Bidirectional calibration-robust booking control under heterogeneous cancellation risk

Reservation-based capacity systems often rely on predicted cancellation or show-up probabilities when deciding how much demand to accept before service realization. Existing hotel booking and overbooking models typically treat these probabilities as fixed inputs or merge different forecast errors into a single uncertainty source. Less is known about how a capacity-constrained decision-maker should protect a segment-level acceptance portfolio when heterogeneous show-up probabilities are imperfectly calibrated and aggregate residual load remains distributionally ambiguous. This paper develops a calibration-robust optimization framework for this problem. Segment-level probability miscalibration is modeled by a weighted budgeted uncertainty set, while ambiguity in the aggregate residual load law is modeled by a 1-Wasserstein ambiguity set. The robust booking control model combines asymmetric vacancy and overflow costs with a distributionally robust conditional value-at-risk (CVaR) service-risk constraint. The main theoretical contribution is the bidirectional calibration principle: The locally adverse probability perturbation is segment specific and reverses when the marginal value of capacity crosses a segment-specific value-to-load threshold. We derive finite reformulations and a certified scenario-and-piece cutting-plane algorithm with finite convergence for the finite piecewise-linear model. Controlled hotel-booking experiments check the direction-reversal mechanism, quantify the separate and joint effects of the two uncertainty layers, and evaluate scalability. A public hotel booking data set is used only as a real-world-calibrated illustration of segment heterogeneity and calibration bands. The results show that separating calibration risk from aggregate load ambiguity changes the portfolio composition, capacity protection, and the auditable price of robustness.