1. Introduction
Reinsurance is a transaction whereby one insurance company (the reinsurer) agrees to indemnify another insurance company (the reinsured, cedent or primary company) against all or part of the loss that the latter sustains under a policy or policies that it has issued. For this service, the ceding company pays the reinsurer a premium, and there are many premium calculation principles (e.g., [
1,
2]; see also [
3,
4]).
Mathematically, let X be the loss for an insurer from a policy or a group of policies. Assume that under a reinsurance treaty, a reinsurer covers the ceded part of the loss, say , where , for a premium . The primary insurer’s retained loss is denoted by . Commonly-used forms of reinsurance treaties are the excess-of-loss treaty, where  with deductible level (attaching point) ; and the quota-share treaty, where  with a constant (share) .
Optimal forms of reinsurance have been studied extensively in the literature. Most of the results obtained are from the cedent’s point of view. That is, the question asked is: for a given premium principle, what is the optimal functional form and/or parameter values of the ceded function 
f, such that the cedent’s expected utility is maximized or its risk minimized? For example, by maximizing the cedent’s expected utility, Arrow [
5] concluded that “given a range of alternative possible reinsurance contracts, the reinsured would prefer a policy offering complete coverage beyond a deductible.” Borch [
6] showed that for a fixed premium and expected reinsurance payments, the variance of the cedent’s losses is minimized by the excess-of-loss reinsurance policy. In recent years, various solutions to the optimal reinsurance problem have been obtained where the value-at-risk (VaR) and the tail-value-at-risk (TVaR) have been used to measure the cedent’s risk level (e.g., [
7,
8,
9,
10,
11,
12,
13] and the references therein).
Borch [
14] argues that “there are two parties to a reinsurance contract, and that an arrangement which is very attractive to one party may be quite unacceptable to the other.” However, as pointed out by [
15], optimal forms of ceded functions considering both the cedent and the reinsurer had scarcely been discussed until quite recently. For example, Ignatov et al. [
16] study the optimal reinsurance contracts under which the finite horizon joint survival probability of the two parties is maximized. Kaishev and Dimitrova [
17] derive explicit expressions for the probability of joint survival up to a finite time of the cedent and the reinsurer, under an excess of loss reinsurance contract with a limiting and a retention level. Golubin [
18] studies the problem of designing the Pareto-optimal reinsurance policy by maximizing a weighted average of the expected utility of the insurer and the reinsurer. Dimitrova and Kaishev [
19] introduce an efficient frontier type approach to setting the limiting and the retention levels, based on the probability of joint survival. Cai et al. [
20] analyse the optimal reinsurance policies that maximize the joint survival probability and the joint profitable probability of the two parties and derive sufficient conditions for optimal reinsurance contracts within a wide class of reinsurance policies and under a general reinsurance premium principle. Using the results of [
20], Fang and Qu [
21] derive optimal retentions of combined quota-share and excess-of-loss reinsurance that maximize the joint survival probability of the two parties. Cai et al. [
22] study the optimal forms of reinsurance policies that minimize the convex combination of the 
s of the cedent and the reinsurer under two types of constraints that describe the interests of the two parties. For the determination of the optimal excess of loss contract considering the dependency between the losses of the insurer and the reinsurer, we refer to [
23] and the references therein.
A closely-related problem to optimal reinsurance is the so-called optimal transfer of risks among partners, where everybody’s interests are considered simultaneously. The usual approach is to identify Pareto-optimal treaties, whereby no agent can be made better off without making another agent worse off. For results in this area, we refer to, e.g., [
6,
7,
24,
25] and the references therein.
In this paper, we determine Pareto-optimal reinsurance policies under which one party’s risk, measured by its VaR, cannot be reduced without increasing that of the other party in the reinsurance contract. We consider two classes of ceded functions:
	  
     and:
	 
