Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Risk Measures and Portfolio Capital Allocation

Abstract. Portfolio capital allocation attributes a single aggregate assessment of risk to its contributing positions, contracts, business lines, territories, and risk factors. The task requires a mathematical account of the aggregate functional, an explicit allocation objective, and evidence about the joint loss distribution. This monograph examines the foundations of risk measures, differentiability and Euler's theorem, marginal and incremental contributions, coherent and convex allocations, value-at-risk and expected-shortfall sensitivities, covariance-based models, Shapley and optimization-based alternatives, tail discontinuities, and applications to insurance portfolios. It derives full-allocation identities, distinguishes the effect of adding business from the assignment of existing aggregate capital, and establishes a reproducible two-line Gaussian example together with a discrete default/catastrophe counterexample. A central finding is that a stable aggregate capital number does not entail stable business-line risk shares: gradient estimation, capital definition, tail-event sampling, dependence assumptions, and constraints determine whether attribution is economically interpretable. The account places internal allocation alongside, rather than inside, the distinct legal calculation of regulatory capital. All numerical examples are original pedagogical constructions with disclosed assumptions. The synthesis extends established mathematical and actuarial scholarship without claiming a novel risk-allocation theory.

Keywords: Risk measure; economic capital; risk allocation; Euler allocation; marginal contribution; incremental capital; expected shortfall; value-at-risk; coherent risk measure; convex risk measure; subgradient; Shapley value; diversification; insurance portfolio; tail dependence; model risk; solvency.

Introduction

Capital is held at the level of a legal entity or group, whereas insurance losses arise through many contracts, risk classes, portfolios, and circumstances. A consolidated insurer may need one amount of resources to absorb a specified horizon of adverse net losses, yet need another set of figures to discuss underwriting capacity, compare reinsurance structures, price the cost of risk-bearing, or measure performance by business line. Capital allocation constructs that second set of figures. It translates a portfolio-level quantity into attributed contributions under a stated convention.

The starting question is mathematically simple: if losses are L1,…,LnL_1,\ldots,L_n, aggregate loss is L=∑iLiL=\sum_i L_i, and a measure ρ\rho assigns capital K=ρ(L)K=\rho(L), how should the numbers K1,…,KnK_1,\ldots,K_n be defined so that ∑iKi=K\sum_i K_i=K and the attribution has the desired meaning? Full allocation alone allows infinitely many answers. One could assign equal shares, shares proportional to premiums, stand-alone capital, marginal sensitivities, bargaining contributions, or an optimization objective. Their agreement requires special conditions.

The coherent-risk framework of Artzner and colleagues (1999) gave a disciplined account of how an aggregate capital functional can respond to diversification. Denault (2001), in his axiomatic study of allocation, connected the allocation problem to cooperative-game concepts and gradient contributions. The subsequent axiomatization by Kalkbrener (2005) clarified conditions under which a specified system of allocation axioms can be satisfied uniquely. Tasche's study of the Euler principle placed a broad collection of marginal risk contributions within a differentiable, homogeneous framework.

These works answer complementary questions. Risk-measure axioms concern the whole portfolio; an allocation rule distributes that quantity; a risk-management decision evaluates the consequences of the distribution. None of these questions, by itself, establishes that an amount is a regulatory capital requirement, a market price for capital, an actual loss reserve, or a transfer of funds. The conceptual separation is essential for insurance, where recoveries, asset values, contractual boundaries, supervisory requirements, risk appetite, and operational frictions interact.

The scope is the mathematical attribution of a given portfolio risk functional. Supervisory calibration, insurer-wide capital policy, and balance-sheet solvency management are adjacent disciplines. The analysis considers regulatory institutions only when they illuminate the difference between measurement and allocation. The research methods combine primary scholarly literature, exact algebra, explicitly constructed numerical cases, a discrete probability counterexample, comparative normative analysis, and an examination of model sensitivity. No proprietary insurer data or new empirical estimation is claimed.

Aggregate Capital Measures

The portfolio and the capital functional

Let (Ω,F,P)(\Omega,\mathcal F,\mathbb P) be a probability space and let LiL_i denote a monetary loss from unit ii during a common horizon, with positive amounts representing adverse outcomes. The total modeled loss is

L(w)=∑i=1nwiLi,K(w)=ρ(L(w)),L(w)=\sum_{i=1}^n w_iL_i, \qquad K(w)=\rho(L(w)),

where wi≥0w_i\geq0 scales exposure and ρ\rho maps a class of integrable or bounded random losses into a real capital amount. In a loss convention, monetary translation reads ρ(L+c)=ρ(L)+c\rho(L+c)=\rho(L)+c. In a net-worth convention, the sign of translation is reversed. Fixing this sign convention is necessary before comparing equations from different texts.

An internal capital metric might apply ρ\rho to gross claims, net claims after reinsurance, claims above expected cost, changes in net asset value, or an entire one-year balance-sheet result. Those are distinct random variables. A charge of ρ(L)\rho(L) and a charge of ρ(L)−E[L]\rho(L)-\mathbb E[L] differ by expected loss; their Euler allocations differ by each business line's expected loss. The treatment of the expectation must therefore be consistent across aggregation, allocation, and reported performance.

A coherent risk measure under the adapted loss convention satisfies monotonicity, cash translation, positive homogeneity, and subadditivity. Homogeneity means ρ(λL)=λρ(L)\rho(\lambda L)=\lambda\rho(L) for λ≥0\lambda\geq0, while subadditivity gives ρ(X+Y)≤ρ(X)+ρ(Y)\rho(X+Y)\leq\rho(X)+\rho(Y). The foundational coherence formulation of 1999 explains how these properties constrain aggregation. Positive homogeneity supports Euler decomposition; subadditivity permits diversification credit. Neither property alone states how that credit should be shared.

A convex risk measure replaces positive homogeneity and subadditivity as separate axioms with convexity, allowing nonlinear exposure effects and penalties for concentration. The theoretical scope of such functionals is developed in Föllmer and Schied's convex risk-measure analysis. Convexity is useful when doubling exposure does not double capital, but it also means Euler summation need not recover the entire amount. A risk function can be mathematically well defined even when it fails the axioms of coherence. A defensible allocation identifies which properties the chosen function actually possesses.

