A decomposition of general premium principles into risk and deviation
aa r X i v : . [ q -f i n . R M ] J u l A DECOMPOSITION OF GENERAL PREMIUM PRINCIPLES INTORISK AND DEVIATION
MAX NENDEL, FRANK RIEDEL, AND MAREN DIANE SCHMECK
Abstract.
In this paper, we provide an axiomatic approach to general premiumprinciples giving rise to a decomposition into risk, as a generalization of the expectedvalue, and deviation, as a generalization of the variance. We show that, for every pre-mium principle, there exists a maximal risk measure capturing all risky componentscovered by the insurance prices. In a second step, we consider dual representations ofconvex risk measures consistent with the premium principle. In particular, we showthat the convex conjugate of the aforementioned maximal risk measure coincides withthe convex conjugate of the premium principle on the set of all finitely additive prob-ability measures. In a last step, we consider insurance prices in the presence of anot neccesarily frictionless market, where insurance claims are traded. In this setup,we discuss premium principles that are consistent with hedging using securizationproducts that are traded in the market.
Key words:
Principle of premium calculation, risk measure, deviation measure, con-vex duality, superhedging
AMS 2010 Subject Classification: Introduction
In classical risk theory, a premium principle is a map that assigns a real number H ( X ) to a random variable X . Here, H ( X ) is the premium for insuring the claim X ,see B¨uhlmann [3], Deprez and Gerber [5], Young [27], or, for textbook references, Rolskiet al. [22] and Kaas et al. [12]. In this approach, it is assumed that the probabilitydistribution of any loss is known. Frequently, however, the probability distributionis not known exactly. The issue of Knightian or model uncertainty has entered thecenter stage in recent years. The International Actuarial Association acknowledges theimportance of such uncertainty in Chapter 17 of the risk book [11]: ’ Risk is the effectof variation that results from the random nature of the outcomes being studied (i.e.,a quantity susceptible of measurement).
Uncertainty involves the degree of confidencein understanding the effect of perils or hazards not easily susceptible to measurement.’Model uncertainty is also widely recognised, for example, in the context of life insurance,cf. Biagini et al. [2], Bauer et al. [1], Milevsky et al. [17], and Schmeck and Schmidli[23].In this paper, we thus take a more general position, and model insurance claims asmeasurable functions, thus being closer to the actual real world contract. In particular,we do not assume ex ante that the probability distributions of losses are known to theinsurer. For a class C of bounded claims, we impose only two very natural conditionson premium principles. We require that there is no unjustified risk loading, i.e a shift Date : July 21, 2020.Financial support through the German Research Foundation via CRC 1283 is gratefully acknowledged.The authors thank Hans-Ulrich Gerber, Marcelo Brutti Righi, and Ruodu Wang for their helpfulcomments and remarks. of a loss by a known amount is priced correctly, or H ( X + m ) = H ( X ) + m for all X ∈ C and m ∈ R , (P1)compare Deprez and Gerber [5] and Young [27]. In the textbook Kaas et al. [12, Section5.3.1] Property (P1) is also referred to as a consistency condition. In the context ofmonetary risk measures, property (P1) is, up to a sign, usually referred to as cashadditivity , see e.g. F¨ollmer and Schied [8]. Our second natural requirement has theform H ( X ) ≥ H (0) = 0 for all X ∈ C with X ≥ . (P2)Condition (P2) simply states that an insurer will not be willing to pay money for insur-ing pure losses , i.e. claims with only positive outcomes, a property neccesary to avoid aruin with certainty. Since typically insurance claims have only positive outcomes, onecould, loosely speaking, interpret (P2) as a condition stating that insurance premia arealways nonnegative, a standard requirement, see e.g. Young [27]. Notice that, (P2) is,for example, implied by monotonicity.Our first main result shows that every insurance premium can be written as H ( X ) = R ( − X ) + D ( X ) for all X ∈ C, where R is a monetary risk measure (compare, e.g., F¨ollmer and Schied [8]) and D isa deviation measure (compare Rockafellar and Uryasev [21]) . Therefore, the simpleaxioms (P1) and (P2) immediately give a lot of structure and contain most knownexamples that are used in practice. We would like to point out that (P2) does notcontradict the standard no-ripoff condition, cf. Deprez and Gerber [5], Kaas et al. [12,Section 5.3.1], or Young [27], H ( X ) ≤ max X for all X ∈ C, (1.1)and that the latter can be added if necessary, leading to D = 0 in the decomposion of H (see Proposition 3.6). In the classic case, when the probability distribution is known,a typical insurance premium consists of the sum of the fair premium and a multiple ofthe variance or standard deviation, compare [22]. As the expected loss is a risk measureand the variance a deviation measure, we thus show that one can think of insurancepremia as generalizations of this basic approach in a very general way.It is natural to ask in what sense the risk and the deviation measure can be identifieduniquely. In general, this is not the case. However, we show that the premium principlecan be uniquely decomposed into a maximal risk measure R Max (capturing all riskycomponents of the insurance claim) and a minimal deviation measure D Min measuringthe claim’s pure fluctuations. Moreover, we show that R Max can be explicitly readout of the premium principle H and, additionally, can be extended to the space of all bounded random variables. That is, it is possible to explicitly filter the (maximal) riskcontribution to the premium principle H and price other insurance contracts (that arenot contained in C ) in a consistent way. The minimal deviation, i.e. the differencebetween the premium and the maximal risk measure can be seen as a margin forcompensating the parts of the claim that cannot be quantified as pure risk.We show that the classic premium principles as the aforementioned variance or stan-dard deviation principle or the well-known economic principles can be subsumed underour framework. We also discuss generalizations of these classic premium principles to We also refer to Liu et al. [14] for an overview on convex risk functionals, a class containing, both,risk and deviation measures and to Righi [20] for a detailed discussion on compositions between riskand deviation measures.
