Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Risk Ordering and Stochastic Dominance

Abstract. Stochastic orders compare uncertain outcomes through inequalities that remain valid over specified classes of preferences, threshold losses, or aggregation operations. Their importance for insurance lies in the distinction between a loss distribution that is uniformly smaller, one that creates a smaller expected shortfall above every deductible, and one that is less dispersed around the same mean. These are different mathematical claims with different consequences for underwriting, reinsurance, reserving, capital assessment, and insured welfare. This monograph reconstructs first-order stochastic dominance, second-order dominance for wealth, increasing convex order for losses, convex order, integrated survival functions, and quantile-integral representations. It presents expectation and coupling characterizations, proves key relationships from elementary integral identities and Jensen's inequality, and identifies precise requirements for each comparison. Original finite-support calculations show that variance rankings need not establish convex dominance, that a value-at-risk ranking can reverse across confidence levels even when convex order is unambiguous, and that insurer aggregation depends on dependence structure rather than marginal loss distributions alone. Further analysis extends these comparisons to stop-loss layers, comonotonic bounds, compound and common-shock portfolios, and empirical inference. The study distinguishes established mathematical theorems from illustrative computations and from institutional uses requiring assumptions about incentives, contracts, estimation and law. Its synthesis establishes how distributional ordering provides partial yet actionable comparisons without replacing a decision criterion, a validated probabilistic model, or a complete account of financial risk.

Keywords: Stochastic order; first-order stochastic dominance; second-order stochastic dominance; increasing convex order; convex order; mean-preserving spread; stop-loss premium; integrated survival function; martingale coupling; actuarial risk; comonotonicity; dependent aggregation; expected shortfall; quantile function; empirical dominance; insurance pooling; reinsurance; uncertainty.

Introduction

Insurance risk is encountered as a set of possible payments and the probability law governing them. A comparison of expected claims alone permits some decisions but leaves the distribution of large claims unresolved. An assessment of variance adds information about squared deviations but can conceal the direction of risk in different parts of a distribution. An insurer deciding whether to retain a deductible layer, a reinsurer comparing portfolios, and a household comparing residual wealth after a loss can therefore arrive at distinct answers from the same pair of distributions. Stochastic ordering supplies a rigorous language for asking what a comparison proves and for which class of consequences it is valid.

The early economic foundations include the ordering rules of Hadar and Russell (1969), the efficiency analysis of Hanoch and Levy (1969), and Rothschild and Stiglitz's definition of increasing risk (1970). These contributions develop distinct but connected routes from cumulative distribution inequalities to preference comparisons and changes in riskiness. Subsequent mathematical treatments, particularly the reference work Stochastic Orders by Shaked and Shanthikumar (2007), establish a much wider taxonomy of orders and their closure properties. The present work concentrates on those relations necessary for comparing insurance losses and consequences under explicit assumptions.

Throughout, X and Y denote integrable real-valued random variables on possibly different probability spaces. When used as losses, smaller numerical values are preferred unless a different decision objective is specified. A negative observation may represent a gain or recovery; the theory of orders admits such outcomes even though conventional gross insurance claims are nonnegative. A wealth variable is W = w - L when initial wealth w is fixed and L is the financial loss. The orientation of all order symbols will be given in terms of the loss variable. In particular, X preceding Y in an increasing-convex loss order means that Y is at least as costly under every increasing convex loss evaluator for which expectations exist. A statement about the utility of W reverses that orientation.

This is a comparative theoretical investigation. Primary scholarly literature supplies the established foundations; finite probability tables are independent mathematical calculations made for this monograph. Insurance implications are derived with conditions identified, and findings about published standards are used only where an actual primary source supports them. Neither an ordering theorem nor a toy calculation is presented as a validated capital requirement or a policy-specific premium.

Distribution Functions and Risk Orders

Defining the comparison object

The cumulative distribution function of X is F_X(t) = P(X <= t); its survival function is S_X(t) = P(X > t). The left-continuous quantile function is Q_X(p) = inf{t: F_X(t) >= p} for 0 < p < 1. These definitions make no assumption of continuity, so they apply to discrete insurance payments, mixtures with a point mass at zero, and bounded contractual payoffs. For an integrable variable, the stop-loss transform at retention d is

ΠX(d)=E⁡[(X−d)+]=∫d∞SX(t) dt,\Pi_X(d)=\operatorname E[(X-d)_+]=\int_d^\infty S_X(t)\,dt,

where a_+ = max(a,0). The identity follows pointwise from (x-d)_+ = integral from d to infinity of 1{x > t} dt, followed by Tonelli's theorem. It remains valid with atoms at d, and the strict inequality in the survival function does not affect the Lebesgue integral. An actuary can interpret this functional as the pure expected payment of an unlimited excess-of-loss layer attaching at d; contractual expense, discounting, credit risk, limits and premium loadings require separate treatment.

A stochastic order is a relation on probability laws rather than on the numerical labels of the same observations. It asks whether a family of inequalities holds uniformly. Write X <=st Y for usual or first-order stochastic order, X <=icx Y for increasing convex order, and X <=cx Y for convex order. These relations are partial orders on distributions modulo equality in law: two different laws can be incomparable, and a shared expectation or a lower variance does not force them into a complete ranking. Detailed mathematical reference treatments are given by Müller and Stoyan (2002) and Denuit et al. (2005).

Relationship of risk measures to preference classes

A risk statistic T(X) gives one numerical value, such as E[X], Var(X), a percentile or an expected excess. A stochastic order instead makes an infinite family of comparisons. One scalar functional can witness that an alleged order fails, while agreement at a single scalar value seldom proves an order. The distinction is crucial where a formal decision rule attaches consequences to large losses at multiple deductibles or to the entire distribution of retained wealth.

