Survival Probability in the Life Annuity Insurance Model with Stochastic Return on Investments

Abstract

We investigate the survival probability in the life annuity, or pension, insurance model when whole of the surplus (or a xed its part) is invested into a risky asset with the price following the geometric Brownian motion. For the case of exponential distribution of revenue sizes, we formulate a singular boundary value problem for linear integrodi erential equation and prove that the survival probability as a function of the initial surplus is the unique solution of this problem. Moreover, asymptotic representations for the survival probability both for small and large values of initial surplus are obtained. The e cient algorithm for the numerical calculation of the survival probability is described. Using computational experiments, we show that in the pension insurance business risky investments play a very important role in strengthening of the insurers solvency in a zone of small sizes of the surplus.

Keywords

• life annuity insurance
• dual risk model
• survival probability
• investment
• risky asset
• geometric Brownian motion
• exponential distribution of revenue sizes

-

Introduction and Statement of the Problem We consider the life annuity insurance model [9] (so called \dual risk model", see, e. g., [2]), where the surplus or equity of a company (in the absence of investments) is of the form

Rt = u

N(t) ct + X Zk; t k=1 0: ( 1 ) Here Rt is the surplus of a company at time t 0; u is the initial surplus, c is the life annuity rate (or the pension payments per unit of time), assumed to be deterministic and xed. N (t) is a homogeneous Poisson process with intensity > 0 that, for any t > 0, determines the number of random revenues up to the time t; Zk (k = 1; 2; : : : ) are independent identically distributed random variables with a distribution function F (z) (F (0) = 0, EZ1 = m < 1) that determine the revenue sizes and are assumed to be independent of N (t). These revenues arise at the moments of the death of policyholders.

In comparison with the classical non-life collective risk model (so called Cramer-Lundberg (CL) model, see [9]), the circumstances in ( 1 ) are reversed: in the classical model second and third summands have opposite signs. Since the \claims" in the dual model are negative (the jumps of the process ( 1 ) are positive), this model is also called the insurance model with negative risk sums or compound Poisson model with negative claims [3].

Let now the whole surplus be continuously invested into risky asset with price St following the geometric Brownian motion dSt =

Stdt +

StdBt; t

0; ( 2 ) where is the expected return rate, is the volatility, Bt is a standard Brownian motion.

Then the resulting surplus process Xt is governed by the equation dXt =

Xtdt +

XtdBt + dRt; t

0; ( 3 ) with the initial condition X0 = u, where Rt is de ned by ( 1 ).

Remark 1. For the case when only xed part of the surplus is invested into risky asset while the rest part is invested into a risk free asset with xed return rate, the equation ( 3 ) is valued with modi ed parameters for the corresponding surplus process (see, e. g., [4]).

Denote '(u) = P (Xt 0; t 0) the survival probability (i. e., the probability that bankruptcy will never happen); (u) = 1 '(u) is the ruin probability. Let '0(u) = P (Rt 0; t 0) be the survival probability for the process ( 1 ) and 0(u) = 1 '0(u) be the corresponding ruin probability (in the absence of investments).

For the case of exponential distribution of revenue sizes, the following theorem is proved in [10].

Theorem 1. Let F (z) = 1 e z=m, m > 0, := 2 = 2 1. Then: 1) if > 0 then (u) = Ku (1 + o( 1 )); u ! 1, for some K > 0; 2) if 0 then (u) = 1 for any u.

The formulation of this theorem is exactly the same as in [8], [14] for the non-life insurance model.

For the ruin probability 0(u), it is easy to obtain an integro-di erential equation (IDE) using the obvious modi cations of the \di erential argument" (see, e. g., [9]). In the case of exponential distribution of revenue sizes and if the safety loading is positive, i. e., the inequality m > c is valid, this IDE has an exact solution satisfying boundary conditions limu!+0 0(u) = 1, limu!1 0(u) = 0; this solution has the form 0(u) = exp f ( =c 1=m)ug :

Such slow decay as in Theorem 1 is in contrast with the exponential representation ( 4 ) in corresponding model without investment. This fact leads to the following conclusion, which has been made earlier for various models with non-heavy-tailed distributions of claims: investment of the whole surplus into risky assets in a zone of large values of the surplus impair the insurers solvency. The same conclusion remains true when only xed part of the surplus is invested into risky assets (so called simple investment strategies), see [4] and references therein.

