Sharing catastrophe risk when capacity and contract rules get in the way

Why the cheapest layer-by-layer choice can fail

Catastrophe insurance spreads losses among insurers, reinsurers, catastrophe-bond investors and sometimes public backstops. But each participant operates within limits. A reinsurer may have a maximum commitment, a bond may cover only losses within a specified band, and an insurer may have to retain part of every loss layer.

Without these restrictions, a familiar mathematical solution assigns each layer to whichever participant assesses it as least costly. Once capacity or shared transfer limits bind, that shortcut fails: assigning one layer can exhaust resources needed elsewhere.

In their preprint, Optimal robust comonotonic risk sharing under operational constraints, Carole Bernard, Christopher Blier-Wong and Jean-Gabriel Lauzier ask how to compute the best allocation under interacting restrictions—and how much those restrictions cost.

Turning contracts into a linear program

The authors restrict attention to comonotonic allocations. Each participant’s assigned loss is a nondecreasing function of the total loss, starts at zero and, together with everyone else’s allocation, accounts for the entire loss. A participant’s obligation therefore cannot fall simply because the aggregate loss rises.

Participants evaluate their exposure using distortion risk measures, which transform loss probabilities before calculating a risk assessment. These measures include Expected Shortfall and Value-at-Risk. The objective is to minimize the sum of participants’ assessments, not the physical damage caused by a catastrophe.

The method divides aggregate losses into an ordered grid. Its decision variables specify how much of each successive loss layer each participant holds. Under the monotonicity restriction, every participant’s risk assessment becomes a weighted sum of these layer amounts. Capacity caps, eligible coverage bands, retention requirements and other restrictions expressible as linear equalities or inequalities fit into the same framework.

The resulting problem is a linear program that standard optimization software can solve. The finite-grid result does not require all participants’ distortion functions to represent risk aversion. With capacity caps alone, it also takes the familiar form of a transportation problem: loss layers supply amounts that must be distributed among participants with limited room.

What a binding limit is worth

The optimization’s dual supplies supporting marginal prices for layers and shadow values for operational constraints. A capacity shadow value measures how much the minimum total risk assessment would fall with a small expansion, after re-optimizing the whole arrangement.

That qualification matters. Additional public coverage might replace a catastrophe bond in one band, allowing the bond to move higher and replace the insurer elsewhere. Comparing only the public backstop and insurer on the original layer would miss this rearrangement.

The authors also address points where the optimized value has a kink. There, the benefit of expanding a limit can differ from the cost of tightening it. The smallest optimal multiplier gives the marginal benefit of relaxation; the largest gives the marginal cost of tightening, where those changes remain feasible. An arbitrary multiplier returned by a solver need not describe both directions.

Supporting prices are not automatically market premiums. Under general operational restrictions, they need not even generate a normalized, nonnegative probability pricing measure. The allocation problem also leaves participant budgets and compensating transfers unresolved.

Testing disagreement between catastrophe models

The robust extension chooses one allocation that minimizes the worst total assessment across a finite collection of scenarios. Scenarios can change loss probabilities, participants’ distortion functions, or both. Adding a variable representing the worst assessment and a constraint for each scenario preserves the linear-programming structure.

Dual scenario weights identify which worst-case assessments contribute to supporting prices. These weights combine assessment coefficients; they are not estimated probabilities that particular catastrophe models are correct.

The illustration uses a hypothetical pool, two reinsurers, a catastrophe-bond tranche and a public backstop. Three constructed loss scenarios represent two vendor-style models and a climate-conditioned model. The parameters are illustrative, not calibrated to an actual insurance program.

A central result is that additional public capacity becomes less valuable as other opportunities for rearranging coverage disappear. In the nominal example, its marginal value drops sharply when the bond and backstop together fill their eligible band. Further capacity eventually becomes worthless because the backstop already covers its entire permitted band.

Model uncertainty changes this pattern. The robust marginal value of public capacity initially exceeds its nominal value, but later falls below it as a different scenario determines the worst case. This does not contradict the robust total value being at least each scenario’s separately optimized value: total levels and marginal changes are different quantities.

Where the conclusions stop

These are mathematical and illustrative results, not evidence that a particular public backstop should expand. The cost of providing public capacity must be supplied separately.

The model is single-period and uses only aggregate-loss information. Its optimum is not necessarily best among allocations outside the comonotonic class. Robustness covers the specified finite scenarios, not every possible model or their mixtures. The example also caps losses at 250 billion, making treatment of the extreme tail consequential.

For bounded losses and appropriately preserved constraints, the authors prove that finer grids approximate the continuous problem’s optimal value. That guarantee does not resolve the separate risk of truncating an unbounded catastrophe-loss distribution.