regression

Conformalized quantile regression (CQR) is a statistical method that combines the strengths of quantile regression and conformal prediction to generate prediction intervals with guaranteed coverage probabilities, even for data with complex heteroscedasticity. Current research focuses on improving CQR's adaptability to various data structures, including time series and graph data, and extending its application to diverse machine learning models like Deep Operator Networks and graph neural networks. This approach offers a distribution-free, robust method for uncertainty quantification, impacting fields ranging from chip manufacturing and autonomous driving to hyperparameter optimization and time series forecasting by providing reliable prediction intervals without strong distributional assumptions.