Quantiles, shortfall, and capital interpretation

For confidence level α∈(0,1)\alpha\in(0,1) define the lower-quantile value-at-risk of a loss as

VaR⁡α(L)=inf⁡{x:P(L≤x)≥α}.\operatorname{VaR}_{\alpha}(L)=\inf\{x:\mathbb P(L\leq x)\geq\alpha\}.

Value-at-risk supplies a quantile of the distribution, rather than an average of all losses exceeding that quantile. In particular, adding a portfolio component can shift the relevant quantile abruptly. VaR is positively homogeneous but generally fails subadditivity.

Expected shortfall is defined using the integrated quantile representation

ES⁡α(L)=11−α∫α1VaR⁡u(L) du,\operatorname{ES}_{\alpha}(L)=\frac{1}{1-\alpha}\int_{\alpha}^{1}\operatorname{VaR}_{u}(L)\,du,

when the integral is finite. The integrated definition behaves consistently with point masses at the quantile, which matter in insurance models with deductible triggers, capped layers, and binary catastrophe outcomes. Acerbi and Tasche's distribution-sensitive analysis of shortfall demonstrates why apparently interchangeable conditional-tail expressions can differ for discontinuous losses. Under an atom-free aggregate loss distribution with no ambiguity at the relevant threshold, ES⁡α(L)=E[L∣L≥qα]\operatorname{ES}_{\alpha}(L)=\mathbb E[L\mid L\geq q_\alpha]; with atoms, tail mass must be allocated fractionally.

Rockafellar and Uryasev's conditional value-at-risk optimization provides the equivalent variational representation

ES⁡α(L)=min⁡t∈R{t+11−αE[(L−t)+]},\operatorname{ES}_{\alpha}(L)=\min_{t\in\mathbb R}\left\{t+\frac{1}{1-\alpha}\mathbb E[(L-t)_+]\right\},

with suitable integrability conditions. This form makes ES amenable to scenario optimization and subgradient calculation. It is especially useful when the portfolio contains many lines and the insurer estimates a discrete simulation distribution. The minimizer need not be unique, so there may be several valid scenario-weight allocations even when the total ES is unique.

The choice between a quantile, an expected-shortfall functional, a standard-deviation measure, and a scenario capital bound is substantive. It determines which states exert marginal influence on the portfolio and how additional capacity alters aggregate risk. One cannot infer an allocation method solely from a generic reference to economic capital; the loss basis, time horizon, percentile, conditioning information, recoveries, and treatment of expected loss must be specified.

Risk-measure comparison

Feature Quantile VaR Integrated expected shortfall Standard deviation of loss
Functional focus Distribution percentile Average upper quantile tail Dispersion across all states
Positive homogeneity Yes Yes Yes
Subadditivity for arbitrary losses No Yes (integrable losses) Yes
Monotonicity in loss Yes Yes Not generally
Euler derivative Requires distribution smoothness Requires uniqueness or a selected subgradient Exists away from zero variance
Tail sensitivity Threshold neighborhood Whole upper tail Variance-weighted whole distribution

The table describes pure statistical functionals, not the capital formula of any particular regulator. Standard deviation is a useful aggregation benchmark but is not a coherent loss risk measure because a sure higher loss can have zero variance. VaR and ES require one common horizon and a coherent specification of the underlying random variable for cross-business attribution.

Marginal Risk Contributions

Three distinct questions

A stand-alone capital figure is Si=ρ(Li)S_i=\rho(L_i). It measures a unit as though detached from the rest of the portfolio, subject to the standalone model's own assumptions. Stand-alone figures usually do not sum to consolidated capital. Their excess D=∑iSi−ρ(∑iLi)D=\sum_iS_i-\rho(\sum_iL_i) is sometimes called diversification benefit, but DD may be negative for non-subadditive measures or altered models.

An incremental contribution is Δi=ρ(L)−ρ(L−Li)\Delta_i=\rho(L)-\rho(L-L_i), measuring the change in total risk when a complete unit is removed while holding other positions and assumptions fixed. This is a finite portfolio change, and incremental contributions generally fail to sum to KK. Their ordering can change with the definition of the counterfactual, especially when removing a large unit changes reinsurance contracts, market clearing, or model calibration.

A marginal contribution is the derivative of the capital functional with respect to a small scaling change: mi(w)=∂K(w)/∂wim_i(w)=\partial K(w)/\partial w_i, where the derivative exists. Its value is expressed per unit of exposure. The product wimi(w)w_im_i(w) is the component's Euler risk contribution. This distinction between marginal sensitivity, full-position Euler attribution, and complete removal avoids a common misuse of the word 'marginal.'

For a homogeneous differentiable functional the Euler allocation follows directly; for a nonsmooth functional, subgradients offer an extension. Tasche's analysis of portfolio risk contributions explicitly studies the estimation of such quantities and their use in performance measurement. Each statistic answers a different managerial question: incremental risk concerns a particular entry or exit, the derivative concerns a local exposure change, and full allocation concerns accountability for an already chosen total capital number.

The derivative identity

For L(w)L(w) linear in positions, define

KiE(w)=wi∂ρ(∑jwjLj)∂wi.K_i^{\mathrm E}(w)=w_i\frac{\partial\rho(\sum_jw_jL_j)}{\partial w_i}.

At actual exposure weights wi=1w_i=1, this becomes KiE=∂iK(1)K_i^{\mathrm E}=\partial_iK(\mathbf1). The derivative depends on the joint law: a line's own loss distribution alone is insufficient. In contrast, an incremental measure depends on the counterfactual ρ(L−Li)\rho(L-L_i); it need not equal a line integral of the gradient evaluated at the final portfolio weights. Indeed, if KK is smooth, then

ρ(L)−ρ(L−Li)=∫01∂K(1,…,t,…,1)∂wi dt.\rho(L)-\rho(L-L_i)=\int_0^1\frac{\partial K(1,\ldots,t,\ldots,1)}{\partial w_i}\,dt.

This integral holds where the stated regularity permits the fundamental theorem of calculus. Euler's marginal contribution evaluates the derivative at t=1t=1, whereas the incremental change averages along a path in which only component ii varies. The two coincide under a constant derivative along that path, which is more restrictive than portfolio-level homogeneity.