In this monograph every mathematical implication is conditional on the relevant integrability and on the class of test functions. Claims about all increasing functions may be evaluated first for bounded functions; claims about unbounded functions additionally require the expectations to be well defined. Convex-order characterizations use finite first moments. Variance comparisons require finite second moments. Examples are discrete and so satisfy all moment conditions.

A compact map of the principal orders

Loss ordering Distributional characterization Test-function interpretation Typical insurance reading
X <=st Y F_X(t) >= F_Y(t) for every t E[g(X)] <= E[g(Y)] for all nondecreasing g All threshold exceedance probabilities are no larger for X
X <=icx Y E[(X-d)+] <= E[(Y-d)+] for every real d E[g(X)] <= E[g(Y)] for increasing convex g Every pure unlimited excess-of-loss expectation is no larger for X
X <=cx Y X <=icx Y and E[X] = E[Y] E[g(X)] <= E[g(Y)] for convex g Same mean with less dispersion in the convex-order sense

The table is a summary of established equivalences, not a definition of insurance welfare. The wealth-side counterpart of increasing convex loss order is dominance under increasing concave utility. Convex order additionally removes differences in mean and isolates comparative dispersion; it still cannot specify whether two decision makers have the same initial wealth, borrowing constraints, premiums or contract terms.

First-Order Dominance

The threshold theorem

The first-order stochastic order X <=st Y holds if and only if F_X(t) >= F_Y(t) for every t. Equivalently, P(X > t) <= P(Y > t) at each threshold. This is a direct statement about the entire distribution of losses. For every bounded nondecreasing function g, E[g(X)] <= E[g(Y)]; the extension to integrable transformations uses the usual expectation conditions. Conversely, choosing g(x) = 1{x > t} for arbitrary t recovers the survival inequalities. This establishes the functional and distributional equivalence without making differentiability assumptions.

A shift by a positive constant is a transparent example. Let X be 0 or 3, each with probability 1/2, and let Y be 1 or 4, each with probability 1/2. Then Y has the same law as X+1. At every threshold X has at least as much probability accumulated to the left as Y, and X <=st Y. The means are 1.5 and 2.5 respectively; the variances are equal, both 2.25. A criterion based only on variance declares a tie and misses the strict loss deterioration captured by first-order dominance.

Quantile and coupling representations

First-order dominance is also equivalent to Q_X(p) <= Q_Y(p) for almost every p in (0,1). Let U be uniform on (0,1) and construct X* = Q_X(U), Y* = Q_Y(U). The two constructed random variables have the required marginal laws, and X* <= Y* almost surely. Thus a monotone coupling exists even if the actual insured exposures are not jointly observed. This mathematical construction gives a powerful sufficiency proof: increasing g preserves the pointwise inequality. It does not claim that the real-world losses X and Y are observed in the same states or under the constructed dependence.

The general theory of existence of couplings with prescribed marginals is connected to Strassen's marginal-distribution theorem (1965). The distinction between a possible coupling and actual dependence will become central in comparing aggregate insurer losses.

Transformations and comparison robustness

If X <=st Y and h is a nondecreasing measurable function, h(X) <=st h(Y). If Z is independent of X and Y in the appropriate comparative constructions, X+Z <=st Y+Z: condition on Z and translate the threshold. These closure properties allow transparent analysis of parallel benefits, expense adjustments, or contractual functions that preserve monotonicity. For instance, applying a capped nondecreasing claims-payment function min((L-d)_+, m) preserves a first-order ordering when the same contract is applied to both loss laws.

Conversely, a change in benefit design can reverse the relevant target variable. Net losses after a deductible, the insurer's indemnity, and the policyholder's total payment are distinct monotone transformations and may need separate comparisons. A ranking of gross claims does not identify who bears which layer or whether contract prices differ. The functional definition tells us exactly where those institutional questions enter.

Second-Order Dominance and Expected Utility

Wealth and the increasing-concave order

For wealth outcomes A and B, say that A second-order stochastically dominates B in the increasing-concave sense when E[u(A)] >= E[u(B)] for every increasing concave utility function u for which the expectations exist. Define an increasing-concave order A >=icv B for clarity. This order is commonly expressed using integrals of cumulative distribution functions: for appropriate integrability,

A⪰icvB⟺∫−∞tFA(s) ds≤∫−∞tFB(s) dsfor every t.A\succeq_{\mathrm{icv}} B \quad\Longleftrightarrow\quad \int_{-\infty}^{t}F_A(s)\,ds\leq\int_{-\infty}^{t}F_B(s)\,ds\quad\text{for every }t.

The functional class is broader than a single subjective utility curve. It establishes that all nonsatiated risk-averse expected-utility decision makers rank A no worse than B, subject to common nonfinancial consequences and the wealth representation itself. Classical formal treatments are the Hanoch–Levy choice analysis and the Hadar–Russell dominance rules.

If A = w-X and B = w-Y for identical starting wealth w, this condition is equivalent to X <=icx Y. Indeed, for any increasing concave u, the function g(l) = -u(w-l) is increasing and convex; hence E[g(X)] <= E[g(Y)] means E[u(w-X)] >= E[u(w-Y)]. This sign conversion is fundamental. A loss distribution more favorable in increasing convex order generates wealth more favorable for all increasing concave utilities. Second-order dominance stated for wealth must not be copied into loss tables without reversing the preference meaning.

The integrated-distribution proof

For an integrable wealth variable A, the integrated distribution functional at threshold t has a simple interpretation:

∫−∞tFA(s) ds=E⁡[(t−A)+].\int_{-\infty}^{t}F_A(s)\,ds = \operatorname E[(t-A)_+].