At the same time, the studies of optimal investment strategies, which maximize the survival probabilities in various settings of a problem for the CL model, show that the risky assets play a crucial role in strengthening of the insurer's solvency in a zone of small sizes of the surplus (see, e. g., [5] and references therein). The same conclusions concern the simple investment strategies for CL model and some its modi cations, see [6].

The main goal of our paper is to identify the impact of simple investment strategies on the solvency in the dual risk model not only in the case of large surplus levels but also in the case of its small levels. For this purpose, we use the approach based on so called su ciency theorems [4], which state that the solutions of singular problems for linear IDEs, generated by in nitesimal operators of the resulting surplus processes, de ne the corresponding survival probabilities. This approach is rather di erent from the one used in [10] and eliminates the need a priory to prove the twice continuously di erentiability of the survival probability as well as the justi cation of boundary conditions. Solving the singular problem for IDE, we calculate the survival probability as a function of the initial surplus on the whole nonnegative semi-axis by means of the proposed algorithm; some results of numerical experiments are described. These results allow us to make conclusions about impact of the simple risky investment strategies on the solvency in the life annuity insurance model. 2

Some Preliminary Results

Recall at rst that the in nitesimal generator A (see, e. g., [11]) of the process Xt de ned by ( 3 ) has the form (Af )(u) = 1 2u2f 00(u) + f 0(u)( u 2 c) f (u) +

Z 1 0 f (u + z) dF (z); ( 5 ) for any function f from a certain subclass D of the space C2(IR+) of real-valued, twice continuously di erentiable on (0; 1) functions.

In the assumption '(u) 2 D, where '(u) is the survival probability of the process Xt, some considerations based on the generalized Ito's formula and the complete probability formula allow us to write the following equation for u > 0: (A')(u) = 0 (6) (see [10]; for the corresponding equation in the CL model with stochastic return on investment, see, e. g., [13]).

For the case of exponential distribution of revenue sizes, namely when As noted in [10], the life annuity insurance case is rather di erent from the nonlife insurance one because the change of two signs to the opposite ones in the equation de ning the dynamics of the reserve leads to technical complications. In contrast to the non-life insurance case, the risk process ( 3 ) may leave the positive half-axis only in a continuous way. In [10] it was emphasized that the main di culty in deriving the IDE is to prove the smoothness of the ruin probability. The other di culty is to establish lower and upper asymptotic bounds for ruin probability in order to identify boundary conditions at in nity for the survival probability as the solution of the IDE. The smoothness of the ruin probability is studied in [10] using a method based on integral representations; as a tool for the proof of the asymptotic bounds for the ruin probability some theorems of the renewal theory are used.

As mentioned above in the introduction, in this paper we apply the approach based on su ciency theorems (developed in [4] for the CL model and its modi cation with stochastic premiums; see also the earlier paper [13]) which allows us to avoid the a-priori proof of the the smoothness of the ruin probability as well as the justi cation of the boundary conditions at in nity.

For this purpose, we need a few preliminary propositions.

Lemma 1. For the survival probability '(u) of the process ( 3 ) with the initial condition X0 = u, the following relation is valid: '(0) = 0; (10) i. e., the ruin occurs immediately at zero initial surplus.

The proof of this lemma is obvious due to the negativity of the deterministic component and we omit it. We provide below several statements concerning the properties of the solutions to IDE (8) satisfying the various conditions. Lemma 2. Let in IDE (8) the parameters c, , , m be xed positive numbers and is arbitrary xed number. Then if there exists a solution '(u) to IDE (8) with conditions lim '(u) = 0; u!+0 lim '(u) = 1; u!1 then this solution is unique.

This lemma can be easily proved by contradiction using the linearity of IDE.

