New conformal prediction method improves spatial event forecasting

Conformal prediction is a powerful framework for uncertainty quantification, but applying it to spatial events—like cyclone genesis or earthquake epicenters—has been challenging because predictions are sets of points rather than scalar values. In a new preprint, Collin Nill, Trevor Harris, and Jason Adams introduce Manifold Constrained Conformal Prediction (MCCP). Their approach represents spatial point clouds as empirical measures and scores them using sliced Wasserstein distance. To ensure practical validity, they constrain the resulting distribution-valued prediction sets to be supported only near the training data manifold, deriving a coverage lower bound that can be made small with a data-adaptive selection criterion.

Because the intersected sets are not analytically tractable, the authors propose a modified flow-based sampling procedure to represent and apply them as ensembles. In numerical experiments on synthetic data, tropical cyclone genesis, and earthquake occurrences, MCCP achieves near-nominal coverage (e.g., 90% nominal coverage with actual coverage within 1–2%) while significantly reducing both energy distance and manifold distance compared to highest predictive density region (HDR) baselines and generative model baselines. The method offers a principled way to quantify uncertainty in high-stakes geospatial forecasting tasks, where accurate risk assessment is critical.