The test function u_t(a) = -(t-a)_+ is increasing and concave. If A dominates B for all increasing concave utilities, evaluating this family yields the integrated-CDF inequalities. Conversely, increasing concave functions can be approximated or represented on bounded intervals by affine components and nonnegative mixtures of these kinked functions; passage to the limit under the appropriate integrability assumptions yields the full expected-utility ranking. The result is a genuine equivalence rather than a visual heuristic about cumulative curves.

Mean shifts and pure risk changes

Increasing-concave wealth dominance may arise from more wealth, less downside dispersion, or both. A 100-unit sure gain over an otherwise identical prospect is a first-order improvement and consequently an increasing-concave improvement. When E[A]=E[B], the comparison isolates a mean-preserving reduction in risk: A is preferred by every concave utility, increasing or otherwise. In loss orientation, a constant loss 10 is less risky in convex order than a 50/50 lottery paying 0 or 20, even though both have expected loss 10. For the wealth variables w-10 and w-L, Jensen's inequality gives u(w-10) >= E[u(w-L)] for every concave u, with strictness under the usual nondegeneracy and strict-concavity conditions.

The economic meaning of curvature must be expressed at the appropriate wealth level. Pratt's measure of local risk aversion (1964) uses -u''(w)/u'(w) for twice-differentiable utility. Its role is to characterize local preference sensitivity; its numerical value does not replace the global distributional comparison, which holds across the entire class of increasing concave functions. Rothschild and Stiglitz's economic consequences of increasing risk (1971) similarly illustrate why a riskier distribution and the behavioral response to it are separate propositions: saving, insurance demand and other choices depend on economic structure as well as on preferences.

Higher-order preferences outside the central comparison

Increasing-concave utility and convex ordering are the principal second-order objects of this investigation. More restricted utility classes, including conditions on successive derivatives, lead to higher-order dominance relations. Bawa (1975) develops optimal rules for ordering uncertain prospects with particular attention to decreasing absolute risk aversion and third-order dominance. Those results are recognized as a separate research line: third-order preference restrictions should not be represented as automatically following from second-order inequalities or from a single moment ranking.

What second-order dominance does not establish

An expected-utility ranking requires the specified utility class and the correct monetary payoff. Nonlinear taxes, credit limits, health consequences, ambiguous probabilities or multiple nonmonetary objectives can change the decision problem. Observing that a distribution has a lower variance does not prove the integrated-CDF inequalities. For an individual insurer, even the expected-utility comparison of terminal surplus can depend on premiums, available liquidity, investment cash flows and who is legally obligated to pay claims. A distributional theorem should therefore be used as a testable component of the decision, not substituted for a complete model of that decision.

Convex Ordering and Mean-Preserving Spreads

Convex expectation as the definition of dispersion

For integrable X and Y, X <=cx Y means E[phi(X)] <= E[phi(Y)] for all convex phi for which expectations exist. Because both x and -x are convex affine functions, this definition implies E[X] = E[Y]. The equality of means is an indispensable part of the order, not an optional descriptive statistic. Any unrestricted statement that one unequal-mean loss law is a mean-preserving spread of another is internally inconsistent.

A useful equivalent characterization is

X≤cxY⟺E⁡[X]=E⁡[Y]  and  ΠX(d)≤ΠY(d)for every real d.X\leq_{\mathrm{cx}}Y \quad\Longleftrightarrow\quad \operatorname E[X]=\operatorname E[Y] \;\text{and}\; \Pi_X(d)\leq\Pi_Y(d)\quad\text{for every real }d.

The forward implication follows because (x-d)_+ is convex. The converse uses the representation of convex functions as an affine term plus a nonnegative mixture of stop-loss hinges. The affine terms cancel under equal means. Thus a single family of excess-loss comparisons establishes every convex expectation inequality, and a violation at any deductible disproves convex order. This result is foundational to the actuarial treatments in Actuarial Theory for Dependent Risks and Stochastic Orders.

The increasing convex order drops the equal-mean requirement. X <=icx Y if and only if E[(X-d)+] <= E[(Y-d)+] for every d. Taking d to negative infinity and using integrability shows E[X] <= E[Y]. Such a comparison therefore allows the second loss to have both a larger mean and a more adverse excess-loss profile. In mathematical insurance literature the phrase stop-loss order commonly denotes this increasing convex relation, while convex order denotes its equal-mean specialization. This identification is documented directly in the positive-dependence and stop-loss study.

Mean-preserving noise and martingale transport

Let X be integrable and Z another integrable random variable satisfying E[Z|X] = 0. Put Y = X+Z. Then E[Y]=E[X], and conditional Jensen's inequality gives

E⁡[ϕ(Y)∣X]=E⁡[ϕ(X+Z)∣X]≥ϕ(X+E⁡[Z∣X])=ϕ(X).\operatorname E[\phi(Y)\mid X] =\operatorname E[\phi(X+Z)\mid X] \geq\phi\bigl(X+\operatorname E[Z\mid X]\bigr) =\phi(X).

Taking expectations proves X <=cx Y. This is the precise probabilistic content of adding a conditionally mean-zero spread. Ordinary zero covariance, E[XZ] = 0, alone does not imply E[Z|X]=0, and should not be substituted for the stronger condition. Independence with E[Z]=0 is sufficient but not necessary. In the converse direction, under standard real-valued integrability conditions a convex-order relation admits a joint construction of variables X*,Y* with the prescribed marginal laws satisfying E[Y*|X*] = X*. This is the martingale-coupling perspective associated with Strassen (1965). The coupling need not reproduce the real-world joint distribution of measured losses.

The classical mean-preserving-spread treatment of Rothschild and Stiglitz (1970) is central to this equivalence. The same argument explains why merely adding an uncertain but independently priced contract to a portfolio can increase convex dispersion even when its expected net value is zero. Whether the party values that dispersion negatively depends on the utility of wealth and on the contract's allocation of gains and losses.