Why performance attribution needs a decision objective

Suppose one business line has modest expected profit and a large positive capital contribution, while another has a slightly smaller profit but a capital-reducing hedge effect. Dividing profit by allocated capital can be informative if the denominator represents the scarce resource in the relevant decision and the numerator is defined on a matching risk basis. A negative Euler contribution is mathematically possible for a hedge and can make ordinary performance ratios misleading. For portfolio expansion, the numerator and denominator should be incremental over the same decision path, with liquidity, risk appetite, contract costs, and capacity restrictions considered separately.

The allocation axioms articulated in Denault's capital allocation study and Kalkbrener's formal treatment give principled restrictions, but mathematical fairness cannot determine a single commercial cost-of-capital charge without a policy on what counts as a relevant business contribution. An insurer that assigns a corporate diversification benefit to an underwriting segment is making an internal attribution decision. Whether the segment may actually consume the resulting capital depends on enforceable transferability and legal-entity constraints.

Euler Allocation

Euler's theorem and full attribution

The relevant mathematical result is Euler's theorem for differentiable positively homogeneous functions. Suppose KK is continuously differentiable around a nonzero exposure vector ww and satisfies K(λw)=λK(w)K(\lambda w)=\lambda K(w) for all positive λ\lambda along the relevant ray. Differentiate with respect to λ\lambda at λ=1\lambda=1:

ddλK(λw)∣λ=1=∑iwi∂K(w)∂wi=K(w).\frac{d}{d\lambda}K(\lambda w)\bigg|_{\lambda=1} =\sum_iw_i\frac{\partial K(w)}{\partial w_i}=K(w).

Consequently, ∑iKiE(w)=K(w)\sum_iK_i^{\mathrm E}(w)=K(w). The proof is a chain-rule identity. It does not require independence of the LiL_i, normality, subadditivity, or a regulatory correlation matrix. Its indispensable conditions are appropriate homogeneity and differentiability of the actual aggregate functional. The link between gradients and allocation is developed in Tasche's Euler treatment, while the formal axiom structure is compared in Kalkbrener's allocation analysis.

Full allocation is not a claim that the components would jointly need these amounts if each were operated independently. Nor does it imply Ki≥0K_i\geq0 for every line: adding a natural hedge can lower aggregate risk, producing a negative derivative. When ρ\rho is coherent, its monotonicity refers to losses ordered pointwise; it does not imply that increasing exposure to an arbitrarily signed random component always raises risk.

Subadditivity and marginal bound

If ρ\rho is subadditive and positively homogeneous, then for any h>0h>0,

ρ(L+hLi)≤ρ(L)+hρ(Li).\rho(L+hL_i)\leq\rho(L)+h\rho(L_i).

If the directional derivative at LL exists, dividing by hh and taking h↓0h\downarrow0 yields mi≤ρ(Li)m_i\leq\rho(L_i). Thus the Euler contribution is no greater than the scaled standalone risk under the same homogeneous convention. The inequality is a local consequence of subadditivity. It does not guarantee a nonnegative contribution or a specific division of diversification savings. Where the risk measure is VaR, the inequality need not hold because VaR is not generally subadditive.

The total subadditive diversification benefit can be written as

D=∑iρ(Li)−ρ(L)=∑i[ρ(Li)−KiE]D=\sum_i \rho(L_i)-\rho(L)=\sum_i\left[\rho(L_i)-K_i^{\mathrm E}\right]

when Euler full allocation applies. This identity assigns a component's implied diversification benefit relative to its own standalone figure. A different, equally full allocation would produce different component credits without changing the portfolio's total DD.

Convex or otherwise nonhomogeneous capital

Suppose management defines K(w)=ES⁡α(L(w))+B(w)K(w)=\operatorname{ES}_{\alpha}(L(w))+B(w), where BB is an explicitly specified capital surcharge that is nonlinear in business volume. For example, B(w)=γ(∑iwi)2B(w)=\gamma(\sum_iw_i)^2 with γ>0\gamma>0. Then the Euler sum of BB is 2B2B, although the aggregate surcharge is only BB. Euler sensitivities remain meaningful as local indicators, but they no longer provide a full allocation of total capital. A fixed surcharge B(w)=bB(w)=b presents the opposite case: its Euler sum is zero, leaving an unallocated constant bb.

A transparent reconciliation has at least three legitimate forms. Management may allocate the homogeneous core by Euler and separately allocate the surcharge by an explicit policy; it may use an Aumann–Shapley path method, provided the path begins from a well-defined baseline; or it may accept incomplete marginal attribution for sensitivity analysis and use a separate accounting allocation for the total. Each choice should identify its residual and baseline. An arbitrary final rescaling of Euler derivatives to make their sum equal total capital changes their marginal interpretation.

The distinction between proportional exposure scaling and nonlinear policy cost helps explain the more general allocation frameworks considered by Dhaene and coauthors (2012). Optimization-based allocation objectives can be appropriate where a company is seeking a policy-specific sharing of capital rather than a pure local derivative of aggregate risk.

Nonsmooth risk functions

For a coherent or convex risk functional, nondifferentiability can arise from scenario switches, identical losses at quantiles, active insurance layers, or competing worst-case models. For a convex homogeneous finite-dimensional KK, select a subgradient g∈∂K(w)g\in\partial K(w). The subgradient inequality gives K(z)≥K(w)+g⋅(z−w)K(z)\geq K(w)+g\cdot(z-w), and homogeneity implies g⋅w=K(w)g\cdot w=K(w): use z=0z=0 and z=2wz=2w to obtain the two inequalities that force equality. Thus Ki=wigiK_i=w_ig_i is a full allocation for every selected supporting subgradient, but the allocation can be nonunique. That conclusion is mathematically narrower than asserting that one unique Euler gradient exists.

In an expected-shortfall model with discrete scenarios, a subgradient can be represented by scenario weights placed on the worst losses, with enough fractional weight at the threshold to fill exactly the upper-tail probability mass. When several scenarios tie at the threshold, more than one valid weighting may produce the same total ES yet different component allocations. The robust integrated-tail distinction established by Acerbi and Tasche (2002) prevents the common mistake of treating every threshold tie as a wholly included tail observation.

