Private Prediction Intervals

Machine learning models often return a single answer; conformal prediction adds a range that is designed to contain the true outcome at a chosen rate, such as 90%. Differential privacy adds noise so no one record can be reverse-engineered from the output. (arxiv.org) A new April 16, 2026 paper from Jiamei Wu, Ce Zhang, Zhipeng Cai, Jingsen Kong, Bei Jiang, Linglong Kong, and Lingchen Kong combines those two ideas in a method they call Differentially Private Conformal Prediction, or DPCP. The paper says the method gives end-to-end privacy guarantees while producing prediction sets with conformal coverage properties. (arxiv.org) Earlier private conformal methods usually split data into one part for training and another for calibration. The 2021 “Private Prediction Sets” paper by Anastasios Angelopoulos, Stephen Bates, Tijana Zrnic, and Michael I. Jordan used that split-conformal setup with a privatized quantile step. (arxiv.org) The new paper argues that split calibration wastes samples and widens intervals. Its core move is a non-splitting conformal procedure that uses the stability of differentially private mechanisms to connect private calibration to the “oracle” version that would use all the data. (arxiv.org) That matters in settings where the interval itself is part of the product, not a side statistic. In health care, finance, and criminal justice, the authors write, overconfident point predictions can turn model errors into unsafe decisions. (arxiv.org) The paper reports that, under the same privacy budget, DPCP can produce tighter prediction sets than existing private split-conformal approaches. A separate March 2026 paper from Purdue University and the University of Pittsburgh reached a similar conclusion for a full-data private conformal framework, reporting sharper sets than split-based private baselines. (arxiv.org 1) (arxiv.org 2) The tradeoff is familiar in privacy research: stronger privacy usually means more noise, and more noise usually means less precise outputs. DPCP is aimed at reducing that precision loss without dropping the differential privacy guarantee. (arxiv.org 1) (arxiv.org 2) The paper is an arXiv preprint, not a peer-reviewed journal article, and its abstract points to synthetic and real-data experiments rather than a production deployment. For teams building audited machine-learning systems, the practical question is whether private uncertainty estimates can stay narrow enough to be useful. (arxiv.org) If that holds up, the change is simple to describe: keep the privacy guarantee, stop throwing away data for calibration, and return a narrower range around each prediction. That is the problem this line of work is trying to solve. (arxiv.org)