REMIUM PRINCIPLES, RISK, AND DEVIATION 3
Knightian uncertainty, and we discuss the more modern notions of quantile-based pre-mia involving Value at Risk or Expected Shortfall, cf. Rolski et al. [22, Section 3.1.3]and Kaas et al. [12, Section 5.6]. Similar to Castagnoli et al. [4], we discuss the case,where H ( X ) = E P ( X ) + Amb P ( X ) (1.2)with a fixed baseline model P ∈ P and whereAmb P ( X ) := 12 sup Q , Q ′ ∈P E Q ( X ) − E Q ′ ( X )measures the ambiguity of the model. Castagnoli et al. [4] consider premium principlesof the form (1.2) together with the no-ripoff condition (1.1), which as we will show,implies that the set of priors P is dominated by the reference measure P (see Proposition3.7).In a second step, we assume that, in addition to (P1) and (P2), the premium principle H is convex or sublinear. We then derive a dual representation of R Max in terms of theFenchel-Legendre transform of H . In the sublinear case, we show that there exists amaximal set P of probability measures (priors) satisfying H ( X ) ≥ E P ( X ) for all X ∈ C and P ∈ P . (1.3)The latter can be seen as a generalized version of a safety loading, see Castagnoli etal. [4] and Young [27]. In the case, where P = { P } consists of a single prior, one ends upwith the classical condition to avoid bankruptcy according to the principle of poolingrisk in a large group. We therefore see that, in the sublinear case, the notion a premiumprinciple H covers a certain amount of model uncertainty or ambiguity in terms of theprior. In view of equation (1.3), the set P can be seen as the set of all priors thatare covered by the premium principle in the sense that the premium principle avoidsbancrupcy under each model P ∈ P . We will therefore also refer to P as the set of allplausible models. In a last step, we discuss the relation of the maximal risk measureto superhedging in presence of a competitive market, that is used by the insurer tohedge against certain risks using portfolios or securization products that are traded inthe market. In the spirit of F¨ollmer and Schied [7], we derive equivalent conditionsensuring that the premium principle is consistent with superhedging.The paper is structured as follows. In Section 2, we introduce the setup and nota-tions, provide the decomposition of a premium principle into risk and deviation, givean explicit description of the maximal risk measure R Max , and discuss various examplesillustrating our notion of a premium principle. Section 3 is devoted to the study of con-vex and sublinear premium principles. In this context, we discuss dual representations,multiple priors, and baseline models. In Section 4, we address the connection betweenmarket consistency of insurance premia and hedging using securization products thatare traded in a competitive market. The proofs can be found in the Appendix A.2.
Premium principles and their decompositions
Model and Notation.
Let (Ω , F ) be a measurable space. Denote the spaceof all bounded, real-valued measurable functions by B b = B b (Ω , F ). Let C ⊂ B b represent the set of insurance claims covered by a premium policy. We assume that0 ∈ C and that X + m ∈ C for all X ∈ C and m ∈ R , where, in the notation, wedo not differentiate between real constants and constant functions (with real values). MAX NENDEL, FRANK RIEDEL, AND MAREN DIANE SCHMECK
Thus, we also consider claims with possibly negative values. We call every measurablefunction X ∈ B b a claim . We use the notationmax X := sup ω ∈ Ω X ( ω ) and min X := inf ω ∈ Ω X ( ω ) . We denote by ≤ , both, the usual order on the reals and the pointwise order on B b .2.2. Premium Principles and a Basic Decomposition.
The central object in ouranalysis is the following notion of a premium principle.
Definition 2.1.
We say that a map H : C → R is a premium principle on C if(P1) H ( X + m ) = H ( X ) + m for all X ∈ C and m ∈ R .(P2) H ( X ) ≥ H (0) = 0 for all X ∈ C with X ≥ H (0) = 0, implies that H ( m ) = m for all constantclaims leading to the common assumption of no unjustified risk loading , cf. Deprezand Gerber [5] and Young [27]. Concerning Property (P2), note that the condition H (0) = 0 is natural for insurance claims. A typical policy insures losses in the sensethat the claim is either zero or positive. (P2) ensures that the company or the markettake a nonnegative premium for sure damages. It is thus a minimal requirement for asensible notion of premium policy.Recall that a map R : B b → R is a (monetary) risk measure (see e.g. F¨ollmer andSchied [8]) if(R1) R (0) = 0 and R ( X + m ) = R ( X ) − m for all X ∈ B b and m ∈ R ,(R2) R ( X ) ≤ R ( Y ) for all X, Y ∈ B b with X ≥ Y .A map D : C → R is a deviation measure (cf. Rockafellar-Uryasev [21]) if(D1) D ( X + m ) = D ( X ) for all X ∈ C and m ∈ R ,(D2) D (0) = 0 and D ( X ) ≥ X ∈ C .Let R : B b → R be a risk measure and D : C → R be a deviation measure. Then,one readily verifies that the sum H ( X ) := R ( − X ) + D ( X ) , for X ∈ C, defines a premium principle on C . It is quite remarkable that this decomposition intoa monetary risk measure and a deviation measure characterizes all premium principles. Theorem 2.2.
A map H : C → R is a premium principle if and only if H ( X ) = R ( − X ) + D ( X ) for all X ∈ C, where R : B b → R is a risk measure and D : C → R is a deviation measure. The theorem shows that premium principles can be decomposed into a net premium or safety loading that takes care of the claim’s risk and a fluctuation loading that pricesthe variability of the damage. In the classic case when a prior probability distribution P is given, the typical premium consisting of the sum of the expected loss E P ( X ) and(a multiple of) the variance of X under P is a case in point. Note that the expectedloss is a risk measure and the variance a deviation measure.It is natural to ask in what sense the risk and the deviation measure can be identifieduniquely. The following theorem provides a partial answer in the sense that it decom-poses the premium principle into a maximal risk measure R Max (capturing all riskycomponents of the insurance claim) and a minimal deviation measure D Min measuringthe claim’s fluctuations that cannot be captured by any risk measure.
REMIUM PRINCIPLES, RISK, AND DEVIATION 5
Theorem 2.3.
Let H : C → R be a premium principle. Define R Max ( X ) := inf (cid:8) H ( X ) | X ∈ C, X + X ≥ (cid:9) , for X ∈ B b . The map R Max : B b → R defines a risk measure, and R Max ( − X ) ≤ H ( X ) for all X ∈ C . Moreover, D Min ( X ) := H ( X ) − R Max ( − X ) defines a deviation measure on C ,and H ( X ) = R Max ( − X ) + D Min ( X ) for all X ∈ C. For every other decomposition of the form H ( X ) = R ( − X ) + D ( X ) , for X ∈ C , witha risk measure R and a deviation measure D , we have R ≤ R Max and D ≥ D Min . Theorem 2.3 shows that one can identify uniquely a maximal risk measure and a minimal deviation measure whose sum forms the premium principle. The risk measure R Max solves a variational problem that is, at least in spirit, akin to the idea of super-hedging in finance, it computes the minimal premium that one has to pay for a claim X ∈ C that covers the loss given by X in every state of the world. Note that this riskmeasure is defined on the whole space of claims B b , the theorem thus provides a naturalextension of the premium principle H to the whole space of claims. In particular, weobtain an algorithm to extend a given premium principle to the set of all claims.2.3. Examples.
We illustrate how classic and new approaches of insurance pricing canbe subsumed under our framework.2.3.1.