A variance comparison that fails

A necessary consequence of X <=cx Y, when second moments exist, is Var(X) <= Var(Y): apply the convex function x^2 and use equal means. The converse fails. Consider the finite loss laws

Loss outcome Law A probability Law B probability
-1 0 1/4
0 1/2 0
2 0 1/2
4 1/2 0
5 0 1/4

Both laws have mean 2. Law A has variance 4; law B has variance 4.5. A dispersion comparison based only on variance ranks A below B. Yet the stop-loss transforms cross. At deductible d=0, Pi_A(0)=2 and Pi_B(0)=2.25. At deductible d=2, Pi_A(2)=1 and Pi_B(2)=0.75. Therefore A <=cx B is false because the second inequality fails, while B <=cx A is false because the first fails. The laws are convex-order incomparable. This example shows exactly why mean-variance comparisons cannot generally be promoted to all-convex-function rankings. Restrictions to particular distribution families, such as appropriately ordered common location-scale families, can produce stronger implications, but those restrictions must be stated and verified.

Quantile-integral criteria

For integrable variables the upper integrated quantile function is K_X(p)=integral from p to 1 of Q_X(u)du. The following equivalences are especially useful in numerical actuarial work:

X≤icxY⟺∫p1QX(u) du≤∫p1QY(u) dufor every 0<p<1,X\leq_{\mathrm{icx}}Y\quad\Longleftrightarrow\quad \int_p^1Q_X(u)\,du\leq\int_p^1Q_Y(u)\,du \quad\text{for every }0<p<1,

where the endpoint p=0 recovers the comparison of means; and X <=cx Y precisely when these upper-integral inequalities hold together with equality of means. The equivalent lower-tail representation uses the reverse inequality for integrals from zero to p under equal means. The statements connect stop-loss contracts with ranked outcomes but should not be confused with pointwise quantile dominance, which is the stronger first-order order. A law can have a larger quantile in some probability ranges and still be convex-order smaller overall.

The proof follows from the dual representation Pi_X(d)=sup over p in [0,1] of {integral_p^1 Q_X(u)du -(1-p)d}, with suitable treatment of tied quantiles, and the corresponding inverse optimization. Both sides are convex transforms of the same distributional information. The practical implication is that checking an entire upper quantile-integral curve is stronger than comparing one percentile or one expected shortfall level.

Stop-Loss Comparisons and Insurance Contracts

Pure layer payments

A stop-loss or excess-of-loss indemnity at retention d and unlimited cover is I_d(L)=(L-d)_+. Its pure expected payment is Pi_L(d). If X <=icx Y, every such expected payment for X is no greater than for Y. The statement is universal in retention and rests solely on the loss laws. It does not require an insurer to charge the pure premium, nor does it determine the premium loading or contract availability. With nonnegative L, the special case d=0 gives E[L]; with contractual limits the expected layer payment becomes

E⁡[min⁡((L−d)+,m)]=ΠL(d)−ΠL(d+m),m≥0.\operatorname E[\min((L-d)_+,m)] =\Pi_L(d)-\Pi_L(d+m),\qquad m\geq0.

This difference of two stop-loss transforms need not preserve their separate ordering. Even if Pi_X(t) <= Pi_Y(t) for every t, subtracting two such inequalities does not prove that a limited layer on X has a smaller expectation for every d and m. Indeed, two convex-ordered laws can have crossing survival functions and changing local layer costs. This is a consequential distinction between an unlimited stop-loss comparison and a bounded layer comparison.

Consider X=10 surely and Y equal to 0 or 20 with probabilities 1/2 each. Then X <=cx Y. For an unlimited layer attaching at 10, Pi_X(10)=0 and Pi_Y(10)=5. For a limited layer attaching at 0 with width 5, the expected payments are E[min(X,5)]=5 and E[min(Y,5)]=2.5. The direction has reversed because the capped payment is an increasing concave function of loss rather than a convex one. This construction prevents the common error of claiming that convex order ranks every capped insurance layer or every claims transformation.

Stop-loss profiles as a research tool

The function d -> Pi_X(d) is nonnegative, decreasing and convex. Its right and left slopes, where applicable, encode survival probabilities; at continuity points Pi'_X(d)=-P(X>d). The change in slope across a mass point recovers its probability mass. Consequently the entire stop-loss profile contains the distribution, while an isolated deductible quote reveals only one functional. For distributions with finite means, two identical stop-loss profiles imply identical laws. This makes a full retention curve a natural diagnostic object for underwriting comparison, exposure modeling and reinsurance layer design.

To compare two portfolios, first determine whether their displayed stop-loss profiles cross. If they do, neither increasing convex order relation holds. If the curves do not cross, verify which one lies below the other and whether equality of means is also present. The observed premiums must be adjusted for contract wording, default exposure, expenses, discounting, commissions, reinstatement provisions, collateral and margin before being interpreted as estimates of the pure functions.

Deductibles and wealth

An insured retaining min(L,d) under a simple deductible contract pays an increasing concave function of gross loss, whereas the insurer paying (L-d)_+ bears an increasing convex function. Thus the same convex-order comparison has a clear result for the unlimited insurer layer but may provide no uniform comparison for the policyholder's retained capped loss. The total insured cost also includes the premium; if premium schedules differ, a separate wealth comparison is required. Contract selection, moral hazard and correlated claims further alter the law to be compared.

The broader actuarial literature on contract structure has developed with a direct focus on stop-loss insurance; the 2025 fifth edition of Föllmer and Schied's Stochastic Finance explicitly includes a new treatment of stop-loss insurance contracts. That bibliographic fact is not used to import any unexamined theorem from the book. The formal results in this chapter follow from the definitions, elementary expectations and fully specified loss examples.

