Cat XL pricing reduces, eventually, to a single question: is the tail of this portfolio worse than the tail of that portfolio? You cannot always answer it with a Monte Carlo. Sometimes you need to know whether portfolio A will always produce heavier maximum losses than portfolio B, regardless of the particular simulation draw. That is precisely the territory stochastic ordering inhabits — and a recent paper by Sangita Das (arXiv:2603.24640, March 2026) pushes that theory further into the territory relevant to reinsurance pricing.
We think this paper is more useful than it first appears, but you need to understand exactly what it does and does not deliver.
What the paper actually does
Das studies a specific problem: you have two heterogeneous portfolios, each containing an independent collection of risks. Portfolio 1 has N₁ claims with severities T₁, …, T_{N₁}. Portfolio 2 has N₂ claims with severities T₁, …, T{N₂}. Claim counts N₁ and N₂ are themselves random — Poisson, negative binomial, or more general distributions. The question is when can you say that max{T₁, …, T{N₁}} is stochastically larger than max{T₁, …, T_{N₂}}?
The answer, under Das’s theorems, depends on two things working in the same direction: the claim count distribution of portfolio 1 must dominate portfolio 2’s in convex order (roughly: more volatile claim counts in portfolio 1), and the severity distributions must be suitably ordered. The tool for establishing this joint ordering is multivariate chain majorization — a generalisation of the classical majorization that handles the structure of heterogeneous risk vectors.
The paper also addresses minima (min{T₁, …, T_{N₁}}), which matters for working layer excess of loss where frequency is high but you care about the minimum claim once you have an occurrence.
The key theoretical move is that you do not need to know the full joint distribution of (N, T). You need only know that certain marginal orderings hold. This is a non-trivial result: it says the tail of the aggregate maximum is determined by relatively coarse properties of the components.
Why this matters for cat XL
Cat XL pricing operates by estimating the expected loss to a layer (say, £50m xs £50m per occurrence) from a portfolio over an exposure period. The standard industry workflow is RMS or AIR event sets, loss amplification factors, and a lot of judgment calls about tail weight. The statistical question underneath all of this is: how confident are we that our modelled maximum loss distribution is heavier or lighter than a benchmark portfolio’s?
This is where stochastic ordering provides something Monte Carlo cannot easily give you: a partial order on portfolios that holds in all samples, not just on average. If portfolio A’s maximum loss stochastically dominates portfolio B’s, then for every quantile — not just the mean — A’s loss exceeds B’s. You do not need a simulation to confirm this; the ordering is a structural property.
In reinsurance treaty pricing this matters for two applications.
Portfolio comparison in renewal discussions. When a cedant presents a portfolio change — say, they have added a block of coastal residential property to a book that was previously dominated by inland commercial — you want to know whether the tail has gotten worse. Das’s framework says: look at whether the claim count distribution has become more volatile (convex order) and whether the severity distribution has shifted in hazard rate order or reversed hazard rate order. If both conditions hold, the maximum claim has unambiguously worsened. This gives you a principled basis for a rate-on-line adjustment that does not depend on curve-fitting assumptions.
Retrocession structuring. If you are accumulating across multiple cedants and want to assess which books contribute most to your aggregate maximum exposed loss, the ordering results let you rank books without running a full joint simulation of the panel. Again — structural reasoning rather than simulation-dependent point estimates.
What this means for pricing teams
The honest view is that this paper is primarily theoretical. It extends known stochastic ordering results to the case of random claim counts, which is the practically relevant case — fixed claim counts are a fiction. The multivariate chain majorization framework is elegant but not something you will implement from scratch in a pricing exercise.
The practical extraction is this: the paper gives you three conditions that are checkable from your data.
-
Convex order on claim counts. Does portfolio A’s claim count distribution have higher variance than portfolio B’s, at the same mean? This is the condition for portfolio A to have a “more volatile” frequency. Test it: compute E[N], Var[N] for both portfolios. Convex order is not the same as variance ordering (it requires all convex functions to be ordered, not just the second moment), but for distributions in the same family (both Poisson, both negative binomial), variance ordering typically implies convex order.
-
Hazard rate order or reversed hazard rate order on severities. For heavy-tailed claims in the Pareto or Burr XII family, hazard rate ordering reduces to comparing tail indices: the distribution with the smaller tail index (heavier tail) dominates in hazard rate order. If you have fitted severity models to both portfolios, comparing tail indices is a one-line calculation.
-
The ordering directions must be consistent. Both conditions must point in the same direction for the theorem to apply.
If these three conditions hold, you have a model-free result that portfolio A’s maximum claim distribution is heavier than portfolio B’s. You can quote this to underwriters and actuarial committees without reference to any simulation.
Connection to insurance-severity
When you need to verify condition 2 above — that one portfolio’s severity distribution dominates another’s in hazard rate order — you need to know both tail indices with some precision.
insurance-severity ships BladtTailScore, a strictly proper scoring rule for tail index estimation based on Bladt & Øhlenschlæger (arXiv:2603.24122). The key property: BladtTailScore evaluates candidate tail indices on normalised upper order statistics rather than the full sample, which is exactly the right domain for comparing Pareto-family tails. Score optimisation recovers the Hill estimator as a special case, but the scoring framework lets you compare any two Pareto-family models on the same data with confidence intervals.
The workflow for applying Das’s ordering conditions:
from insurance_severity import BladtTailScore, pareto_qq
import numpy as np
# Confirm Fréchet domain first
r2 = pareto_qq(claims_portfolio_a, k=200)
# R² > 0.95 required for BladtTailScore to be valid
scorer = BladtTailScore()
k_grid = np.arange(50, 300, 10)
# Compare two candidate tail indices
ranking = scorer.rank(
y=claims_portfolio_a,
gamma_candidates=[0.40, 0.50, 0.60], # heavier tail = higher gamma
k_grid=k_grid,
)
# gamma_hat_A from ranking; repeat for portfolio B
# If gamma_hat_A > gamma_hat_B: portfolio A has heavier tail
# Combined with claim count variance comparison → stochastic order applies
Once you have gamma estimates for both portfolios with confidence intervals, the stochastic ordering argument from Das provides the theoretical backing for saying the ordering is structural, not a simulation artifact.
Where this paper falls short
We want to be direct about the limitations, because the paper does not emphasise them.
The ordering results are sufficient conditions, not necessary conditions. Portfolios can have ordered maxima without satisfying the majorization conditions. If your portfolios fail the conditions, you cannot conclude they are unordered — only that this particular tool does not apply.
The claim count ordering (convex order) is a strong condition. For real portfolios, claim counts are typically modelled with Negative Binomial, and the convex order between two NBs depends on both the mean and variance in a non-trivial way. You cannot just compare variance.
The paper treats claim counts and severities as independent. In reality, frequency and severity are often positively correlated — high-exposure risks tend to produce both more claims and larger ones. When that correlation is present, the theorem structure requires modification.
Finally, the paper works in the Fréchet domain throughout. For books where large losses follow a lognormal (household contents, perhaps some motor books), the results do not apply. Check your domain before invoking the theory.
The verdict
This is solid mathematical actuarial science that gives pricing teams a tool they do not currently have: a model-free way to rank portfolios by tail risk based on structural properties rather than point estimates. For cat XL pricers comparing books in treaty negotiations, the three conditions are worth checking as a first-pass screening before you build the full event loss table. When the conditions hold, you have a strong result. When they do not hold, you fall back to simulation — but now you know you are in the simulation-dependent regime, which is itself useful information.
The paper is submitted to Ricerche di Matematica, not a practitioner journal, and shows it. Extracting the usable core requires work. We have done that work above.