Classic Premium Principles under a Given Probabilistic Model.
Example 2.4 (Ad hoc premium principles under a given model) . The benchmarkpremium principle is the fair premium principle given by H ( X ) = E P ( X ) , for X ∈ B b , where P is a fixed probability measure on (Ω , F ). Here, R Max = E P ( − · ) and D Min = 0.In practice, since the fair premium contains no premium for taking risk, insurers usuallyadd a safety loading, e.g. in terms of the variance H ( X ) = E P ( X ) + θ P ( X ) , for X ∈ B b , with a constant θ ≥
0. Here, R = E P ( − · ), and D = θ var P ( · ) is a decomposition of H into risk and deviation. However, as we will see in Example 3.4, for θ >
0, the maximalrisk measure R Max is given by R Max ( X ) = max Q ∈P E Q ( − X ) − θ G ( Q | P ) , where P consists of all probability measures Q , which are absolutely continuous w.r.t. P and satisfy G ( Q | P ) := var P (cid:18) d Q d P (cid:19) < ∞ .G is the so-called Gini concentration index , see e.g. Maccheroni et al. [15],[16].
Example 2.5 (Economic premium principles) . Let P be a probability measure on(Ω , F ) and ℓ : R → R be a strictly increasing loss function . One can then consider, forexample, the safety equivalent H ( X ) := ℓ − (cid:0) E P [ ℓ ( X − min X )] (cid:1) + min X, for X ∈ B b . Often, one considers a random initial endowment Z ∈ B b , which could be interpretedas an existing portfolio of insurance contracts. Assuming a continuously differentiable MAX NENDEL, FRANK RIEDEL, AND MAREN DIANE SCHMECK loss function ℓ , the premium p := H ( X ) is then computed by requiring that the newinsurance contract together with the premium p (infininitesimally) does not change theexpected loss. We thus have0 = lim h → E P (cid:0) ℓ [ Z + h ( X − p )] (cid:1) − E P (cid:0) ℓ ( Z ) (cid:1) h = E P (cid:0) ℓ ′ ( Z ) · ( X − p ) (cid:1) , which leads to the so-called Esscher transform (cf. [6] or, in the context of optionpricing, Gerber and Shiu [9]) H ( X ) := E P (cid:0) Xℓ ′ ( Z ) (cid:1) E P (cid:0) ℓ ′ ( Z ) (cid:1) , for X ∈ B b . Notice that, for P -a.s. constant Z , this leads to the mean value principle. In the caseof an exponential loss function ℓ ( x ) := α ( e αx −
1) with α >
0, this leads to H ( X ) := E P (cid:0) Xe αZ (cid:1) E P ( e αZ ) , for X ∈ B b , see also B¨uhlmann [3]. For Z = X , we obtain the celebrated Esscher principle (see e.g.B¨uhlmann [3] and Deprez and Gerber [5]) H ( X ) := E P (cid:0) Xe αX (cid:1) E P ( e αX ) , for X ∈ B b . Another condition, when considering a random endowment Z , is given by the followingmodification of the safety equivalent (cf. Deprez and Gerber [5]) E P (cid:0) ℓ ( Z + X − p ) (cid:1) = E P (cid:0) ℓ ( Z ) (cid:1) . For the exponential loss function ℓ ( x ) := α ( e αx −
1) with α >
0, this leads to H ( X ) = 1 α log E P (cid:0) e α ( Z + X ) (cid:1) E P (cid:0) e αZ (cid:1) , for X ∈ B b . Model Uncertainty.
The recent history brought the issue of model uncertaintyto center stage; in particular, it has become clear that working under the assumptionof a single probability distribution can be too optimistic for insurance companies. Newregulations thus ask insurers to take various models into account (stress testing).
Example 2.6 (Model uncertainty) . Instead of a single probability measure P on (Ω , F ),we now consider a nonempty set P of probability measures on (Ω , F ). The set P canbe seen as a set of plausible models, and we thus end up with a setup, where we havemodel uncertainty w.r.t. the models contained in P . Then, one can consider robustversions of the aforementioned premium principles by regarding worst case scenarios.Examples include:(i) A robust variance principle H ( X ) = sup P ∈P E P ( X ) + θ sup P ∈P var P ( X ) , for X ∈ B b , with θ ≥ H ( X ) := sup P ∈P E P (cid:0) Xe αX (cid:1) E P ( e αX ) , for X ∈ B b , REMIUM PRINCIPLES, RISK, AND DEVIATION 7 or a robust safety equivalent with exponential utility function and random endow-ment Z ∈ B b H ( X ) = sup P ∈P (cid:18) α log E P (cid:0) e α ( Z + X ) (cid:1) E P (cid:0) e αZ (cid:1) (cid:19) , for X ∈ B b , for α > H ( X ) := sup P ∈P E P (cid:0) ℓ ( X − min X ) (cid:1) + min X, for X ∈ B b , with a nondecreasing loss function ℓ : R → R .(iv) Alternatively and particularly in the case of parameter uncertainty, one can con-sider, for a Polish space Ω (endowed with the Borel σ -algebra F ), a probabilitymeasure µ : Σ → [0 ,
1] (second-order prior), where Σ = Σ( P ) denotes the Borel σ -algebra on P endowed with the vague topology, and take a a mean value w.r.t. µ . In the simplest case, where ℓ ( x ) = φ ( x ) = x , this corresponds to a Bayesianprediction. This approach can be modified by considering a continuous nonde-creasing loss function ℓ : R → R and another nondecreasing function φ : R → R (second-order loss function). Then, one obtains the so-called smooth ambiguitymodel (cf. Klibanoff et al. [13]) H ( X ) := Z P φ (cid:0) E P (cid:2) ℓ ( X − min X ) (cid:3)(cid:1) µ (d P ) + min X, for X ∈ C b , where C b denotes the space of all continuous and bounded functions Ω → R . Example 2.7 (Ambiguity indices) . Again, we consider a nonempty set P of probabilitymeasures on (Ω , F ). In contrast to the previous example, we now fix a baseline model P ∈ P , which can be seen as the (due to some case-dependent reasons) most plausiblemodel. The idea is now to consider a safety loading, where we differentiate between riskand uncertainty. The risk premium can then be given, for example, by the variance orthe (average) value at risk, and the premium for uncertainty is given by (cf. Castagnoliet al. [4]) Amb P ( X ) := 12 sup Q , Q ′ ∈P E Q ( X ) − E Q ′ ( X ) , for X ∈ B b . Then, Amb P as an uncertainty premium together with the variance as a risk premiumleads to the premium principle H ( X ) = E P ( X ) + θ P ( X ) + γ Amb P ( X ) , for X ∈ B b , (2.1)with γ, θ ≥
