New ALO method dramatically speeds up conformal prediction for ML

Conformal prediction is a powerful framework for uncertainty quantification, but its computational cost limits adoption. Traditional methods require leave-one-out refits for every training point. Recent improvements like Jackknife+ and Jackknife-minmax reduce the cost but still demand retraining each observation. In a new preprint, Jiachen Cong and Jingbo Liu propose using approximate leave-one-out (ALO) estimators to sidestep this bottleneck. The ALO approach, borrowed from high-dimensional statistics, approximates the effect of removing a single data point without actually retraining the model. The authors adapt the theory to conformal prediction, where leave-i-out residuals are needed for predictions at new points, not just at training covariates. They provide asymptotic proofs of coverage and efficiency, and simulation results confirm that the ALO-based methods achieve performance nearly identical to exact methods while dramatically reducing runtime.

The impact could be substantial for practitioners deploying conformal prediction on large-scale systems. By cutting the computational overhead from O(n) refits to a single pass, the ALO method enables uncertainty quantification for datasets with millions of points where exact methods are infeasible. The work also opens the door to further optimizations, as ALO estimators can be combined with existing fast conformal prediction variants. The paper is available on arXiv (2606.31915) and demonstrates that theoretical rigor does not have to come at the cost of practical efficiency.