Abstract
This paper introduces Conformal Thresholded Intervals (CTI), a novelconformal regression method that aims to produce the smallest possibleprediction set with guaranteed coverage. Unlike existing methods that rely onnested conformal framework and full conditional distribution estimation, CTIestimates the conditional probability density for a new response to fall intoeach interquantile interval using off-the-shelf multi-output quantileregression. CTI constructs prediction sets by thresholding the estimatedconditional interquantile intervals based on their length, which is inverselyproportional to the estimated probability density. The threshold is determinedusing a calibration set to ensure marginal coverage. Experimental resultsdemonstrate that CTI achieves optimal performance across various datasets.