0. We would like to point out that, in the setup chosen by Castagnoliet al. [4], it is only possible to consider premium principles with an ambiguity indexfor sets P , which are symmetric to the baseline model P , i.e. if 2 P − Q ∈ P for all Q ∈ P . However, our notion of a premium principle includes premium principles basedon ambiguity indices for any nonempty set P and every choice of the baseline model.Instead of Amb P in (2.1), one could, for example, also consider the ambiguity indexAmb ′P ( X ) := 12 sup Q , Q ′ ∈P d W (cid:0) Q ◦ X − , Q ′ ◦ X − (cid:1) , where d W denotes the Wasserstein distance. MAX NENDEL, FRANK RIEDEL, AND MAREN DIANE SCHMECK
Example 2.8 (Quantile-based premium principles) . Let ε ∈ (0 , P be a probabilitymeasure on (Ω , F ), and P − X ( λ ) := inf { a ∈ R | P ( X ≤ a ) ≥ λ } , for X ∈ B b and λ ∈ (0 , . Then, we could consider the ε -quantile principle , cf. Rolski et al. [22, Section 3.1.3] orKaas et al. [12, Section 5.6], H ( X ) = V@R ε P ( − X ) = P − X (1 − ε ) , for X ∈ B b , as a possible premium principle, where V@R ε P is also known as the value at risk under P at level ε , cf. F¨ollmer and Schied [8]. Here R = V@R ε P ( · ) and D = 0. A majordrawback, when considering the value at risk is that it is typically not convex andthus does not reflect diversification effects. Therefore, one often considers the expectedshortfall or average value at risk AV@R ε P at level ε , given byAV@R ε P ( X ) := 1 ε Z ε V@R γ P ( X ) d γ, for X ∈ B b , instead of V@R ε P . AV@R ε P is convex and positive homogeneous (of degree 1), cf. F¨ollmerand Schied [8]. Alternatively, for θ ≥
0, one can consider the so-called absolute deviationprinciple , cf. Rolski et al. [22, Section 3.1.3], H ( X ) = E P ( X ) + θ E P (cid:0)(cid:12)(cid:12) X − P − X (cid:0) (cid:1)(cid:12)(cid:12)(cid:1) , for X ∈ B b , as a modification of the standard deviation principle. In this case, R ( X ) = E P ( − X )and D ( X ) = θ E P (cid:0)(cid:12)(cid:12) X − P − X (1 / (cid:12)(cid:12)(cid:1) for X ∈ B b . Notice that D ( X ) = θ (cid:18) AV@R P ( − X ) + AV@R P ( X ) (cid:19) = θ Amb Q ( X )is (up to a constant) an ambiguity index, where Q consists of all probability measures Q ≪ P whose density d Q d P is P -a.s. bounded by 2, cf. Example 3.5. In Example 3.5, wefurther show that, for θ ≥
1, the maximal risk measure R Max is given by R Max ( X ) = AV@R θ P ( X ) , for X ∈ B b . Example 2.9 (Choquet integrals) . Wang, Young, and Panjer [26] derive an axiomaticcharacterization of premium principles in a competitive market setting that results ina representation using Choquet integrals. Consider a set function γ : F → [0 ,
1] with γ ( ∅ ) = 0, γ (Ω) = 1, and γ ( A ) ≤ γ ( B ) for all A, B ∈ F with A ⊂ B . Then, we considerthe premium principle given by the Choquet integral w.r.t. γH ( X ) := Z ∞ min X γ (cid:0) { X > t } (cid:1) d t + min X for X ∈ B b . Wang, Young, and Panjer [26] show that, under certain axioms, every premium principle H can be represented as a Choquet integral w.r.t. a distorted probability γ = g ◦ P fora probability measure P and a distortion function g (a nondecreasing function on [0 , g (0) = 0 and g (1) = 1). In this case, H ( X ) = Z ∞ min X g (cid:0) P X ( t ) (cid:1) d t + min X, where, for t ≥ P X ( t ) := P ( X > t ). REMIUM PRINCIPLES, RISK, AND DEVIATION 9 Dual representation of convex premium principles and baselinemodels
Premium principles should generally reflect the benefits of diversification and theaversion to uncertainty. In this section, we thus consider convex premium principles,generalizing the approach of [5] who assume that the probability distribution of claimsis known. We identify the maximal risk measure in the premium’s decomposition as aconvex risk measure, cf. F¨ollmer and Schied [7]. Throughout this section, we assumethat C is a linear space with R ⊂ C . We denote the set of all finitely additive probabilitymeasures on (Ω , F ) by ba . We say that a premium principle H : C → R is convex if H ( λX + (1 − λ ) Y ) ≤ λH ( X ) + (1 − λ ) H ( Y ) for all λ ∈ [0 ,
1] and
X, Y ∈ C. In this case, we denote the convex dual of H by H ∗ ( P ) := inf X ∈ C E P ( − X ) + H ( X ) ∈ [ −∞ ,
0] for P ∈ ba . Theorem 3.1.
Let H : C → R a convex premium principle. Then, the maximal riskmeasure R Max in the decomposition of H satisfies R Max ( X ) = max P ∈ ba E P ( − X ) + H ∗ ( P ) for all X ∈ B b . Moreover, H ∗ ( P ) = inf X ∈ B b E P ( X ) + R Max ( X ) for all P ∈ ba , (3.1)By the previous theorem, the convex dual H ∗ of the premium principle correspondsto the penalty function of its maximal risk measure. H ∗ thus represents the confidencethat the insurer puts on a particular model P within the class of all possible models.In the sequel, we will refer to P := (cid:8) P ∈ ba | H ∗ ( P ) < ∞ (cid:9) as the set of all plausible models .If the premium principle is also scalable in the sense that it is positively homogeneous, R Max is even a coherent risk measure.
Corollary 3.2.
Let H : C → R be a sublinear premium principle, i.e. H is a con-vex premium principle, and H ( λX ) = λH ( X ) for all X ∈ C and λ > . Then, therepresenting maximal risk measure R Max is a coherent risk measure, i.e. R Max ( X ) = max P ∈P E P ( − X ) for all X ∈ B b , where the set of plausible models is given by P = (cid:8) P ∈ ba | ∀ X ∈ C : E P ( X ) ≤ H ( X ) (cid:9) . Proof.