Quantiles, Tail Integrals, and Risk Measures

VaR and expected shortfall answer different questions

For a loss L, define VaR_alpha(L)=Q_L(alpha) for 0<alpha<1. Define the coherent integrated-tail version of expected shortfall by

ES⁡α(L)=11−α∫α1QL(u) du.\operatorname{ES}_{\alpha}(L) =\frac{1}{1-\alpha}\int_{\alpha}^{1}Q_L(u)\,du.

This quantile-integral definition handles point masses without requiring an ad hoc conditioning convention at the cutoff. Acerbi and Tasche's analysis of expected shortfall (2002) emphasizes exactly this issue: alternative conditional-tail definitions that agree for continuous distributions can diverge at atoms. That issue is especially material for insurance, where zero claims and benefit limits generate discrete masses. The present monograph therefore specifies the integral convention whenever ES is computed.

If X <=icx Y, the integrated quantile criterion yields ES_alpha(X) <= ES_alpha(Y) for every alpha in (0,1). The converse holds if all these inequalities hold and integrability is maintained; taking alpha toward zero also recovers E[X] <= E[Y]. For convex order, equal means must additionally be verified. VaR at one alpha is not generally ordered by convex order, and its ranking can reverse as alpha changes. The original coherent-risk-measure discussion of Artzner et al. (1999) and the detailed Acerbi–Tasche treatment provide the appropriate distinction between quantile and integrated-tail measures.

A single comparison with reversed VaR outcomes

Let X be a sure loss of 10. Let Y be 0 with probability 0.95 and 200 with probability 0.05. Both expected losses are 10. Jensen's inequality gives X <=cx Y. At confidence alpha=0.90, VaR_alpha(X)=10 while VaR_alpha(Y)=0. At alpha=0.99, VaR_alpha(X)=10 while VaR_alpha(Y)=200. The value-at-risk ranking reverses across thresholds while the convex-order relation remains unchanged. By contrast, ES_0.90(X)=10 and ES_0.90(Y)=100; ES_0.99(X)=10 and ES_0.99(Y)=200. The ES inequalities remain consistent with increasing convex order.

All outcomes are illustrative monetary units and not statistical estimates. The example gives a decisive logical counterexample to any theorem claiming that a mean-preserving spread must increase every fixed loss quantile. It also shows why a capital measure tied to a particular percentile cannot be identified with a preference order across all convex functions, even if the measure is operationally useful for a specified supervisory or business objective.

Link to risk-measure theory

Risk-measure axioms and stochastic orders meet at the expectation functional. A law-invariant functional monotone with respect to increasing convex order will rank X no larger than Y whenever X <=icx Y; a functional lacking this property may reverse that ranking. Expected shortfall satisfies the order comparison just established. For a single VaR, no such general conclusion follows. Coherence, cash additivity, positive homogeneity, subadditivity and monotonicity are distinct functional axioms; none should be silently substituted for a stochastic-order theorem. The conceptual separation is important: a regulatory capital rule can be well specified while remaining sensitive to one tail region in a way that a whole-distribution preference comparison is not.

Extreme-tail completeness and its limits

The family {ES_alpha: 0<alpha<1} encodes the upper integrated quantile curve. This makes it a complete characterization of increasing convex order, not evidence that estimation across all confidence levels is feasible from a finite sample. High quantiles require extrapolation, scenario models or extreme-value assumptions when data are sparse. Observed agreement in a finite grid of confidence levels is a provisional numerical result; if two transforms cross between grid points, the population order may fail. Convexity, numerical error control, distributional support, and uncertainty in parameters determine how much can be inferred from computations.

Insurance Pool Applications and Dependence

Why marginal distributions do not determine the aggregate

A portfolio loss is S=X_1+...+X_n. Knowledge of each marginal law F_i does not determine the law of S unless a dependence structure is specified. The mean E[S]=sum E[X_i] is fixed under all joint constructions with these marginals, while variance, tail quantiles, stop-loss expectations and the likelihood of simultaneous extreme claims can change substantially. Dependence enters the aggregate through the joint law, not solely through the pairwise correlation coefficients. In the two-dimensional case copulas represent dependence separately from marginal distributions; Nelsen's Introduction to Copulas (2006) develops this mathematical architecture. The distinction is essential to catastrophe accumulation, common policyholder exposure, correlated inflation and common financial shocks.

At fixed marginal laws, the comonotonic sum S^c=sum Q_i(U), with one common uniform U, is maximal in convex order among possible sums under the standard integrability assumptions. This central actuarial result is proved by Kaas et al. (2002) and developed across the comonotonicity theory paper and its applications companion by Dhaene and collaborators. The maximality concerns convex expectations of the sum at fixed marginals. It does not mean that every percentile is highest under comonotonicity, nor does it identify a universal minimum among other multivariate dependences. Additional dependence constraints are needed for stronger bounds.

A simple proof of the upper bound for discrete marginals clarifies the mechanism. Given two marginal loss vectors, pairing their ordered values in the same rank maximizes the expectation of any convex function of their sum: replacing two crossed pairs (a,d) and (b,c), with a<=b and c<=d, by aligned pairs (a,c) and (b,d) preserves their combined mean while placing the new two sums farther apart, and hence weakly raises the convex expectation. Repeating swaps across finite samples produces comonotonic ranks. Approximation extends the intuition to integrable general distributions. A full mathematical treatment handles measure-theoretic details and multiple variables, as in Kaas et al. (2002).