To formulate some further auxiliary propositions, we will use also the following limiting conditions: ul!im+0 j'0(u)j < 1; lim u'00(u) = 0: u!+0 lim u'0(u) = 0; u!1 lim u2'00(u) = 0: u!1 Lemma 3. Let in IDE (8) the parameters c, , , m be xed positive numbers and is arbitrary xed number. If there exists the solution '(u) of IDE (8), satisfying conditions (11), (12), then (11) (12) (13) (14) (15) (16) 1 2u2'000(u) + 2 u + 2u c

'00(u) + 1 2m 2u2 + de ned on IR+, with conditions (11){(13). u m c

'0(u) = 0; (17) 0 '(u) 1; moreover, The rst part of this statement may be proved by contradiction (see the proof of a similar assertion in [6]). Let us prove the second part of the statement. Indeed, from the IDE (8) and conditions (11) and (12) we have the relation

Z 1 m 0 wherefrom, taking into account the proved above relation (14), the conditions (11) and positiveness of c and other parameters, we conclude that (15) is valid.

The following lemma is essential auxiliary statement for further study of the initial problem.

Lemma 4. Let in IDE (8) all the parameters c, , , m be xed positive numbers. Then the singular IDE problem (8), (11){(13) is equivalent to the singular problem for ODE Proof. Let '(u) be satisfying IDE (8). Let us show that it satis es also ODE (17). It is easy to check that, for the operator (9), the following relation is valid: d du 1 2u2'000(u) + 2u'00(u) + ( u 2 The obvious linear combination of IDEs (8) and (19), which exclude the integral term (Jm')(u), leads to ODE (17).

Conversely, let now '^(u) be satisfying ODE (17) and conditions (11) and (13). Let us show that it satis es also the IDE (8). Denote g(u) the left-hand side of the equation (8) with function '^(u). Then g(u) = '^(u) + (Jm'^)(u) + ( u c)'^0(u) + 1 2u2'^00(u); 2 g0(u) = [(Jm'^)(u)

'^(u)] + ( Hence, g0(u) m g(u) m =

)'^0(u) + + ( u + 2u c)'^00(u) + 1 2u2'^000(u):

2 1 2m 2u2 '^00(u) + 1 2u2'^000(u); 2 and, in view of the fact that the function '^(u) is a solution of ODE (17), we obtain

The solution of ODE (20) has the form u c m

'^0(u) + + u + 2u c g0(u) g(u) m

= 0: g(u) = C exp (u=m); u > 0; where C is an arbitrary constant. It is easy to see that for the function '^(u), which satis es conditions (11) and (13), the following relation is valid: lim u!+1 m u 1 Z 1 '^(s) exp ( (s u)=m)ds = 1: (20) (21) Then, taking into account the de nition of g(u), equality (22) and conditions (11), (13) we conclude that the equality limu!1 g(u) = 0 holds. Consequently, in view of positiveness m, the constant C in (21) should be equal to zero for this solution, i. e., g(u) 0. Thus, '^(u) is the solution of IDE (8).

It remains to note that the whole set of conditions is the same for the two considered problems for IDE and ODE, and lemma is proved.

To establish a connection between the original problem of the survival probability investigation and a singular problem for IDE, we need also the following statement which we call the su ciency theorem [4].

Theorem 2. Let in IDE (8) all the parameters be positive numbers and the inequality

2 > 2 be ful lled. Suppose IDE (8) has a twice continuously di erentiable on (0; 1) solution '(u) subject to conditions 0 '(u) 1; The proof of this theorem is completely analogous to the proof of Theorem 3.1 in [4]. 3

Main Theorem

For the considered case of the exponential distribution of revenue sizes, we establish the following statement.

Theorem 3. Let F (z) be of the form (7), all the parameters , 2, m, c, be xed positive constants, and let the condition (23) be satis ed. Then the following assertions hold:

(I) the survival probability '(u) of the process ( 3 ) with initial condition X0 = u is the solution to the singular boundary value IDE problem (8), (11); (II) this solution is unique and satis es the following relations: 0 '(u) 1; (23) (24) (25) (26) (27) (28) where and

0 (IV) '(u) has the asymptotic representations ul!im+0 j (u)j < 1; lim u (u) = 0; u!1 lim u 0(u) = 0; u!+0 lim u2 0(u) = 0; u!1 Z 1

(s)ds = 1; '(u)

D1

1 ! u + X Dkuk=k ; k=2 u +0; D1 = '0(+0); D2 = ( D3 =

D2(2 + + c=m) =c; 2 + c=m)

=m =(2c); Dk+1 = [Dk(k(k 1) 2=2 + k + c=m) where (u) = '0(u) is the solution of the following singular problem for ODE: where K > 0 is a constant;

(V) as u ! +0, the behavior of the survival probability derivatives depends on the relations between the parameters, in particular on a sign of the coe cient ir = ( )m c: ( 1 ) if ir 0, then limu!+0 '00(u) 0, moreover, the solution ' is concave on IR+; ( 2 ) if ir < 0, then limu!+0 '00(u) > 0, the solution ' is convex in a some neighborhood of zero and has an in exion point.