This follows directly from Theorem 3.1 together with the observation that sub-linearity implies that H ∗ ( P ) ∈ {−∞ , } for all P ∈ ba . (cid:3) Notice that, in the sublinear case, for all probabilistic models P ∈ P and all claims X ∈ B b , H ( X ) ≥ E P ( X ) . In other words, the premium principle H incorporates a so-called safety loading undereach plausible model P ∈ P , cf. Castagnoli et al. [4], Young [27], and Deprez and Gerber [5]. In the next step, we analyze in more detail the minimal deviation measure of thepremium’s decomposition. For P ∈ P , D P ( X ) := H (cid:0) X − E P ( X ) (cid:1) , for X ∈ C, (3.2)defines a deviation measure, that we call the fluctuation loading under P . We have H ( X ) = E P ( X ) + D P ( X ) for all X ∈ C. (3.3)The deviation measures D P can thus be seen as the profit for accepting the aleatoricrisk of X under the model P ∈ P . Equation (3.3) is a model-dependent decompositioninto a risk measure and a deviation measure. We have the following relation betweenthe minimal deviation measure D Min and the family of model-dependent deviationmeasures ( D P ) P ∈P . Corollary 3.3.
Let H : C → R be a convex premium principle. Then, D Min ( X ) = min P ∈P D P ( X ) − H ∗ ( P ) for all X ∈ C. Proof.
By Theorem 3.1, D Min ( X ) = H ( X ) − R Max ( − X ) = min P ∈P H (cid:0) X − E P ( X ) (cid:1) − H ∗ ( P ) for all X ∈ C. The statement now follows from Equation (3.2). (cid:3)
Example 3.4.
Let P be a probability measure, θ >
0, and consider H ( X ) := E P ( X ) + θ P ( X ) for all X ∈ B b . Let Q ∈ P . The condition H ∗ ( Q ) < ∞ imples that Q is countably additive andabsolutely continuous w.r.t. P . Let Z := d Q d P . We can write H ∗ ( Q ) = inf X ∈ B b E Q ( − X ) + H ( X ) = inf X ∈ B b E P (cid:20) X (cid:18) − Z + θ (cid:0) X − E P ( X ) (cid:1)(cid:19)(cid:21) With the help of Cauchy-Schwarz inequality, one can then show that H ∗ ( Q ) = inf (cid:26) α E P (cid:0) X (cid:1) (cid:12)(cid:12)(cid:12)(cid:12) α ≤ , αX = 1 − Z + θ (cid:0) X − E P ( X ) (cid:1)(cid:27) , compare the Appendix of [16] for details. Notice that αX = 1 − Z + θ (cid:0) X − E P ( X ) (cid:1) implies that E P ( X ) = 0, which, in turn, implies that X = (cid:0) α − θ (cid:1) − (1 − Z ). Hence, H ∗ ( Q ) = inf α ≤ α (cid:18) α − θ (cid:19) − var P (cid:18) d Q d P (cid:19) . Notice that dd α α (cid:0) α − θ (cid:1) − = 0 if and only if α = − θ . We therefore obtain that H ∗ ( Q ) = − θ var P (cid:18) d Q d P (cid:19) . is (up to the factor − θ ) the Gini concentration index . By Theorem 3.1, R Max ( X ) = max Q ∈P E Q ( − X ) − θ var P (cid:18) d Q d P (cid:19) for all X ∈ B b . This follows from the inequality R Max ( − X ) ≤ H ( X ), for X ∈ B b , together with [8, Proposition 4.21]. REMIUM PRINCIPLES, RISK, AND DEVIATION 11
Example 3.5.
Let P be a probability measure, θ ≥
0, and consider H ( X ) = E P ( X ) + θ E P (cid:0)(cid:12)(cid:12) X − P − X (cid:0) (cid:1)(cid:12)(cid:12)(cid:1) , for X ∈ B b . Then, by [8, Lemma 4.46], E P (cid:0)(cid:12)(cid:12) X − P − X (cid:0) (cid:1)(cid:12)(cid:12)(cid:1) = E P (cid:16)(cid:0) X − P − X (cid:0) (cid:1)(cid:1) − (cid:17) + E P (cid:16)(cid:0) X − P − X (cid:0) (cid:1)(cid:1) + (cid:17) = 12 (cid:0) AV@R P ( − X ) + AV@R P ( X ) (cid:1) Recall that, for ε ∈ (0 , ε P ( X ) = max Q ∈Q /ε E Q ( − X ) for all X ∈ B b , where, for α ≥ Q α denotes the set of all probability measures Q ≪ P whose densityis P -a.s. bounded by α , cf. [8, Theorem 4.47]. Therefore, the set P related to R Max isgiven by the set of all probability measures Q ∗ of the form Q ∗ = P + θ (cid:0) Q − Q ′ (cid:1) (3.4)with Q , Q ′ ∈ Q . Therefore, P consists of all probability measures Q ∗ ≪ P with1 − θ ≤ d Q ∗ d P ≤ θ P -a.s. (3.5)In particular, for θ ≥ P = Q θ , which implies that R Max ( X ) = AV@R θ P ( X ) for all X ∈ B b . In fact, by the previous argumentation, it follows that every Q ∗ ∈ P is of the form (3.4),which, in turn, implies that it satisfies (3.5). Now, assume that Q ∗ ≪ P is a probabilitymeasure, which satisfies (3.5). For θ = 0, it follows that Q ∗ = P ∈ P . Hence, assumethat θ >
0, and define Z := 2 θ (cid:18) d Q ∗ d P − (cid:19) . Then, | Z | ≤ P -a.s., E P ( Z ) = 0, and, by H¨older’s inequality, α := E P ( | Z | )2 ≤
1. Define Y := Z + + 1 − | Z | Y ′ := Z − + 1 − | Z | . Then, 0 ≤ Y ≤ ≤ Y ′ ≤ P -a.s. Moreover, Y − Y ′ = Z P -a.s. and E P ( Y ) = E P ( Z + ) + 1 − E P ( | Z | )2 = 1 = E P ( Z − ) + 1 − E P ( | Z | )2 = E P ( Y ′ ) . Hence, Equation (3.4) is satisfied with Q := Y d P and Q ′ := Y ′ d P We say that a premium principle H : C → R is monotone if H ( X ) ≤ H ( Y ) forall X, Y ∈ C with X ≤ Y . Throughout the remainder of this section, we discussthe relation to monotone sublinear premium principles that Castagnoli et al. considerin [4]. More precisely, we show that, in the convex case, replacing Axiom (P2) in thedefinition of a premium principle by a so-called internality condition , cf. [4], implies themonotonicity of the premium principle, and thus, together with positive homogeneity,leads to the objects considered in [4]. A similar result can be found in Deprez andGerber [5, Theorem 3]. Proposition 3.6.