Gradient versus Aumann–Shapley paths

Aumann and Shapley's non-atomic game-value framework supplies a conceptual route to infinitesimal contribution sharing. For a suitably smooth KK, a ray-path attribution relative to a baseline K(0)=0K(0)=0 takes the form

Ai(w)=wi∫01∂iK(tw) dt.A_i(w)=w_i\int_0^1\partial_iK(tw)\,dt.

The sum is K(w)−K(0)K(w)-K(0) by the chain rule along the path. Under degree-one homogeneity, ∂iK(tw)=∂iK(w)\partial_iK(tw)=\partial_iK(w) for t>0t>0, so Ai=KiEA_i=K_i^{\mathrm E}. This equivalence depends on the selected path and differentiability. For nonhomogeneous risk functions the integral differs from the endpoint derivative and is a method of allocating the change from the baseline. It is essential to say which method is being used rather than treating the words 'Euler,' 'gradient,' and 'Aumann–Shapley' as universal synonyms.

Diversification and Attribution

Covariance structure and analytical contributions

A transparent benchmark is the standard deviation functional. Let Σ\Sigma be the covariance matrix of component losses and ww the vector of exposure weights. Then

σ(w)=wTΣw,∂σ∂wi=(Σw)iσ(w),Kiσ=wi(Σw)iσ(w)\sigma(w)=\sqrt{w^{\mathsf T}\Sigma w},\quad \frac{\partial\sigma}{\partial w_i}=\frac{(\Sigma w)_i}{\sigma(w)},\quad K_i^\sigma=\frac{w_i(\Sigma w)_i}{\sigma(w)}

for σ(w)>0\sigma(w)>0. Summing the last expression gives σ(w)\sigma(w). A contribution reflects both the unit's variance and its covariances with the other units. Under a Gaussian joint loss model with mean μ(w)=wTμ\mu(w)=w^{\mathsf T}\mu,

VaR⁡α(L(w))=μ(w)+zασ(w),ES⁡α(L(w))=μ(w)+φ(zα)1−ασ(w),\operatorname{VaR}_\alpha(L(w))=\mu(w)+z_\alpha\sigma(w),\qquad \operatorname{ES}_\alpha(L(w))=\mu(w)+\frac{\varphi(z_\alpha)}{1-\alpha}\sigma(w),

where zα=Φ−1(α)z_\alpha=\Phi^{-1}(\alpha) and φ\varphi is the standard normal density. Euler contributions become wiμi+cαwi(Σw)i/σ(w)w_i\mu_i+c_\alpha w_i(\Sigma w)_i/\sigma(w), using cα=zαc_\alpha=z_\alpha for VaR and cα=φ(zα)/(1−α)c_\alpha=\varphi(z_\alpha)/(1-\alpha) for ES. This is an exact Gaussian-model identity, not a justification for normally distributed catastrophe, casualty, cyber, or long-tail insurance claims.

Fully reproduced two-line insurance portfolio

Consider two hypothetical net-loss business lines with a common one-year horizon and jointly Gaussian losses. Line A has expected loss 33 and standard deviation 88; line B has expected loss 44 and standard deviation 1212; the correlation is 0.250.25. All monetary figures below use arbitrary common units, interpreted as millions if desired. The covariance is 8×12×0.25=248\times12\times0.25=24. Thus total mean loss is 77, total variance is 64+144+2(24)=25664+144+2(24)=256, and total standard deviation is 1616.

The standard-deviation Euler contributions are

KAσ=(64+24)/16=5.5,KBσ=(144+24)/16=10.5.K_A^\sigma=(64+24)/16=5.5,\qquad K_B^\sigma=(144+24)/16=10.5.

Using z0.99≈2.326347874z_{0.99}\approx2.326347874 and φ(z0.99)/0.01≈2.665214220\varphi(z_{0.99})/0.01\approx2.665214220, the 99% VaR and ES calculations follow. Values are rounded independently for presentation; the exact formulas determine totals.

Quantity Line A Line B Consolidated
Mean loss 3.000 4.000 7.000
Stand-alone standard deviation 8.000 12.000 16.000 (aggregate)
Stand-alone VaR at 99% 21.611 31.916 —
Euler VaR at 99% 15.795 28.427 44.222
Stand-alone ES at 99% 24.322 35.983 —
Euler ES at 99% 17.659 31.985 49.644

The 99% VaR portfolio capital is 7+16(2.326347874)≈44.22167+16(2.326347874)\approx44.2216. The Euler contributions are 3+5.5(2.326347874)≈15.79493+5.5(2.326347874)\approx15.7949 and 4+10.5(2.326347874)≈28.42674+10.5(2.326347874)\approx28.4267. The 99% ES portfolio value is 7+16(2.665214220)≈49.64347+16(2.665214220)\approx49.6434, with allocation 3+5.5(2.665214220)≈17.65873+5.5(2.665214220)\approx17.6587 and 4+10.5(2.665214220)≈31.98474+10.5(2.665214220)\approx31.9847. Stand-alone VaR sums to approximately 53.52753.527 and stand-alone ES sums to approximately 60.30560.305; the differences from aggregate capital reflect modeled diversification. These figures are generated from the stated parameters, rather than fitted to insurer observations.

Under this normal model, line A's portfolio covariance contribution is 8888 and line B's is 168168. Increasing correlation would increase σ\sigma and thereby change both aggregate capital and the distribution of its capital shares. A reinsurance recovery that changes one line's variance or dependence could have a different effect on its standalone capital, its Euler contribution, and its incremental capital. This explains why a portfolio allocation cannot be constructed reliably from standalone standard deviations without the covariance structure.

Cooperative-game comparison

A different fairness objective treats the business lines as participants in a cooperative cost game with coalition costs v(S)=ρ(∑i∈SLi)v(S)=\rho(\sum_{i\in S}L_i) and v(∅)=0v(\varnothing)=0. In a two-line game, the Shapley allocation is

ϕA=12[v({A})+v({A,B})−v({B})],ϕB=12[v({B})+v({A,B})−v({A})].\phi_A=\tfrac12\big[v(\{A\})+v(\{A,B\})-v(\{B\})\big],\quad \phi_B=\tfrac12\big[v(\{B\})+v(\{A,B\})-v(\{A\})\big].