Sketch of the proof. At rst, we need to establish the existence and uniqueness of the solution to the problem (29){(32) and to study its asymptotic behaviors for large and small values of u. For this purpose, we have to investigate the singular problems (29), (30) and (29), (31) separately, taking into account that ODE (28) has irregular singular points at zero and in nity (about singular points for ODEs see, e. g. [15]).

By using methods of the investigation of ODEs with singular points [15], [12] we obtain asymptotic representation for families of solutions to this singular problems (see also [6] and references therein for analogous investigation (30) (31) (32) (33) (34) (35) (36) (37) with application of these methods for CL model with investment in details). As result we have that ODE (29) for small u > 0 has a two-parameter family of solutions (u; D1; C1) and for these solutions the following asymptotic representation holds: (u; D1; C1) = = D1 (1 + 1(u)) + C1 exp 2c=( 2u) u 2 = 2 (1 +

2(u)); +0, the function u +0; u ! 0: (38) 1(u) can be (39) Here C1 is a parameter, 2(u) = o( 1 ), u ! represented by asymptotic series where the coe cients Dk, k = 2; 3; : : : , may be found from the recurrence relations (34){(36).

It is obvious that the conditions (30) hold for all the solutions of the family (u; D1; C1), i. e., for all the solutions to ODE (29).

Under condition (23), ODE (29) has a one-parameter family of solutions (u; C2) which are integrable at in nity. For these solutions, the following asymptotic representation holds as u ! 1: (u; C2) = C2u 2 = 2 (1 + o( 1 )); 0(u; C2) =

C2u 2 = 2 (1 + o( 1 )): (40) 2 2u It is obvious that the conditions (31) hold for all solutions of this family (i. e., for all integrable at in nity solutions of (29)) i the condition (23) is ful lled.

Thus, if the condition (23) is ful lled, then there exists one-parameter family of solutions to the problem (29)-(31). All the solutions of the equation (29) are bounded at zero, and all bounded and integrable at in nity solutions belong to this family and have the asymptotic representations (38) and (40). The condition (32) extracts the unique solution from this family. It is clear that this solution satis es the conditions (30) and (31). Then, taking into account Lemma 4, it is easy to see that the function '(u) de ned by the formula (28) is the solution to the IDE problem (8), (11){(13) and has the asymptotic representations (33) and (37). Therefore, according to Lemma 3, the relations (14) also hold and, in view of Theorem 2, for any u 2 IR+, the value of '(u) is the survival probability for the process ( 3 ) with initial state X0 = u. In accordance to Lemma 2 (about the uniqueness) this probability as a function of u satis es conditions (12) and (13) with necessity. In view of Lemma 3, we have also that the inequalities (15) take place, and, in accordance to the asymptotic representation (33), we conclude that '00(+0) and the expression + c=m are of the same sign. The sketch of the proof is completed. 4

Numerical Examples

The studies given in previous sections allow us to suggest computationally simple and theoretically justi ed algorithm for numerical calculation of the survival probability in the considered model. This algorithm requires to solve the singular Cauchy problem from in nity (29), (31) with the normalizing condition (32). Then it remains to use the relation (28) (recall that all the solutions of ODE (29) are bounded as u ! +0 and all the integrable at in nity solutions form the one-parameter family). To solve numerically the problem (29), (31) we realize previously the equivalent transfer of the limit conditions (31) from in nity to a large nite point using the results [7], [12]. For the rst approximation, such approach yields boundary condition at a nite point u = u1 1 as follows 2 2u1 (u1) (the same relation follows also from (40)).