Let H : C → R a convex map that satisfies (P1). Then, the followingstatements are equivalent: (i) H is a monotone premium principle,(ii) H ( X ) ≤ H (0) = 0 for all X ∈ C with X ≤ , i.e. H is internal. Notice that (ii) together with (P1) implies the standard no-ripoff condition (1.1).The following proposition is a partial extension of Theorem 3 in Castagnoli et al. [4],where statement (i) is a reformulation of Axiom P.7 in [4].
Proposition 3.7.
Let H : C → R be a sublinear premium principle, and define Amb P ( X ) := 12 (cid:0) R Max ( − X ) + R Max ( X ) (cid:1) = 12 max Q ,Q ′ ∈P E Q ( X ) − E Q ′ ( X ) for X ∈ B b . Then, for every P ∈ P , the following statements are equivalent:(i) E P ( X ) = (cid:0) R Max ( − X ) − R Max ( X ) (cid:1) for all X ∈ B b ,(ii) P is symmetric with center P , i.e. P − Q ∈ P for all Q ∈ P ,(iii) Amb P ( X ) = max Q ∈P E P ( X ) − E Q ( X ) ,(iv) D P ( X ) = D Min ( X ) + Amb P ( X ) for all X ∈ C .In this case, R Max is dominated by P , i.e. every Q ∈ P is absolutely continuous w.r.t. P ,and P is countably additive if and only if every Q ∈ P is countably additive.Remark . We would like to point out what are the implications of Proposition 3.7 inview of [4]. Notice that, in contrast to assumption (P2) in our definition of a premiumprinciple (Definition 2.1), the internality axiom P.1 together with the subadditivityrequirement P.3 in [4] already implies the monotonicity of the premium functional H .Therefore, Proposition 3.7 shows that in [4] only very particular types of ambiguity sets P , namely symmetric ones, and only a very particular choice of the baseline model P canbe considered for the premium calculation. Consequently, the case of an asymmetric P does not fall into the setup chosen in [4]. Moreover, the symmetry of P implies that allelements of P are absolutely continuous w.r.t. the baseline model P excluding setups,where the set P is undominated. However undominated sets of plausible models appearquite naturally, for example, when considering a Brownian Motion with uncertainty inthe volatility parameter, see e.g. Peng [18],[19] or Soner et al. [24],[25]. Hence, replacingthe internality axiom P.1 in [4] by the apparently similar assumption (P2) in Definition2.1 has a huge impact. In the following example, we describe a basic setup that leadsto a nonsymmetric set P . Example 3.9.
Consider the setup (2.1) from [4] with Ω = N , endowed with the σ -algebra F = 2 N (power set). For n ∈ N , we consider the measure P n := 1 n n X k =1 δ k , where δ k denotes the Dirac measure with barycenter k ∈ N . We then consider themonotone premium principle H ( X ) := sup n ∈ N E P n ( X ) = sup n ∈ N n n X k =1 X ( k ) , for X ∈ B b . One readily verifies that the set P consists only of probability measures P of the form P = X n ∈ N λ n δ n (3.6)with a nonincreasing sequence ( λ n ) n ∈ N ⊂ [0 ,
1] summing up to 1. Assume that thereexisted some P ∈ P with 2 P − P n ∈ P for all n ∈ N . Then, P is of the form (3.6) with a REMIUM PRINCIPLES, RISK, AND DEVIATION 13 nonincreasing sequence ( λ n ) n ∈ N ⊂ [0 , P − P n ∈ P for all n ∈ N ,which, in particular, means that2 λ n − n ≥ λ n +1 for all n ∈ N . (3.7)However, Equation (3.7) implies that λ n +1 = λ + n X k =1 (cid:0) λ k +1 − λ k (cid:1) ≤ λ − n X k =1 k → −∞ , as n → ∞ , leading to a contradiction. By Proposition 3.7, we may therefore conclude that thereexists no P ∈ P with H ( X ) = E P ( X ) + Amb P ( X ) for all X ∈ B b . That is, the left right-hand side of the previous equation does not define a premiumprinciple in the sense of Castagnoli et al. [4], whereas it defines a premium principle inthe sense of Definition 2.1.4.
Superhedging and market consistency
The integration of insurance and finance has been a central issue of research in thelast years. In this section, we consider insurance premia that are consistent with agiven financial market (or liquidly traded insurance contracts). We will identify themaximal risk measure in the premium’s decomposition as the so-called superhedgingrisk measure.The financial market is modeled by a linear subspace M ⊂ C , where C is againassumed to be a linear space, and a nonnegative linear price functional F : M → R .Assuming M to be a linear space and F : M → R to be linear corresponds to a com-petitive market without frictions. We would like to point out that our model can alsobe used for markets with frictions. That is, the linearity of the price functional F canbe replaced by sublinearity, and M can be assumed to be a convex cone instead of alinear subspace. In this case, F would resemble the ask price for securization productsthat are traded in the market or, in other words, the price an insurer has to pay for“selling” the risk of a claim to the market. Nonnegativity is a no arbitrage conditionas it requires F ( X ) ≥ X ≥
0. Without loss of generality, we assume that F (1) = 1,i.e. the interest rate that is implicit in F is zero. We call M = (cid:8) P ∈ ba (cid:12)(cid:12) ∀ X ∈ M : E P ( X ) = F ( X ) (cid:9) the set of martingale measures for the financial market.Throughout this section, we consider a sublinear premium principle. We assume thatthe premia charged by our insurer coincide with market prices on M , i.e. H | M = F .The condition H | M = F expresses the fact that the insurer cannot charge a premiumabove market prices due to competition. We introduce the set M := (cid:8) X ∈ M | F ( X ) = 0 (cid:9) of all claims that are traded on the market with price 0. In the sequel, we consider thesuperhedging risk measure R ∗ ( X ) := inf (cid:8) m ∈ R (cid:12)(cid:12) ∃ X ∈ M : m + X + X ≥ (cid:9) for all X ∈ B b . The superhedging risk measure amounts to the cost of staying on the safe side with thehelp of the products that are already being traded liquidly in the market. Notice that R ∗ is well-defined, since M is nonempty. Moreover, R Max ≤ R ∗ since H | M = F . Proposition 4.1.
Let H be sublinear. Then, the following statements are equivalent:(i) The maximal risk measure in the decomposition of H is the superhedging func-tional R ∗ , i.e. R Max = R ∗ .(ii) The premium principle H is based on the use of securization products, i.e., forall X ∈ C , there exists some X ∈ M with X ≤ X and H ( X ) ≥ F ( X ) .(iii) The plausible models for H coincide with the martingale measures, i.e. P = M . Appendix A. Proofs
Proof of Theorem 2.3.