Note the inclusion 
, which has been verified by [
13]. Furthermore, for every 
 both 
f and 
 are Lipschitz continuous, and they are comonotonic.
The requirements that the ceded function 
f is non-decreasing and that the bounds 
 hold for all 
x are needed in 
 and 
 to avoid the moral hazard problem in reinsurance. The additional requirement of the convexity of 
f in 
 essentially requires that 
 approaches infinity linearly when 
 and thus disallows the popular layered reinsurance policies. Nevertheless, this class includes the important quota-share and the excess-of-loss reinsurance policies. Note also that both classes are of interest in the more general context of economic theory with two agents having conflicting interests. Optimal reinsurance problems with admissible classes 
 and 
 have been studied extensively in the literature, and we refer to [
13] for an informative review.
For simplicity of discussion, we assume that the reinsurance premiums are determined by the expected premium principle:
	  
      where 
 is the safety loading. Hence, the cedent’s total loss becomes:
	  
     and the reinsurer’s total loss under the reinsurance contract is:
In this paper, we use VaR to measure the insurer’s and reinsurer’s risk level. A natural starting point for measuring the (joint) risk of the cedent and the reinsurer is a bivariate risk measure, such as the bivariate 
 ([
26]) of the pair 
 and 
. However, since the ceded loss 
 and the retained loss 
 are comonotonic (see [
27,
28] for a very detailed discussion of the concept of comonotonicity with applications), the set of values of the bivariate 
s of 
 and 
 is determined by values of the univariate 
 of 
 and 
. Therefore, the Pareto-optimal reinsurance policies could be determined by minimizing a linear combination of the univariate 
s of 
 and 
. We note in this regard that the optimization criterion of minimizing linear combinations of the risks of the cedent and the reinsurer was adopted by [
7,
22]. Our arguments provide an additional economic meaning to such criteria.
Although VaR is not sub-additive in general, it was shown that it is sub-additive in the deep right tail in many cases of interest (e.g., [
29]). General results related to optimal forms of reinsurance (risk exchanges) using the so-called distortion risk measures exist in the literature, and we refer to [
7,
8,
25]. The distortion risk measures are very general and include VaR, TVaR and proportional hazards transforms as special cases. The feature of the current paper is that we extend the geometric approach of [
12] to our optimization problem that considers the interests of the two parties. The geometric proofs facilitate intuition and enable us to avoid lengthy and complex mathematical arguments. We derive closed-form and user-friendly formulas for the optimal reinsurance policies and thus provide a convenient route for practical implementation of our results.
The rest of the paper is organized as follows. 
Section 2 provides preliminaries and shows (cf. [
25]) that the form of Pareto-optimal reinsurance policies can be determined by minimizing linear combinations of the cedent’s and the reinsurer’s risks. In 
Section 3 and 
Section 4, we determine optimal reinsurance forms and derive the corresponding optimal parameters when the feasible classes of ceded functions are 
 and 
, respectively. There, we also provide illustrative numerical examples. 
Section 5 provides further insights regarding the results of our numerical examples. 
Section 6 concludes the paper.
  2. Preliminaries
Let 
 and 
 denote the cumulative distribution function (c.d.f.) and the survival function of 
X, respectively. Furthermore, let 
 and 
 denote the c.d.f.’s of 
 and 
, respectively. Then, the individual 
s of the cedent and the reinsurer under the reinsurance contract are:
	  
      and:
	  
      respectively. To consider the risk of both the cedent and the reinsurer, we propose to use the bivariate lower orthant 
 introduced by [
26], which is:
	  
For any ceded function 
, the random variables 
 and 
 are comonotonic, and so:
	  
Therefore, when the “joint” risk of the cedent and the reinsurer is measured by their bivariate lower orthant VaR, one could work with the marginal VaRs of  and , instead of the much more complicated joint VaR.
In the following, we assume that the probability levels in the VaRs used by the cedent and the reinsurer are possibly different, say 
 and 
, respectively, and then determine the Pareto-optimal reinsurance policies (ceded functions 
f) in the sense that one party’s risk, measured by its VaR, cannot be reduced without increasing the other party’s VaR. Mathematically, let 
 denote a ceded function in an admissible set 
, such as 
 or 
. Let the corresponding cedent’s and reinsurer’s total losses under the ceded function 
 be denoted by 
 and 
, respectively. Then, 
 is a Pareto-optimal reinsurance policy if there is no ceded function 
 belonging to the admissible set 
, such that:
	  
      and:
	  
	  with at least one of the inequalities being strict. To find the Pareto-optimal reinsurance policies, we utilize the following proposition.
Proposition 1. All Pareto-optimal reinsurance policies f in , , can be determined by solving the problem:where .  Proof.  Similar to the discussion on page 90 of [
30], one method to find Pareto-optimal decisions is to choose two positive constants 
 and find:
		 