For general ODE systems with pole-type singular points, the theory of boundary condition transfer from singular points is developed; such transfer can be realized by construction of the stable initial manifolds, or the Lyapunov manifolds of conventional stability, at the neighborhoods of singular points (see, e. g., [1] and references therein). On the application of such approach in actuarial mathematics, see [6] and references therein.

Numerical experiments (see in particular Figs. 1, 2) show that risky investments improves survival probability at a zone of small values of initial surplus in the case of positive safety loading (Fig. 2). Moreover, risky investments make survival possible in the case of negative safety loading, see Fig. 1 (in this case survival is impossible without investments). To study the impact of investments with stochastic return on survival probability in the life annuity insurance model we use the approach based on so called su ciency theorem and the existence theorems for the corresponding singular problems for IDEs (see [4]). This uni ed approach eliminates need to proof regularity of the survival probability as well as to use its upper and lower bounds. Moreover, the solving of above singular problem for IDE leads to calculation of the survival probability on all the non-negative semi-axis. We reduce the problem (8), (11) to a certain initial problem from in nity for some second order ODE with respect to the derivative of the survival probability with normalizing condition. As a result of calculations, we conclude in particular that if the value of safety loading ( m c) in the model ( 1 ) is negative or su ciently small and the surplus is small too, then the use of the risky investments allows to increase the survival probability signi cantly. 6. Belkina, T.A., Konyukhova, N.B., Kurochkin, S.V.: Dynamical insurance models with investment: Constraint singular problems for integro-di erential equations.

Comput. Math. Math. Phys. 56( 1 ), 47-98 (2016) 7. Birger, E.S., Lyalikova, N.B.: Discovery of the solutions of certain systems of differential equations with a given condition at in nity, I. U.S.S.R. Comput. Maths. Math. Phys. 5(6), 1{17 (1965); On nding the solutions for a given condition at in nity of certain systems of ordinary di erential equations, II. U.S.S.R. Comput.

Maths. Math. Phys. 6( 3 ), 47{57 (1966) 8. Frolova, A., Kabanov, Yu., Pergamenshchikov, S.: In the insurance business risky investments are dangerous. Finance Stochast. 6( 2 ), 227-235 (2002) 9. Grandell, J.: Aspects of Risk Theory. Springer, Berlin{New York (1991) 10. Kabanov, Yu., Pergamenshchikov, S.: In the insurance business risky investments are dangerous: the case of negative risk sums. Finance Stochast. 20( 2 ), 355{379 (2016) 11. Karatzas, I., Shreve, S.: Brownian Motion and Stochastic Calculus. Springer

Verlag, NY (1988) 12. Konyukhova, N.B.: Singular Cauchy problems for systems of ordinary di erential equations. U.S.S.R. Comput. Maths Math. Phys. 23( 3 ), 72{82 (1983) 13. Paulsen, J., Gjessing, H.K.: Ruin theory with stochastic return on investments.

Adv. Appl. Probab. 29( 4 ), 965{985 (1997) 14. Pergamenshchikov, S. and Zeitouny, O.: Ruin probability in the presence of risky investments. Stochastic Process. Appl. 116( 2 ), 267{278 (2006) 15. Wasow, W.: Asymptotic Expansions for Ordinary Di erential Equations. Dover, New York (1987).

References

1. Abramov , A.A. , Konyukhova , N.B. : Transfer of admissible boundary conditions from a singular point for systems of linear ordinary di erential equations . Sov. J. Numer. Anal. Math. Modelling (VNU Science Press, Utrecht, The Netherlands) 1 ( 4 ), 245 { 265 ( 1986 )

2. Albrecher , H. , Badescu , A. , Landriault , D. : On the dual risk model with tax payments . Insurance Math. Econom . 42 , 1086 { 1094 ( 2008 )

3. Asmussen , S. , Albrecher , H.: Ruin Probabilities . Advanced Series on Statistical Science and Applied Probability , V. 14 . World Scienti c, Singapore ( 2010 )

4. Belkina , T.A. : Risky investment for insurers and su ciency theorems for the survival probability . Markov Processes Relat. Fields . 20 , 505 { 525 ( 2014 )

5. Belkina , T. , Hipp , C. , Luo , S. , Taksar , M. : Optimal constrained investment in the Cramer-Lundberg model . Scandinavian Actuarial Journal 5 , 383 { 404 ( 2014 )