First, notice that R Max : B b → R is well-defined since max( − X ) ∈ C with X +max( − X ) ≥ H ( X ) ≥ H (cid:0) min( − X ) (cid:1) for all X ∈ C with X + X ≥ R Max ( − X ) ≤ H ( X ) for all X ∈ C . Hence, D Min ( X ) = H ( X ) − R Max ( − X ) ≥ X ∈ C . Moreover, H ( X ) ≥ H (0) = 0 for all X ∈ C with X ≥
0, which implies that R Max (0) = 0. In particular, D Min (0) = H (0) − R Max (0) = 0. Wewill now show that R Max defines a risk measure. First, observe that R Max ( X ) ≤ H ( Y )for X, Y ∈ B b with X ≥ Y and Y ∈ C with Y + Y ≥
0. Taking the infimum over all Y ∈ C with Y + Y ≥
0, it follows that R Max ( X ) ≤ R Max ( Y ). Now, let X ∈ B b , m ∈ R and X ∈ C with X + X ≥
0. Then, R Max ( X + m ) ≤ H ( X − m ) = H ( X ) − m. Taking the infimum over all X ∈ C with X + X ≥ R Max ( X + m ) ≤ R Max ( X ) − m . On the other hand, R Max ( X ) − m = R Max ( X + m − m ) − m ≤ R Max ( X + m ) . This also shows that, for X ∈ C and m ∈ R , D Min ( X + m ) = H ( X + m ) − R Max ( − X − m ) = H ( X ) − R Max ( − X ) = D Min ( X ) . Let R : B b → R be a risk measure with R ( − X ) ≤ H ( X ) for all X ∈ C . Then, for all X ∈ B b and X ∈ C with X + X ≥ R ( X ) ≤ R ( − X ) ≤ H ( X ) . Taking the infimum over all X ∈ C with X + X ≥
0, we may conclude that R ( X ) ≤ R Max ( X ) for all X ∈ B b . (cid:3) Proof of Theorem 3.1.
We first show that R Max : B b → R is convex. Let X, Y ∈ B b and λ ∈ [0 , X , Y ∈ C with X ≤ X and Y ≤ Y , R Max (cid:0) λX + (1 − λ ) Y (cid:1) ≤ H (cid:0) λX + (1 − λ ) Y (cid:1) ≤ λH ( X ) + (1 − λ ) H ( Y ) . Taking the infimum over all X , Y ∈ C with X + X ≥ Y + Y ≥
0, we obtain that R Max is convex. Since R Max is a convex risk measure, it follows that, see e.g. F¨ollmerand Schied [8, Theorem 4.12], R Max ( X ) = max P ∈ ba E P ( − X ) + R ∗ Max ( P ) for all X ∈ B b , where R ∗ Max ( P ) := inf X ∈ B b E P ( X ) + R Max ( X ) for P ∈ ba . It remains to show (3.1),i.e. H ∗ ( P ) = R ∗ Max ( P ) for all P ∈ ba . Since R Max ( − X ) ≤ H ( X ) for all X ∈ C , itfollows that R ∗ Max ( P ) ≤ inf X ∈ C E P ( − X ) − R Max ( − X ) ≤ H ∗ ( P ) for all P ∈ ba . REMIUM PRINCIPLES, RISK, AND DEVIATION 15
In particular, there exists some P ∈ ba with H ∗ ( P ) > −∞ . Therefore, R ( X ) := sup P ∈ ba E P ( − X ) + H ∗ ( P ) , for X ∈ B b , defines a risk measure. Since R ∗ Max ( P ) ≤ H ∗ ( P ) for all P ∈ ba , it follows that R Max ( X ) ≤ R ( X ) ≤ H ( X ) for all X ∈ C. By the maximality of R Max , we may conclude that R Max = R . In particular, H ∗ ( P ) ≤ E P ( X ) + R ( X ) = E P ( X ) + R Max ( X ) for all X ∈ C and P ∈ ba . By definition of R ∗ Max , it follows that H ∗ ( P ) ≤ R ∗ Max ( P ) for all P ∈ ba . (cid:3) Proof of Proposition 3.6.
Trivially, (i) implies (ii). We first show that (ii) implies (P2),let X ∈ C with X ≥
0. Then, by Condition (ii),0 ≤ − H ( − X ) ≤ H ( X ) , where the second inequality follows from 0 = 2 H (0) ≤ H ( X ) + H ( − X ). In order toprove the monotonicity, first notice that, due to (P1) and (ii), H ( X ) = H ( X − max X ) + max X ≤ max X for all X ∈ C. (A.1)Now, let X, Y ∈ C with X ≤ Y . Then, by (A.1), for all λ ∈ (0 , H ( X ) ≤ λH ( Y ) + (1 − λ ) H (cid:18) X − λY − λ (cid:19) ≤ λH ( Y ) + max( X − λY ) ≤ λH ( Y ) + (1 − λ ) max X. Letting λ →
1, we obtain that H ( X ) ≤ H ( Y ). (cid:3) Proof of Proposition 3.7.
For all X ∈ B b ,Amb P ( X ) = R Max ( X ) + 12 (cid:0) R Max ( − X ) − R Max ( X ) (cid:1) and max Q ∈P E P ( X ) − E Q ( X ) = R Max ( X ) + E P ( X ) . Therefore, Amb P ( X ) = max Q ∈P E P ( X ) − E Q ( X ) for all X ∈ B b if and only if E P ( X ) = (cid:0) R Max ( − X ) − R Max ( X ) (cid:1) for all X ∈ B b . On the other hand, if E P ( X ) = (cid:0) R Max ( − X ) − R Max ( X ) (cid:1) for all X ∈ B b , then, for all X ∈ B b and Q ∈ P ,2 E P ( − X ) − E Q ( − X ) ≤ E P ( − X ) + R Max ( − X ) = R Max ( X ) , i.e. 2 P − Q ∈ P . Next, assume that 2 P − Q ∈ P for all Q ∈ P . Then, for all X ∈ B b ,12 (cid:0) R Max ( − X ) − R Max ( X ) (cid:1) = 12 (cid:0) max Q ∈P E Q ( X ) + min Q ′ ∈P E Q ′ ( X ) (cid:1) ≤
12 max Q ∈P (cid:0) E Q ( X ) + (2 E P ( X ) − E Q ( X ) (cid:1) = E P ( X ) . Using a symmetry argument, this implies that (cid:0) R Max ( − X ) − R Max ( X ) (cid:1) = E P ( X )for all X ∈ B b . We have therefore established the equivalence (i) - (iii). In order toestablish the remaining equivalence, first observe that, for all P ∈ P , E P ( X ) + D P ( X ) = H ( X ) = R Max ( − X ) + D Min ( X )= 12 (cid:0) R Max ( − X ) − R Max ( X ) (cid:1) + D Min ( X ) + Amb P ( X ) The equivalence between (i) and (iv) in now an immediate consequence of the previousequation. Under (ii), it follows that E Q ( X ) ≤ E P ( X ) for all X ∈ B b with X ≥ Q ∈ P . Choosing X = 1 N for N ∈ F with P ( N ) = 0, it follows that every Q ∈ P isabsolutely continuous w.r.t. P . On the other hand, let Q ∈ P and ( X n ) n ∈ N ⊂ B b with X n +1 ≤ X n for all n ∈ N and inf n ∈ N X n = 0. If P is countably additive, then0 ≤ E Q ( X n ) ≤ E P ( X n ) → n → ∞ , which shows that Q is countably additive. (cid:3) Proof of Proposition 4.1.