Two insured exposures: exact aggregation table

Consider two exposures with fixed losses of 100 each when a claim occurs, each claim probability 0.10. Under independence the aggregate loss S^i has probabilities 0.81 at 0, 0.18 at 100 and 0.01 at 200. Under perfect comonotonicity the aggregate S^c is 0 with probability 0.90 and 200 with probability 0.10. Both aggregates have expected loss 20, but the variances are 1,800 and 3,600 respectively. At retention 100, Pi_i(100)=1 while Pi_c(100)=10. At confidence 0.95, VaR(S^i)=100 and VaR(S^c)=200; at confidence 0.85, VaR(S^i)=100 while VaR(S^c)=0. This latter reversal is consistent with convex-order maximality and shows again why isolated VaR comparisons do not characterize aggregation risk.

Aggregate outcome Independent probability Comonotonic probability
0 0.81 0.90
100 0.18 0
200 0.01 0.10

The 95% expected shortfall is 120 under independence: the worst 5% consists of 1% at 200 plus 4% at 100, so (0.01×200+0.04×100)/0.05=120. It is 200 under comonotonicity, where the worst 5% consists solely of 200 outcomes. Every number follows directly from the table. This is a controlled mathematical comparison, not a calibrated catastrophe model, and the absence of common shock in the independent construction is a substantive assumption.

Positive dependence between independence and comonotonicity

Some dependence hypotheses allow a lower comparison against an independent version of the portfolio. Denuit, Dhaene and Ribas (2001) show for their particular positive cumulative dependence class that an independent aggregate is below the dependent aggregate in stop-loss order, and the comonotonic aggregate is an upper bound. The precise assumptions and statement are available in their positive-dependence stop-loss analysis. It would be incorrect to generalize the lower bound to every joint distribution sharing the same marginals. For some negative dependence structures aggregate claims are materially more stable than under independence; pairing two mutually exclusive fixed Bernoulli payments, for example, eliminates simultaneous claims and may produce an almost constant total.

The dependence assumptions therefore determine which side of the independent reference is attainable. A common-shock model can create higher upper-tail expectations without altering the individual claim marginal law. A quota-share reinsurance treaty transforms each loss linearly and preserves convex-order comparisons under scaling, but a mixture of proportional, attachment-point and limited contracts requires attention to the actual net-loss function. A combined portfolio's order cannot be inferred by simply adding separately estimated orders unless the necessary joint dependence or independence conditions hold.

Convolution, pooling, and diversification

If X <=cx Y and an integrable Z is independent of both under the comparison constructions, then X+Z <=cx Y+Z. The proof conditions on Z: for every convex phi, x -> phi(x+z) is convex, so the inequality holds for each z and can be integrated. The analogous result holds for increasing convex order. Repeated convolution thus preserves order when the independent additions are actually justified. If the added exposure is systematically dependent on one compared portfolio but not the other, the conditioning proof fails; distinct cross-covariances may reverse the conclusion.

Consider n independent, identically distributed nonnegative losses with finite mean mu and variance sigma squared. The aggregate variance grows as n sigma squared and the per-policy average variance falls as sigma squared/n. That frequency-stabilization result is different from asserting that the aggregate of n policies is convex-order below the aggregate of n-1 policies: the means differ and convex order then cannot even apply. A fair comparison needs a fixed monetary scale, such as averages, or a standard exposure size. For each n, S_n/n is convex-order no greater than an individual loss X_1 when observations are identically distributed, because Jensen's inequality yields phi((sum X_i)/n) <= (sum phi(X_i))/n pointwise. Independence is not required for this particular fixed-n convex-order conclusion, though dependence determines how much reduction occurs. This observation clarifies the exact mathematical content of pooling without confusing total and per-exposure losses.

Marginal uncertainty and aggregation bounds

Even when the complete marginal laws are known, the distribution of their sum remains partially identified if dependence is unknown. The classical Makarov bounds (1982) study distribution-function bounds for a sum with specified marginals. Such bounds concern feasible aggregate distribution functions and must be distinguished from convex-order upper bounds and from full joint-model inference. For more than two risks, attainable dependence structures and sharpness at different thresholds become more complicated. This motivates presenting quantitative results as conditional intervals under a stated dependence class when the joint distribution is not credibly established.

A reliance on a single estimated correlation coefficient cannot replace this work: joint tail behavior is not determined by a linear correlation coefficient. Two distributions can have similar second moments and very different rare-event clustering. The monograph's correct claim is that a chosen stochastic order follows only after the required marginal and dependence restrictions have been verified; the act of pooling does not confer a universal favorable ranking on the underlying aggregate distributions.

Reinsurance, portfolio selection, and prices

For unlimited excess-of-loss covers, X <=icx Y implies the pure stop-loss premium for X is no larger at every retention. This provides a direct comparison before market pricing elements. For proportional reinsurance, positive scaling a>=0 preserves first-order, increasing convex and convex order; deterministic translations preserve all three, with mean equality retained for convex order. For aggregate risk transfer with nonlinear limits or reinstatements, a claim amount is a piecewise-defined function that should be tested anew rather than presumed to preserve convex order.

Risk choice also depends on what the insurer optimizes. A reinsurance buyer minimizing a loss-based expected-shortfall measure will rank increasing-convex-ordered losses in the expected direction, provided other cash flows are held fixed. A capital rule based on a selected loss quantile may rank them differently. A contract priced by expected indemnity plus loading may change the comparison of post-premium wealth. It follows that order theory can narrow admissible alternatives while leaving a remaining choice set that must be resolved by utility, budget, liquidity, market prices and applicable regulation.

Statistical Identification and Empirical Testing

Population order and finite-sample uncertainty

All previous theorems are statements about population distributions. Insurance researchers typically observe a finite sample of policy losses, claims counts, annual aggregates or simulated events. Replacing F_X and F_Y with empirical cumulative distribution functions produces an estimated difference, not a proven population order. At a given retention d, a sample stop-loss estimate is the arithmetic mean of (X_i-d)_+; its sampling uncertainty may be substantial when severe claims are rare. Checking several selected retentions is informative but incomplete, because stochastic dominance requires an inequality across the whole relevant domain.