Its interpretation rests on averaging incremental contributions over possible entry orders, according to the cooperative-game valuation in Shapley's original game-value chapter. For the Gaussian VaR example, the result is approximately ϕA=16.958\phi_A=16.958 and ϕB=27.263\phi_B=27.263. Both Shapley and Euler allocations sum to approximately 44.22244.222, but Shapley assigns a somewhat larger share to A. This is a policy difference, rather than numerical error: Euler reports a derivative of the final combined portfolio; Shapley averages the costs of joining different coalitions.

With many insurance divisions, Shapley calculation requires evaluating many coalitions, often a high computational burden; approximations based on random permutations also introduce sampling uncertainty. The approach may become especially meaningful when individual business lines have heterogeneous access to risk-pooling arrangements or transferability restrictions. It still depends critically on whether a standalone or coalition cost is defined on an economically feasible basis.

Optimal-allocation objectives

A portfolio manager may seek capital assignments that minimize deviations of business-line losses from their assigned capital under an explicitly selected loss function. In the optimal allocation principles of Dhaene et al., a weighted optimization setup can represent alternative corporate risk tolerances and reproduce multiple allocation rules. Such a method makes the corporate objective explicit. The theoretical lesson is that full allocation and a chosen measure of fairness do not determine a unique portfolio valuation independent of purpose.

A suitable comparison of approaches therefore asks what mathematical property is being preserved: the derivative of total risk, marginal coalition participation, closeness between a unit's loss and its assigned capital, or a policy-defined distribution of scarcity costs. None should be advertised as providing the other properties automatically. For internal insurance governance, the selected objective is part of the method's specification, alongside the stochastic model and calibration.

Tail Events and Risk Shares

A discrete counterexample to VaR subadditivity

The Gaussian example is differentiable, but the insurance setting frequently contains discontinuous loss indicators. Construct two independent one-year losses XX and YY. Each equals 100100 with probability 0.0060.006 and zero with probability 0.9940.994. Because P(X=0)=0.994>0.99\mathbb P(X=0)=0.994>0.99, both VaR⁡0.99(X)\operatorname{VaR}_{0.99}(X) and VaR⁡0.99(Y)\operatorname{VaR}_{0.99}(Y) are zero. Their sum is therefore zero.

The combined loss Z=X+YZ=X+Y has probabilities

P(Z=0)=0.9942=0.988036,P(Z=100)=2(0.006)(0.994)=0.011928,P(Z=200)=0.0062=0.000036.\mathbb P(Z=0)=0.994^2=0.988036,\quad \mathbb P(Z=100)=2(0.006)(0.994)=0.011928,\quad \mathbb P(Z=200)=0.006^2=0.000036.

The cumulative probability at 00 is below 0.990.99 and the cumulative probability at 100100 exceeds 0.990.99, so VaR⁡0.99(Z)=100\operatorname{VaR}_{0.99}(Z)=100. Thus VaR assigns zero to each unit alone but 100100 to their aggregate. This explicit construction proves that VaR can violate subadditivity. It does not establish that diversification increases risk under every risk measure. The conclusion is consistent with the quantile limitations analyzed in the coherent-measure literature.

Now apply integrated 99% expected shortfall. For one individual loss, the worst 1% contains a probability mass 0.0060.006 at 100100 and 0.0040.004 at zero; hence ES⁡0.99(X)=ES⁡0.99(Y)=60\operatorname{ES}_{0.99}(X)=\operatorname{ES}_{0.99}(Y)=60. For the portfolio, the worst 1% contains all probability 0.0000360.000036 at 200200 and the remaining 0.0099640.009964 at 100100. Therefore

ES⁡0.99(Z)=0.000036(200)+0.009964(100)0.01=100.36.\operatorname{ES}_{0.99}(Z)=\frac{0.000036(200)+0.009964(100)}{0.01}=100.36.

The combined ES, 100.36100.36, is below the standalone sum of 120120. By the exact symmetry of XX and YY, an exchangeable, fully allocated ES attribution assigns 50.1850.18 to each unit. The quantile-threshold ties make the choice of tail scenario weights relevant to differentiated portfolios; symmetry resolves the present example. The expected-shortfall coherence treatment is directly relevant to this boundary case.

Quantile derivatives and conditional component losses

For continuous distributions under sufficient smoothness, an important local identity is

∂∂wiVaR⁡α(L(w))=E[Li∣L(w)=qα(w)],\frac{\partial}{\partial w_i}\operatorname{VaR}_{\alpha}(L(w)) =\mathbb E[L_i\mid L(w)=q_\alpha(w)],

where the conditional expectation at a zero-probability level set is interpreted via a regular conditional distribution and appropriate density limits. This is a conditional expectation at the quantile boundary, not an average over the upper tail. The distinction becomes consequential when portfolios contain rare, heavy events with high losses far beyond the threshold.

Under compatible regularity and an atom-free threshold,

∂∂wiES⁡α(L(w))=E[Li∣L(w)≥qα(w)].\frac{\partial}{\partial w_i}\operatorname{ES}_{\alpha}(L(w)) =\mathbb E[L_i\mid L(w)\geq q_\alpha(w)].

For a scenario distribution with probability atoms, the valid formula replaces the simple conditional expectation by a tail-weighted expectation with fractional selection at the quantile. In both formulas, the conditional distribution is a property of the aggregate portfolio and its joint dependence with each component. This is why loss correlation, dependence modeling and scenario generation are part of attribution, rather than optional add-ons.

Rare-event sampling error

For a simulated portfolio with NN independent annual scenarios, a 99% expected-shortfall estimate draws its primary information from roughly 0.01N0.01N tail observations. At N=10,000N=10{,}000, this is only about 100 aggregate-tail scenarios; for 99.5%, about 50. Component attribution uses the same scarce scenarios and can fluctuate more than the consolidated total if different sublines contribute differently to the worst losses. Autocorrelation, importance-sampling weights, fitted distributions, and scenario ties can alter the effective sample size.

The empirical investigation reported in Boonen's Euler-allocation study finds substantial measurement sensitivity, especially for value-at-risk-based Euler allocations, with lower but still meaningful volatility for expected shortfall. A parameter-fitted distribution can stabilize estimates under an appropriate model but also introduces specification risk. An attractive-looking smooth attribution from a mis-specified dependence model is not necessarily superior to a noisy nonparametric estimate.

