Every risk model built on observed history underestimates the true maximum. This is not a calibration problem. It is a structural property of optimal estimation — and it can now be proved. After every ...