Formal empirical work distinguishes a descriptive sample ordering from a statistical test of a composite population hypothesis. Davidson and Duclos (2000) develop inference for stochastic dominance in the context of welfare, poverty and inequality measurement. Barrett and Donald (2003) provide consistent stochastic-dominance tests, with simulation and bootstrap approaches for higher-order comparisons. The statistical lessons extend to insurance distributions, but their validity relies on each study's sampling conditions and test construction. A bootstrap designed for independent observations should not automatically be applied to spatially dependent catastrophe claims or serially dependent reserving data.

Exact finite-support computation

If two loss distributions have known finite supports, first-order order can be checked by evaluating both CDFs at every distinct support point. Increasing convex order can be checked by evaluating the stop-loss differences at every distinct support point, together with the appropriate far-left asymptote; between adjacent support points those differences are linear. Equal means then establish convex order if and only if increasing convex order has been verified. For exact rational probabilities, integer arithmetic can avoid floating-point ambiguity; for computed probabilities use justified numerical tolerance and account for probability normalization.

This finite-support property is what makes the counterexamples in this monograph logically decisive. In a continuous or heavy-tailed model the retention domain is infinite and numerical checks need analytical reduction, error bounds, or sufficiently controlled approximation. For two sample distributions, an observed crossing at some d is evidence of sample incomparability but may be unstable under sampling variation. A credible empirical conclusion includes confidence regions or simultaneous testing procedures rather than treating the observed curve as a population certainty.

Testing without assuming away the tail

A practical insurance testing protocol begins by specifying the exposure unit, currency, inflation adjustment, observation window, gross or net loss basis, censoring and deductibles, and claim-development adjustment. It then separates frequency, severity, and common-shock sources, fits or estimates the marginal laws, and audits which dependences the data can identify. Distributional comparisons should be supported by diagnostic plots of F_X-F_Y and Pi_X-Pi_Y, an explicit grid or analytical check, and sensitivity bands for sparse upper tails. A high observed deductible premium with few exceedances can be driven by one claim, one parameter assumption or one data revision.

If losses are truncated by policy limits, the observed paid-claim distribution is the distribution of a contract transformation rather than the latent ground-up loss. If the compared policies have different attachment thresholds, ordinary raw-claim CDF comparisons can confound insurance design and exposure severity. If the probability law shifts between training and future underwriting periods, a historical order does not necessarily persist. Nonstationarity, mixture shifts, selection changes, inflation and catastrophe dependence warrant additional structured sensitivity analysis.

Modeling error and partial identification

The relevant output should state what has been established at three distinct levels: whether the estimated laws are ordered; whether the statistical evidence supports the corresponding population assertion at an explicit error criterion; and whether future losses are plausibly governed by the population laws used in that assertion. Statistical significance at the second level does not create the third. Conversely, a full order may be mathematically present in a model while empirical data are too sparse to establish its parameters. The legitimate conclusion in such cases is an interval, an incomplete ranking, or an analysis of robust decisions, with the unresolved assumption made explicit.

Sensitivity, Counterexamples, and Decision Boundaries

Ordering relations are conditional propositions

The following implications hold under their stated integrability conditions:

  • X <=st Y implies X <=icx Y, because increasing convex functions are a subclass of increasing functions.
  • X <=cx Y implies X <=icx Y, because increasing convex functions are a subclass of all convex functions.
  • X <=cx Y implies E[X]=E[Y]; with finite second moments, Var(X)<=Var(Y).
  • X <=icx Y implies Pi_X(d)<=Pi_Y(d) for every real deductible d and ES_alpha(X)<=ES_alpha(Y) for every 0<alpha<1.
  • The converse of the single-variance implication fails, and an isolated VaR or stop-loss inequality does not establish any of the full orders.

The relation between first-order dominance and convex order requires care. If X <=st Y and the two means are equal and finite, the two distributions must be identical: first-order dominance with a strictly different law produces a strict mean inequality under integrability. Thus a strict first-order improvement and a nontrivial pure mean-preserving spread should not be conflated. Moreover, there is no universal inference from X <=icx Y to X <=st Y; the constant-versus-two-point example demonstrates this immediately.

Counterexamples as substantive evidence

The monograph contains four deliberately different logical tests. The positive-translation example establishes first-order dominance while variances tie. The A/B finite-support example shows that a higher variance is insufficient for convex order. The constant/rare-disaster example demonstrates the reversal of VaR ranking while convex order and ES comparisons are stable. The two-policy aggregation example shows that fixed marginals can generate materially different aggregate stop-loss and tail outcomes depending on dependence. Together they isolate mechanisms that single-number risk summaries cannot represent. Their value is proof of conditional logical claims, not empirical generalization to real insurers.

A fifth contract counterexample in the stop-loss chapter distinguishes unlimited excess-of-loss from capped layers. The comparison X=10 surely versus Y in {0,20} shows that unlimited insurer-layer expected cost at attachment 10 is higher under Y, while expected cost of a low capped layer is higher under X. This demonstrates how an economically relevant nonlinear contract payoff may fall outside the convex test-function class. It is a direct rebuttal to an otherwise tempting but invalid generalization from stop-loss order to every insurance-layer price.

Interpreting the hierarchy for insurance decisions

First-order dominance is the strongest simple statement here for monotone gross-loss evaluations: all survival probabilities are favorable. Increasing convex order is more tailored to high-loss obligations and to the preferences of risk-averse wealth holders. Convex order adds the equal-mean condition and thus interprets comparative dispersion. A particular risk measure can respect one of these relations while violating another. The useful unit of scientific reporting is consequently a triple: the pair of distributions, the precise order (or partial order) established, and the class of decision functionals for which the conclusion is valid.