 Without loss of generality, we set 
 and 
 with 
. In more detail, let 
g be a function belonging to 
 and minimizing (
2), then there cannot exist in 
 any function 
 such that 
 and 
 with at least one of the inequalities being strict, because otherwise, we would have:
	  
This is a contradiction to the assumed property of function g.
Furthermore, for any two ceded functions 
, the family 
 of ceded functions defined by 
 is a subset of 
 and satisfies:
	  
      and:
	  
Equation (
3) is satisfied because:
	  
      where the last equality is due to the fact that 
 and 
 are non-decreasing functions of the same random variable 
X and therefore comonotonic. Similarly, Equation (
4) is satisfied. Therefore, Condition C on page 90 of [
30] is satisfied, and we conclude that all Pareto-optimal reinsurance policies in 
 can be found by solving Problem (
2). ☐
In view of Proposition 1, throughout the rest of this paper, we seek optimal reinsurance policies by solving the optimization problem:
      for 
, which is equivalent to minimizing:
	  
As shown by [
13], we have 
, and every function 
 is Lipschitz-continuous and, hence, continuous. Consequently (e.g., [
27]), for every 
, we have 
 and thus, with 
 and 
, the optimization problem becomes:
	  
Since we allow 
, the relationships between the probability levels 
 and 
, as well as 
 need to be discussed. Namely, we have the following observations:
	  
- If  -  and  - , then  - . Thus,
           - when  - , the solution to Problem ( 6- ) is  -  for all  x- ; 
- when , the solution is ; 
- when , the objective function is always zero. 
 
- If  -  and  - , then  -  and  - . Thus,
           - when , the optimal ceded function is ; 
- when  - , the form of the optimal ceded function is similar to the case when  - , with only the risk and the profit of the cedent considered (the solution for the latter case can be found in Case 2 of  Section 3.2-  and  Section 4.2-  below); 
- when , the optimal ceded function is . 
 
- If  -  and  - , then  -  and  - . Thus,
           
Throughout the rest of this paper, we only consider the optimal forms of reinsurance policies under the conditions  and .
Now, we are ready to determine the optimal forms of 
f, the task that makes up the contents of the following two sections. Namely, in 
Section 3, we consider the case when the admissible set of ceded functions is 
 and in 
Section 4 when the admissible set is 
. As noted earlier, both classes are of interest in the broad context of economic theory, with the class 
 being more relevant to reinsurance policies. Nevertheless, the class 
 includes the important quota share and excess-of-loss reinsurance policies that provide natural reference points for analysing the optimal reinsurance policies in 
.
  4. Optimal Reinsurance Policy When 
In this section, we determine optimal reinsurance policies when 
, that is when both 
f and the retained loss function 
 are non-decreasing. Comparing this situation with the earlier 
, we can now deal with non-convex ceded functions, such as 
 for any retention level 
. Mathematically, the problem becomes:
		  
As pointed out in the Introduction, solutions to similar problems exist in the literature, and we refer to [
7,
8,
25] for details and further references. Our contribution in this paper is to generalize the geometric arguments of [
12] to the situation when the interests of both the cedent and the reinsurer are taken into account, and we do so in such a way that allows us to avoid lengthy mathematical arguments and consequently helps us to gain useful intuition. In addition, for all scenarios considered, we are able to provide explicit recipes for determining optimal reinsurance policies.
In 
Section 4.1 below, we derive optimal forms of ceded functions, and in 
Section 4.2, we determine parameter values of the optimal functions. 
Section 4.3 contains an illustrative numerical example, which is a continuation of that of 
Section 3.3. Throughout the rest of this section, we assume 
 and 
.
  4.1. Functional Form of the Ceded Function
We have subdivided our considerations into three cases.
  4.1.1. Case 1: 
Similarly to Case 1 of 
Section 3.1.1, we determine the functional form of the ceded function 
 in the following manner. For any 
, we seek 
, such that 
 and:
			  