A complete validation exercise should therefore compare multiple independent simulation seeds, record numerical confidence intervals or bootstrap ranges where justified, test alternative tail definitions and levels, and vary dependence structures. Bootstrap coverage for rare catastrophes can be poor if extreme-event mechanisms are absent from historical data; statistical resampling cannot manufacture missing peril information. Model uncertainty belongs in the interpretation of a contribution, not merely in a footnote to the aggregate figure.

Concentration and nonlinearity

Dependence structures that agree in linear correlation can differ sharply in joint catastrophe tail probabilities. A portfolio of geographically adjacent properties may experience the same flood, wildfire, or windstorm; combining it with risks that appear uncorrelated during ordinary years can still generate exceptional joint losses under systemwide shocks. A line's Euler attribution for ES then depends on how frequently its claims appear in the aggregate portfolio's tail, not simply its expected losses or standalone severity.

Insurance policies introduce nonlinear functions of underlying physical damage. A catastrophe treaty with retention dd and limit uu pays min⁡{(L−d)+,u}\min\{(L-d)_+,u\}, and a net retained loss can be written as L−min⁡{(L−d)+,u}L-\min\{(L-d)_+,u\}. Changing portfolio scale may shift the distribution across treaty attachment and exhaustion points. Fixed contractual limits can break simple homogeneity of a capital function defined after such recoveries when the exposure scaling also changes contractual economics. The exposure path must preserve, or explicitly change, reinsurance terms. In a model where weights only scale fixed net-loss random variables, homogeneity can still hold; in a model that recomputes a nonlinear recovery as exposure changes, the derivative is generally different.

The practical implication is a controlled distinction between sensitivities to underwriting volume, claim severity, inflation, correlation, and contract structure. One number labeled 'line capital' cannot summarize all those mechanisms. A well-documented allocation can reveal exposure concentration while retaining separate stress tests for threshold effects and model discontinuities.

Insurance Portfolio Cases

Case: personal lines and catastrophe aggregation

Consider an insurer writing automobile, property, and small-business coverage. Auto liability has relatively frequent, heterogeneous claims with a potentially long reporting and settlement tail. Property has shorter claims-settlement timing in some events but large event accumulation across a territory. Small-business policies may include both local property and liability exposures that are correlated with the first two lines. These are qualitative examples; no specific market's observed claim frequencies or correlation coefficients are assumed.

The correct setup models the common portfolio-year result for each line under a consistent gross or net basis. Premium, claims reserves, claims inflation, reinsurance recoveries, counterparty default, asset movements, and expenses must be assigned to the same reporting and time horizon as the capital metric. With a scenario engine producing Li,sL_{i,s} and probabilities psp_s, the portfolio tail is determined from Ls=∑iLi,sL_s=\sum_iL_{i,s}, not by ranking each line's individual tail independently. Each component's ES attribution is calculated using its losses in the aggregate-tail states.

For example, a low-frequency coastal flood scenario may affect both personal property and small business in the same geographic basin. A low stand-alone mean claim amount for small business does not prove a low capital share if those losses reliably occur in the states where the insurer is worst off. Conversely, diversifying into a low-covariance region can lower a line's aggregate marginal contribution even while increasing its premiums and standalone risk. This interpretation follows from the conditional-tail equations above and makes capital attribution an instrument for examining the joint distribution, not a substitute for catastrophe science.

Case: reinsurance and retained risk

Suppose an insurer purchases an excess-of-loss catastrophe cover. Compute both gross loss LiGL_i^{G} and recoveries RiR_i, and define net retained loss LiN=LiG−RiL_i^{N}=L_i^{G}-R_i, including the possibility of recovery delays, reinstatement premiums, and counterparty failure where those risks are relevant and modeled. The capital changes by ρ(∑iLiN)−ρ(∑iLiG)\rho(\sum_iL_i^{N})-\rho(\sum_iL_i^{G}) under the chosen risk measure. This is an aggregate change, which is distinct from the distribution of the post-reinsurance net capital among business lines.

An internal attribution that assigns every recovery to the line generating the ceded claim can differ from the marginal portfolio benefit when a treaty is shared across lines or when recoveries are exhausted by concurrent events. If the treaty attaches at the insurer level, treating each line's hypothetical standalone recovery as a separate available contract can double count protection. The coalition or counterfactual model must preserve legal treaty terms and recognize exhaustion. Portfolio allocation and the economic value of the treaty answer related but distinct questions.

In a stress test, a reinsurer's failure or a dispute over coverage can transform a low retained event into a high retained event. Economic capital on a net basis must therefore state the recoverability assumptions that connect catastrophe severity to insurer solvency. A computationally precise Euler allocation under a model that omits disputed recoveries cannot compensate for that omission.

Case: underwriting entry and capital capacity

Assume a manager considers adding a new book of business. The decision statistic should begin with incremental consolidated risk ρ(L+δLnew)−ρ(L)\rho(L+\delta L_{new})-\rho(L) or a local derivative for infinitesimal δ\delta, with consistent additional premiums, expected claims, expenses, and reinsurance premiums. A full-position Euler allocation of the existing portfolio is useful for periodic performance reporting but may be different from the relevant entry-cost statistic. Under large capacity changes, pricing and contract terms may change in nonlinear ways, requiring a complete counterfactual.

Allocating total enterprise capital in proportion to each line's premiums may be appropriate for a managerial overhead convention, but it is generally unrelated to the derivative of a tail-risk functional. Conversely, mathematically valid tail attribution alone does not specify the price a reinsurer or policyholder will pay for transfer, nor the liquidity requirement associated with settlement timing. A useful operating framework reports at least aggregate capital, assigned capital, incremental capital for proposals, and documented capacity constraints as separate quantities.

Insurance supervision and its proper boundary

European Solvency II distinguishes the regulatory calculation of capital from internal attribution. Under Article 104 on basic SCR, prescribed risk modules include non-life underwriting, life underwriting, health underwriting, market, and counterparty default risks. The standard-formula framework uses prescribed aggregation and module calibration. The implementing basic SCR calculation rule adds elements such as intangible risk according to the applicable version. A simple Gaussian Euler model of accounting losses is therefore not automatically a compliant Solvency II module allocation, even if its algebra resembles a correlation aggregation formula.