First observe that, since H | M = FR ∗ ( X ) = inf (cid:8) H ( X ) (cid:12)(cid:12) X ∈ M, X + X ≥ (cid:9) for all X ∈ B b . Therefore, by Theorem 3.1, statement (i) is equivalent to statement (iii). It remains toshow the equivalence of (i) and (ii). In order to establish this equivalence, first assumethat R Max = R ∗ . Then, R ∗ ( − X ) ≤ H ( X ) for all X ∈ C . By definition of R ∗ , this meansthat, for all X ∈ C , there exists some X ∈ M with X ≤ X and F ( X ) ≤ H ( X ).Now, assume that, for all X ∈ C , there exists some X ∈ M with X ≤ X and F ( X ) ≤ H ( X ). Then, by definition of R ∗ , it follows that R ∗ ( − X ) ≤ H ( X ) for all X ∈ C . The maximality of R Max together with R Max ≤ R ∗ , which is due to H | M = F ,thus implies that R Max = R ∗ . (cid:3) References [1] D. Bauer, M. B¨orger, and J. Ruß. On the pricing of longevity-linked securities.
Insurance Math.Econom. , 46(1):139–149, 2010.[2] F. Biagini, T. Rheinl¨ander, and J. Widenmann. Hedging mortality claims with longevity bonds.
Astin Bull. , 43(2):123–157, 2013.[3] H. B¨uhlmann. An economic premium principle.
Astin Bull. , 11(1):52–60, 1980/81.[4] E. Castagnoli, F. Maccheroni, and M. Marinacci. Insurance premia consistent with the market.
Insurance Math. Econom. , 31(2):267–284, 2002.[5] O. Deprez and H. U. Gerber. On convex principles of premium calculation.
Insurance Math.Econom. , 4(3):179–189, 1985.[6] F. Esscher. On the probability function in the collective theory of risk.
Skand. Aktuarietidskr. ,1932(3):175–195, 1932.[7] H. F¨ollmer and A. Schied. Convex measures of risk and trading constraints.
Finance Stoch. ,6(4):429–447, 2002.[8] H. F¨ollmer and A. Schied.
Stochastic finance . De Gruyter Graduate. De Gruyter, Berlin, 2016. Anintroduction in discrete time, Fourth revised and extended edition of [ MR1925197].[9] H. U. Gerber and E. S. Shiu.
Option pricing by Esscher transforms .[10] I. Gilboa and D. Schmeidler. Maxmin expected utility with nonunique prior.
J. Math. Econom. ,18(2):141–153, 1989.[11] S. Gutterman.
Chapter 17 - Risk and Uncertainty . IAA Risk Book. 2017.[12] R. Kaas, M. Goovaerts, J. Dhaene, and M. Denuit.
Modern Actuarial Risk Theory . Springer, Berlinand Heidelberg, 2008.[13] P. Klibanoff, M. Marinacci, and S. Mukerji. A smooth model of decision making under ambiguity.
Econometrica , 73(6):1849–1892, 2005.[14] F. Liu, J. Cai, C. Lemieux, and R. Wang. Convex risk functionals: representation and applications.
Insurance Math. Econom. , 90:66–79, 2020.[15] F. Maccheroni, M. Marinacci, and A. Rustichini. Ambiguity aversion, robustness, and the varia-tional representation of preferences.
Econometrica , 74(6):1447–1498, 2006.[16] F. Maccheroni, M. Marinacci, A. Rustichini, and M. Taboga. Portfolio selection with monotonemean-variance preferences.
Math. Finance , 19(3):487–521, 2009.[17] M. A. Milevsky, S. D. Promislow, and V. R. Young. Killing the law of large numbers: Mortalityrisk premiums and the sharpe ratio.
J. Risk Insurance .[18] S. Peng. G -expectation, G -Brownian motion and related stochastic calculus of Itˆo type. In Sto-chastic analysis and applications , volume 2 of
Abel Symp. , pages 541–567. Springer, Berlin, 2007.
REMIUM PRINCIPLES, RISK, AND DEVIATION 17 [19] S. Peng. Multi-dimensional G -Brownian motion and related stochastic calculus under G -expectation. Stochastic Process. Appl. , 118(12):2223–2253, 2008.[20] M. B. Righi. A composition between risk and deviation measures.
Ann. Oper. Res. , 282(1-2):299–313, 2019.[21] R. T. Rockafellar and S. Uryasev. The fundamental risk quadrangle in risk management, optimiza-tion and statistical estimation.
Surv. Oper. Res. Manag. Sci. , 18(1-2):33–53, 2013.[22] T. Rolski, H. Schmidli, V. Schmidt, and J. Teugels.
Stochastic processes for insurance and finance .Wiley Series in Probability and Statistics. John Wiley & Sons, Ltd., Chichester, 1999.[23] M. D. Schmeck and H. Schmidli. Mortality options: the point of view of an insurer.
Center forMathematical Economics Working Papers; 616 , 2019.[24] H. M. Soner, N. Touzi, and J. Zhang. Martingale representation theorem for the G -expectation. Stochastic Process. Appl. , 121(2):265–287, 2011.[25] H. M. Soner, N. Touzi, and J. Zhang. Quasi-sure stochastic analysis through aggregation.
Electron.J. Probab. , 16:no. 67, 1844–1879, 2011.[26] S. S. Wang, V. R. Young, and H. H. Panjer. Axiomatic characterization of insurance prices.
Insurance Math. Econom. , 21(2):173–183, 1997.[27] V. R. Young.
Premium Principles . Encyclopedia of Actuarial Science, 2006.
Center for Mathematical Economics, Bielefeld University
E-mail address : [email protected] Center for Mathematical Economics, Bielefeld University
E-mail address : [email protected] Center for Mathematical Economics, Bielefeld University
E-mail address ::