This requires , as well as the entire function  to be as small as possible for a fixed value of .
As we see from 
Figure 3, because 
f is non-decreasing with a slope not exceeding one, the aforementioned requirements are satisfied by the function:
			  
              where 
 can be any constant. The optimal value of 
d will be determined in 
Section 4.2 below. In reinsurance jargon, the above specified optimal form of the reinsurance policy is for the reinsurer to provide coverage over the layer 
.
  4.1.2. Case 2: 
Similarly to Case 2 of 
Section 3.1.1, since the coefficients in front of 
 and 
 in objective Function (
15) are negative, the optimal reinsurance policy is found by seeking 
, such that 
 and:
			  
As we see from 
Figure 4, these requirements are satisfied by the function:
			  
              where 
 can be any constant. Hence, the optimal form of the reinsurance policy is for the reinsurer to provide a coverage except for the layer 
. In other words, the insurer retains losses in the layer 
.
  4.1.3. Case 3: 
In this case, the minimization problem (
15) simplifies to:
			  
When , because the ceded function is non-decreasing, this requires  to be constant on the interval . Therefore, any function  in  with  on , where  is a constant, is Pareto-optimal.
When , because the slope of the ceded function cannot exceed one, the function  increases at the rate of one on the interval . Therefore, any function  in  with  on  is Pareto-optimal.
Finally, when , then the objective function is always constant.
  4.2. Parameter Values of the Optimal Ceded Function
In this section, we obtain parameter values of the optimal ceded functions that we derived in 
Section 4.1. Four cases are considered separately.
  4.2.1. Case 1:  and 
Theorem 5. Under the conditions  and , the optimal ceded function is   with the parameter:- 1.
-  when ; 
- 2.
-  when . 
 In addition, when , then  for all x.
 Proof.  With the function 
 given by Equation (
16), optimization Problem (
15) becomes:
			  
              where:
			  
The derivative:
			  
              is increasing in 
d. Therefore, when 
, then 
 is minimized at 
. When 
, then 
 is minimized at 
. Finally, when 
, then 
 is minimized at 
, and so, 
. ☐
   4.2.2. Case 2:  and 
With the function 
 given by Equation (
16), optimization problem (
15) reduces to:
			  
              where:
			  
We calculate the derivative:
			  
              which is an increasing function in 
d, and so, we have the following theorem.
Theorem 6. Under the conditions  and , the optimal ceded function is  with the parameter:- 1.
-  when ; 
- 2.
-  when ; 
- 3.
-  when  and ; 
- 4.
-  when . 
 If none of the above conditions are satisfied, then  for all x.
 Proof.  We use similar arguments to those in Theorem 2. We illustrate them here by proving Part (1) only. When , the derivative  reaches zero at  and then remains positive for . Therefore,  reaches its minimum at . With this, we conclude the proof of Theorem 6. ☐
   4.2.3. Case 3:  and 
With the function 
 given by Equation (
17), optimization Problem (
15) reduces to:
			  
              where the objective function is:
			  
Thus:
			  
              which leads us to the following theorem, whose proof is similar to that of Theorem 3 and thus omitted.
Theorem 7. Under the conditions  and , the optimal ceded function is: with the parameter:- 1.
-  when ; 
- 2.
-  when ; 
- 3.
-  when  and ; 
- 4.
-  when  and ; 
- 5.
-  when . 
 If none of the above conditions are satisfied, then  for all x.
   4.2.4. Case 4:  and 
With the function 
 given by Equation (
17), optimization Problem (
15) reduces to:
			  
              where:
			  