Under Article 45 on ORSA, an undertaking assesses overall solvency needs in light of its specific risk profile, tolerance limits, and strategy, and examines deviations from assumptions underlying the regulatory SCR. The EIOPA ORSA Guidelines set an institution-specific governance context. These documents do not prescribe one universal internal Euler allocation to individual lines of business. Their significance here is that measurement choices, management interpretation, documentation, and regulatory compliance operate at distinguishable levels.

Regulatory formulas, economic capital models, and allocations are connected by decisions about solvency and resource use, but their numbers need not coincide. Different legal entities may have ring-fenced funds, constraints on fungibility, supervisory interventions, or local capital requirements. An attribution computed as if all capital were freely transferable can misstate available diversification. A group-level allocation must explicitly articulate the legal, contractual, and jurisdictional boundaries of risk transfer.

Governance, Robustness, and Interpretation

A reproducible allocation protocol

An internally auditable procedure begins by identifying the economic object: gross claims, net insurance losses, net-asset change, or unexpected loss. Next it fixes legal entity, horizon, confidence level, currency, discounting, scenario probability measure, and capital basis. It then establishes the full portfolio risk functional and checks whether homogeneity, convexity, subadditivity, and differentiability genuinely apply. The selection of a capital allocation objective follows: risk sensitivity, coalition fairness, optimization criterion, or an administrative cost policy.

For Euler attribution, the analyst must specify weights, the exposure path, gradient or subgradient method, and reconciliation tolerance. With a simulation model, the record should contain scenario generation parameters, tail-event selection, tie treatment, reinsurance recoveries, model version, estimation error and any constraints on negative contributions. For a cooperative game, it should identify the feasible coalitions and how coalition costs were calculated. For optimization-based allocations, it should show the objective, weights, constraints, and sensitivity to competing objectives.

The central control is a reconciliation ledger that records KK, the assigned KiK_i, any explicit residual, and the reason that a residual exists. A zero numerical residual proves bookkeeping consistency under a specified procedure. It does not by itself validate the probability model, justify the economic interpretation, or establish legal capital sufficiency.

Stability of aggregate and component estimates

Consider two calibrated models that yield almost the same portfolio ES but differ on which line causes the most extreme outcomes. Their aggregate capital may agree because the magnitude of total tail losses is similar; their line allocations may diverge because the conditional composition of those losses is different. This is a structural reason why attribution is more sensitive to joint scenario composition than aggregate risk alone. Boonen's empirical sensitivity findings reinforce this concern for estimated Euler allocations.

A useful stability analysis compares the rank, sign, and magnitude of component contributions under plausible alternatives: independent versus dependent catastrophes; alternative climate/peril catalogues where available; different inflation and reserve assumptions; stable versus disrupted reinsurance recoveries; alternative loss-distribution families; and both VaR and ES at justified confidence levels. These are model changes, so their impact should be reported as sensitivity rather than measurement error around a fixed model.

For a sufficiently differentiable KK, the Hessian ∂ijK\partial_{ij}K describes how a marginal contribution changes when another line changes its exposure. Even when every line has a clearly defined first derivative, large cross-derivatives can make the attribution unstable under portfolio changes. If a risk function is nonsmooth, one-sided derivatives and subgradient ranges can be more informative than a single asserted point estimate.

What allocation can and cannot establish

Attribution has at least three evaluative dimensions. Mathematical validity asks whether the formula follows from the stated risk functional and assumptions; empirical credibility asks whether model losses and dependence structures adequately represent the insurer's potential experience; institutional relevance asks whether the assigned capital supports a particular managerial question under actual legal and operational constraints. The three dimensions must be evidenced separately.

This leads to a precise distinction. A mathematical statement such as ∑iKi=K\sum_iK_i=K can be proved. A reported numerical allocation can be independently recalculated under a disclosed synthetic model. A claim that an allocation is fair to policyholders, optimal for pricing, or sufficient to protect an actual insurer requires additional economic, contractual, empirical, and legal grounds. The framework developed by Dhaene et al. (2012) is useful precisely because an explicitly selected objective gives the word 'optimal' a meaning.

When capital allocation informs underwriting limits, the final decision should also consider risk appetite, available capital in the relevant entity, liquidity, insurance policy obligations, operational resilience, profitability under risk, and stress scenarios. The present research explains the analytical infrastructure of attribution; it does not produce an individualized supervisory assessment, actuarial pricing opinion, or risk-management recommendation for a specific insurer.

Attribution Under Model Change and Organizational Boundaries

Mathematical reconciliation across model versions

An insurer can change its loss model while leaving its observed portfolio composition unchanged. Let K(0)=ρ(0)(L(0))K^{(0)}=\rho^{(0)}(L^{(0)}) be the base capital estimate and K(1)=ρ(1)(L(1))K^{(1)}=\rho^{(1)}(L^{(1)}) be the estimate after a model or data update. An exact Euler reconciliation within each version yields ∑iKi(r)=K(r)\sum_iK_i^{(r)}=K^{(r)} for r∈{0,1}r\in\{0,1\}. It follows that ∑i(Ki(1)−Ki(0))=K(1)−K(0)\sum_i(K_i^{(1)}-K_i^{(0)})=K^{(1)}-K^{(0)}. This identity is useful for audit, but it does not identify whether the capital movement was caused by new policies, loss inflation, updated dependence, parameter fitting, revised reinsurance recoveries, or a new risk functional. Those mechanisms can change together.

A controlled comparison therefore specifies a sequence of counterfactual model updates. For example, evaluate the base model with the old portfolio, the base model with new exposures, the revised frequency-severity model with those exposures, and finally the revised dependence and treaty model. The observed sequence decomposes the difference according to the chosen order. Changing the order can change intermediate contributions if factors interact; there is no unique order-free decomposition without an additional sharing convention. A Shapley-style average over update orders could be applied if the cost and available model states are well defined, but such a calculation would represent an analyst-selected attribution of model change and should be labeled as such.

The result matters for governance because an internal report can show a line's assigned risk increasing even when its actual exposure declines. Such a finding need not be contradictory. If the line becomes more correlated with the insurer's systemic tail or the model changes the ranking of aggregate loss scenarios, the conditional component loss within the tail can grow. Analysts should report both exposure change and risk-share change with traceable calculation stages.

