← Back to Machine Learning (Statistics)
stat.ML

How do we fairly split disaster costs without anyone leaving?

Ieva Kazlauskaite

June 4, 2026

When groups pool financial risk from rare disasters, some members may end up worse off and defect. This work formalizes a "Certified Allocation Problem": devise a cost-sharing rule, cap each person's obligation, and prove nobody is harmed—all from finite data without assuming any distribution. Conformal Risk Sharing combines a tunable sharing policy with split conformal calibration to generate distribution-free guarantees for each agent. Tests on precipitation and energy cooperatives show it cuts extreme obligations for high-risk members while protecting others.
Published as Conformal Risk Sharing: Certified Cost Allocation with Participation Guarantees arXiv:2606.06391
Read the original paper →