Thus,
              
              which gives us the following theorem.
Theorem 8. Under the conditions  and , the optimal ceded function is:with the parameter:- 1.
-  when ; 
- 2.
-  when ; 
- 3.
-  when . 
 If none of the above conditions are satisfied, then  for all x.
   4.3. The Illustrative Example Continued
In this subsection, we continue the illustrative example of 
Section 3.3, but now assume that the admissible class of ceded functions is 
.
  4.3.1. Scenario A:  and 
Applying Theorems 5 and 7, we have:
			  
When 
, then:
			  
              where 
 can be any constant.
The values of 
 versus 
 are reported in 
Table 3.
We have the following observations:
			  
- Since the cedent and the reinsurer have more choices when , their VaRs under the optimal reinsurance policy  are lower than the corresponding ones under . In particular, the reinsurer’s risk is reduced significantly even when . 
- For , the reinsurer assumes the “good” risk in the layer , as well as losses greater than . The former layer creates profit, and the latter layer does not contribute to its VaR because the chance of penetration is too small compared with the probability level  used in its VaR. 
- For , the insurer retains the “good” risk in the layer , as well as the losses greater than . The former layer creates profit, and the latter layer does not contribute to its VaR because the chance of penetration is too small compared with the probability level  used in its VaR. 
  4.3.2. Scenario B:  and 
Applying Theorems 6 and 8, we have:
			  
When 
,
              
              where 
 can be any constant. The values of 
 versus 
 are reported in 
Table 4.
  5. A Numerical Comparison of the Optimal Reinsurance Policies in  and 
In 
Section 3.3 and 
Section 4.3, we derived the Pareto-optimal reinsurance policies in 
 and 
, respectively. In this section, we compare the two cases.
In 
Figure 5, we depict 
 and 
 obtained for Scenario A with the proportional reinsurance 
 when 
a varies from zero to one and also with the excess-of-loss reinsurance 
 when the deductible level 
d varies from zero to 
. The following can be concluded from the figure.
- The efficient frontier for the s of the two parties with  is represented by the path from  to  and then to . Note that the points between A and B represent the VaRs of the two parties resulting from the optimal policies obtained with . The points between B and C represent the VaRs of the two parties resulting from the optimal policies obtained with . 
- The efficient frontier for the VaRs of the two parties when  is represented by the path from  to . 
- For the quota-share reinsurance with  where a ranges from zero to one, the VaRs of the two parties go from B to . When , the quota-share reinsurance policy is quite close to the efficient frontier. 
- For the excess-of-loss reinsurance  with d ranging from zero to , the VaRs of the two parties go along the path  with . 
From 
Figure 5, we conclude that if the reinsurer worries about the right-hand tail more than the primary insurer (
), then the difference between the efficient frontiers obtained for 
 and 
 is significant. This means that the convexity requirement in the definition of 
 is quite restrictive to the reinsurer, and the coverage with an upper limit (which is not allowed in 
) is valuable. In the case when the convexity of the ceded function must be required, quota-share policies are quite efficient.
In 
Figure 6, we compare 
 and 
 obtained for Scenario B with the quota-share reinsurance policies 
 when 
a ranges from zero to one and the excess-of-loss reinsurance policies 
 when the deductible 
d ranges from zero to 
.
In particular, we observe the following:
		  
- The efficient frontier for the VaRs of the two parties with  is represented by the path from  to . 
- The efficient frontier for the VaRs of the two parties when  is represented by the path from  to . In fact, it can algebraically be shown that the path from B to A is actually a part of the path from D to C. That is, by allowing , the efficient frontier is extended from the path  to the path . 
- For the quota-share reinsurance with the parameter a ranging from zero to one, the VaRs of the two parties are represented by the path from  to . We see that when , the quota-share reinsurance policies are not efficient. 
- For the excess-of-loss reinsurance with the parameter d ranging from zero to , the VaRs of the two parties change along the path . We see that setting  is quite efficient, whereas setting  is not. 
From 
Figure 6, we conclude that if the primary insurer worries about the right-hand tail more than the reinsurer (
), then the excess-of-loss policies with the deductible level ranging from 
 to 
 provide a good part of the efficient frontier. The quota-share policies are in general inefficient.