The mathematical order should be applied to the risk bearer named in the decision. A policyholder evaluates retained loss, cash premium and wealth; an insurer evaluates contractual indemnity, expenses, reserves and capital; a reinsurer evaluates ceded layers and counterparty obligations. The same physical event produces different random payoffs for these parties. A statement that one contract improves a loss order for one party is an analytical result about that party's specified payoff, not an automatic claim of social welfare improvement.

Distinguishing theory, empirical claim, and editorial synthesis

The mathematical equivalences used in this study are established results attributed to their literature. The tabular examples and algebraic checks are original derivations for exposition. The principal analytical synthesis is the organized mapping of probability-law comparisons to insurance obligations, preference classes, contractual layers, estimation limits and dependence-sensitive aggregation. It creates no named authorial theory or proprietary doctrine. A stronger future empirical claim would require observed insurer or reinsurer data, source-specific adjustment of benefit structures and uncertainty in the tails, and the appropriate inference procedure.

Research Method, Reproducibility, and Evidentiary Limits

Source selection and inspection

The source program prioritizes original journal articles, publisher records, scholarly books from their publishers, and primary research repositories. The checked bibliography encompasses the classical economic ordering literature, martingale-coupling theory, modern stochastic-order textbooks, actuarial dependence papers, coherent risk measures, and statistical tests. Publisher landing pages provide accurate bibliographic metadata, abstract-level findings and issue/date distinctions; where the full article is restricted, this monograph does not claim line-by-line full-text verification. Theorems used as mathematical facts are supported by standard references and also checked by independent argument or exact finite constructions where feasible.

Source publication dates follow the original volume/year rather than the date a host platform digitized a paper. For example, Cambridge lists a later online-hosting date for the Kaas et al. paper, while the original ASTIN Bulletin issue is 2002. The fifth edition of Föllmer and Schied is a 2025 edition and should not be conflated with earlier editions. Where two sources share authors, the short contextual link names the particular work, making source identity visible in the reading flow. The accompanying EVIDENCE.json maps all externally linked scholarly claims to their publisher or research-source records, including inspection scope and caveats.

Analytical reproducibility

All quantitative examples require only the enumerated finite-support probability laws. The complete work can be checked by computing E[L], E[L^2], variance, F_L(t), Pi_L(d), VaR_alpha and ES_alpha from each probability table. Convex-order checks for finite supports reduce to an exact set of hinge comparisons, and the independent-versus-comonotonic two-policy distribution follows from Bernoulli multiplication and a shared Bernoulli event. The central martingale-noise result follows from conditional Jensen's inequality, and the stop-loss identity follows from Tonelli's theorem. No stochastic simulator, third-party proprietary data, random seed or confidential insurance dataset is required to reproduce the stated computations.

The automated manuscript quality gate checks structural identity, a complete bibliographic section, source-ledger connectivity, contextual external links, and SHA-256 fidelity. Human scholarly review additionally checks the sign and domain of every inequality, loss/wealth orientation, counterexamples, paper attribution and the distinction between validated theorem, illustrative calculation and inference. The published-site release check is a separate future operation and is not part of this research archive.

Limits and future empirical investigation

Mathematical ordering is exact when its assumed distributions are exact. The evidentiary challenge in insurance concerns identification of those distributions, settlement processes, legal wording, dependence and future conditions. The study has no access to confidential insurer portfolios, policy-level claims, actuarial rate filings or supervisory capital models; no country-specific adequacy result follows. Further empirical work could estimate robust stop-loss profiles for specified business lines, compare simultaneous confidence regions under varying dependence assumptions, analyze capped versus uncapped layer costs, and test out-of-sample persistence of observed orders under inflation and catastrophe-regime changes. Such projects would extend empirical coverage without implying a newly created theory.

Conclusion

Risk ordering is the disciplined comparison of uncertain outcomes against an explicitly declared class of thresholds, preferences or convex consequences. First-order stochastic dominance compares every loss threshold. Increasing convex order compares every unlimited stop-loss expectation and every integrated upper quantile. Convex order supplements those inequalities with equal means, thereby capturing a mean-preserving change in dispersion. For a fixed initial wealth, increasing convex order on losses is equivalent to an expected-utility improvement for every increasing concave utility on terminal wealth when the loss order is favorable. These identities provide generality that an isolated expected loss, variance or VaR cannot supply.

The original counterexamples establish the practical cost of confusing these concepts. Equal variance can coexist with strict first-order dominance. A larger variance can fail to establish convex ordering. VaR rankings can reverse across confidence levels even as a full convex order holds. A higher expected unlimited stop-loss payment need not imply a higher payment for every limited layer. Two portfolios with identical individual loss distributions can have different aggregate convex risk because dependence changes their joint outcomes. Each statement has been resolved by exact probability calculations or by a proved order implication rather than by an unexplained intuition about greater risk.

In insurance, the relevant scientific question is which random obligation belongs to whom and which ordering theorem applies to it. Gross loss, policyholder retention, insurer indemnity, ceded reinsurance layer, aggregate capital loss and post-premium wealth are different target variables. Their comparisons require contract-consistent definitions and a credible dependence model. Where order relations hold, they narrow permissible decisions for whole classes of economic or actuarial criteria; where the curves cross, the conclusion is a partial order and the remaining tradeoff requires an explicit decision rule.

The study's original contribution is its integrated analytical account of this translation from probability-law inequalities to insurance contracts, pooling and inferential limits, supported by independently derived examples. The established theorems retain their original scholarly attribution. The final scientific conclusion is that comparative risk analysis gains precision when it reports the distributional object, its stochastic order, all integrability and dependence assumptions, and the exact decision functionals that respect the order. This is the appropriate foundation for reproducible research on stochastic dominance in international insurance and risk science.

References