摘要
Uncertainty quantification in machine learning remains a fundamental challenge in safety-critical domains: point estimates provide no reliability information, while traditional probabilistic approaches rely on distributional assumptions that frequently fail in practice. Conformal prediction addresses this by wrapping any model in a distribution-free calibration layer that produces prediction intervals with finite-sample validity guarantees under exchangeability alone. This project investigates conformal prediction across six experimental settings of increasing complexity, examining where the framework succeeds, where its guarantees fall short of practical usefulness, and what determines the gap between the two. We implement and compare multiple conformal methods, including Split-CP, Conformalized Quantile Regression, Scalar Uncertainty CP, and Adaptive Conformal Inference, across tabular regression, spatial modeling, image regression, multi-classclassification, and electricity price forecasting. Three frontiers are investigated: spatial data, where geographic correlation violates exchangeability and a single global quantile produces uneven coverage across regions; heteroscedastic neural networks, where learned uncertainty estimates carry genuine discriminative signal but collapse under distribution shift; and temporal data, where regime nonstationarity renders static calibration inadequate and adaptive recalibration becomes essential. The results confirm that valid marginal coverage is reliably achieved across all settings. Efficiency, however, is not guaranteed by construction. Adaptive methods produce intervals up to 40% narrower than constant-width baselines, but exhibit substantially larger coverage degradation under covariate shift. Base-model quality sets the efficiency ceiling, with the largest gains concentrated at low model complexity and diminishing returns thereafter. In the electricity price setting, we find that a richer four-parameter Johnson’s SU distribution reduces tail misspecification relative to a Normal baseline in stressed regimes, but cannot substitute for adaptive calibration during abrupt regime breaks, and the two are complementary rather than competing strategies. Taken together, the findings argue that conformal prediction is best understood not as a plug-in guarantee, but as a calibration layer whose practical value depends jointly on method choice, base-model quality, and the stationarity of the deployment environment.