Allocation of a loss and allocation of capital costs

A quantitative measure of risk and the economic price of financing that risk are not the same function. Suppose the insurer holds KK units of capital and incurs an illustrative annual financing charge cKcK, with a constant scalar c>0c>0. If the capital metric is differentiable and homogeneous, allocating the financing charge proportionally to Euler capital produces cKicK_i and reconciles exactly to cKcK. Yet the assumption of one constant cc is economic. Different financing constraints, availability of fungible funds, tax treatment, debt covenants, or legal-entity arrangements can mean that the marginal cost of capital varies across businesses.

If the funding-cost function is F(K)F(K), then an infinitesimal increase in exposure changes financing cost by F′(K)∂iKF'(K)\partial_iK, when differentiable. Summing exposure-weighted derivatives gives F′(K)KF'(K)K, which equals F(K)F(K) only for particular functional forms such as a proportional charge without a fixed component. Assigning all of F(K)F(K) then requires an explicit policy about the difference. This algebra explains why even a perfectly reconciled allocation of risk capital cannot silently become a reconciled allocation of the economic cost of that capital.

The same issue appears in pricing. An insurance policy's expected loss, operational expenses, reinsurance cost, taxes, capital charge, and demand-related commercial margin are different terms. A line's tail contribution can inform one of them, under the proper model, but it does not define their sum. The optimal-allocation framework offers a useful reminder that the meaning of a capital assignment is a function of the objective chosen for it.

Nested allocations and organizational consistency

Capital is often allocated hierarchically: first to risk types, then to business lines, then to products, and finally to geographic cells. Suppose a top-level portfolio comprises groups GG and each group contains components i∈Gi\in G. If the same differentiable homogeneous aggregate functional K(w)K(w) is evaluated on a single consistent exposure space, group Euler contributions add up to the component Euler contributions within the group. Indeed,

KGE=∑i∈Gwi∂iK(w),∑GKGE=K(w).K_G^{\mathrm E}=\sum_{i\in G}w_i\partial_iK(w),\qquad \sum_GK_G^{\mathrm E}=K(w).

However, if each group first constructs and allocates a standalone group-specific capital number, then a second allocation at the enterprise level may treat diversification differently. Those independently estimated figures need not sum to the original consolidated risk. A hierarchical dashboard must either propagate the same enterprise gradient into lower levels or reconcile the distinct model levels and diversification charges explicitly.

Nested allocation also changes when reinsurance cover, legal segregation, and exposure currency are modeled at one organizational level but not another. A line-level model that treats a shared catastrophe treaty as an independently replicable protection for every product can manufacture additional recoveries. Consolidated net losses should be built first from legally feasible policy and treaty mechanics, followed by attribution to components under a consistent convention.

Negative contributions and hedging

A negative risk contribution means a position locally decreases the selected aggregate measure at the prevailing portfolio weights. It is not a negative claim liability or a negative amount of legally required regulatory capital. In the standard-deviation model, a component with negative covariance to the aggregate can receive a negative Euler standard-deviation contribution even though its own standalone variance is positive. If managerial accounting prohibits negative allocations, a constrained positive allocation may be introduced, but it no longer equals the raw Euler derivative. The report should retain the unconstrained derivative as a risk signal and display the constrained accounting number separately.

This distinction is especially important for hedges, reinsurance protections, and investments that offset insurance liability risk. A position with negative derivative may still have positive expected cost, counterparty risk, liquidity needs, and basis risk. Those items deserve their own assessment. A numerical rule that mechanically rewards every negative tail contribution with an equivalent commercial subsidy would overstate what the mathematical derivative establishes.

Temporal consistency

A one-year allocation is a snapshot of the loss distribution and risk measure defined at the valuation date. A long-duration life or casualty portfolio may have risk extending far beyond that year. Rolling forward the model changes the information set, the insured exposure, claim reserves, asset holdings, and available reinsurance. The sum of annual allocated capitals is not generally a lifetime capital number, nor does the sum of daily quantiles equal a one-year quantile. Any temporal aggregation requires an explicit stochastic evolution and a suitable dynamic risk-measure framework.

A board-level report should identify whether a comparison is contemporaneous across business lines, a historical comparison of repeated one-year snapshots, or a present-value assessment of future capital costs. The time convention constrains the allowable interpretation of both capital and diversification. Annual supervisory or internal metrics can support decisions about longer-run resilience when combined with scenario analysis; the mathematical attribution alone has the specified horizon.

Conclusion

Portfolio risk attribution is a mathematically conditional interpretation of consolidated capital. The capital functional determines the aggregate quantity, the joint portfolio distribution determines exposure contributions, and the allocation rule determines how a chosen total is assigned to components. Full allocation, fairness, sensitivity, diversification credit, and decision usefulness are distinguishable requirements. Their compatibility depends on assumptions and the purpose of measurement.

For differentiable and positively homogeneous risk measures, Euler's theorem gives an exact decomposition into exposure-weighted gradients. This result explains the strong role of Euler allocation in capital-performance measurement. It also reveals its boundaries: subadditivity is relevant to diversification but is not necessary for the Euler identity; convex functions can be nonsmooth and admit multiple full subgradient allocations; nonhomogeneous capital requirements need an explicit residual or a different path/allocation rule.

The numerical examples establish complementary points. A two-line Gaussian model permits exact VaR and ES attribution, including a direct comparison of Euler and Shapley allocations. A discrete independent-loss construction demonstrates that stand-alone 99% VaR can be zero for two components while aggregate VaR equals 100100, whereas correctly integrated ES retains coherent aggregation. In practical insurance portfolios, shared catastrophes, long-tail liabilities, reinsurance layers, recoverability, legal-entity segmentation, and model uncertainty determine the interpretation of the resulting contributions.

The strongest operational conclusion is that an insurer should preserve the distinction between aggregate capital, full-position allocated capital, marginal sensitivity, and finite incremental capital. It should also preserve the distinction between mathematical capital attribution and binding regulatory solvency requirements. The resulting research contribution is a checked synthesis of established risk-measure theory, exact derivations, illustrative computations, and explicitly bounded insurance applications. No original named doctrine, proprietary allocation model, or empirical novelty is asserted.

References