MSTA Modern Stochastics: Theory and Applications 2351-6054 2351-6046 2351-6046 VTeXMokslininkų g. 2A, 08412 Vilnius, Lithuania MSTA45 10.15559/15-MSTA45 Research Article Ruin probability in the three-seasonal discrete-time risk model GrigutisAndriusandrius.grigutis@mif.vu.lt1 KorvelAgneškaagneska.korvel@mif.vu.stud.lt ŠiaulysJonasjonas.siaulys@mif.vu.lt∗ Faculty of Mathematics and Informatics, Vilnius University, Naugarduko 24, Vilnius LT-03225, Lithuania Corresponding author.

The first author was supported by grant No. MIP-049/2014 from the Research Council of Lithuania.

 2015 3012201524421441 26112015 14122015 14122015 © 2015 The Author(s). Published by VTeX2015 Open access article under the CC BY license.

This paper deals with the discrete-time risk model with nonidentically distributed claims. We suppose that the claims repeat with time periods of three units, that is, claim distributions coincide at times {1,4,7,…} , at times {2,5,8,…} , and at times {3,6,9,…} . We present the recursive formulas to calculate the finite-time and ultimate ruin probabilities. We illustrate the theoretical results by several numerical examples.

 Inhomogeneous model three-seasonal model finite-time ruin probability ultimate ruin probability 91B30 60G50 Introduction

The discrete-time risk model is a classical collective risk model for insurance. In the homogeneous version of this model, the insurer’s surplus at each time n∈N0={0,1,2,…} is defined by the following equality: Wu(n)=u+n− ∑i=1nZi, where u∈N0 is the initial insurer’s surplus, and the claim amounts Z1,Z2,… are assumed to be independent copies of a nonnegative integer-valued random variable Z. This random variable and the initial surplus u generate the homogeneous discrete-time risk model. A typical path of the surplus process Wu(n) is shown in Fig. 1.

Behavior of the surplus sequence Wu(n)

The claim amount generator Z can be characterized by the probability mass function (p.m.f.) zk=P(Z=k),k∈N0, or by the cumulative distribution function (c.d.f.) FZ(x)= ∑k=0⌊x⌋zk,x∈R, where ⌊x⌋ denotes the integer part of x.

The homogeneous discrete-time risk model has been extensively investigated by De Vylder and Goovaerts [5, 6], Dickson [7, 8], Gerber [10], Seal [17], Shiu [19, 18], Picard and Lefèvre [15, 16], Lefèvre and Loisel [11], Leipus and Šiaulys [12], Tang [20], and other authors. The ruin time, the ultimate ruin probability, and the finite-time ruin probability are the main extremal characteristics of any risk model. The first time Tu when the surplus Wu(n) becomes negative or null is called the ruin time, that is, Tu=inf{n∈N:Wu(n)⩽0},∞ifWu(n)>0for alln∈N. 0\hspace{2.5pt}\text{for all}\hspace{2.5pt}\hspace{2.5pt}n\in \mathbb{N}.\hspace{1em}\end{array}\right.\]]]> The ruin probability until time T∈N is called the finite-time ruin probability and is defined by ψ(u,T)=P(Tu⩽T). The infinite-time or ultimate ruin probability is defined by ψ(u)=P(Tu<∞). So, for the ultimate survival probability, we have φ(u)=1−ψ(u)=P(Tu=∞). The presented definitions imply that ψ(u,T)=P( ⋃n=1T{u+n− ∑i=1nZi⩽0})=P(max1⩽n⩽T ∑i=1n(Zi−1)⩾u),ψ(u)=P( ⋃n=1∞{u+n− ∑i=1nZi⩽0})=P(supn⩾1 ∑i=1n(Zi−1)⩾u),φ(u)=P( ⋂n=1∞{u+n− ∑i=1nZi>0}),limT↗∞ψ(u,T)=ψ(u). 0\Bigg\}\Bigg),\\{} & \displaystyle \underset{T\nearrow \infty }{\lim }\psi (u,T)=\psi (u).\end{array}\]]]>

Several formulas and procedures for computing finite-time ruin probability and ultimate ruin probability have been proposed in the literature. Here we present some of them having the recursive form.

For the homogeneous discrete-time risk model, we have (see, for instance, [5, 8, 9]): ψ(u,1)=1−FZ(u),u∈N0,ψ(u,T)=ψ(u,1)+ ∑k=0uψ(u+1−k,T−1)zk,u∈N0,T∈{2,3,…}.

If model (1) is generated by the claim generator Z such that EZ<1 , then the ultimate ruin probability can be calculated by the formulas (see, for instance, [8, 9, 18]): ψ(0)=EZ, ψ(u)= ∑j=1u−1(1−FZ(j))ψ(u−j)+ ∑j=u∞(1−FZ(j)),u∈N.

If the homogeneous discrete-time risk model is generated by Z satisfying condition EZ⩾1 , then we say that the net profit condition does not hold, and, in such a case, we have that ψ(u)=1 for all u∈N0 according to the general renewal theory (see, e.g., [14] and the references therein).

The formulas presented enable us to calculate ψ(u) and ψ(u,T) for u∈N0 and T∈N . Nevertheless, there exist many other methods that allow us to calculate or estimate the finite-time and the ultimate ruin probabilities. Some of them can be found in [1, 13, 16].

The assumption for claim amounts {Z1,Z2,…} to be nonidentically distributed random variables is a natural generalization of the homogeneous model. If r.v.s {Z1,Z2,…} are independent but not necessarily identically distributed, then the model defined by Eq. (1) is called the inhomogeneous discrete-time risk model. For such a model, a recursive procedure for calculation of finite-time ruin probabilities can be found in [2, 3]. For the finite-time ruin probabilities ψ(j)(u,T)=P( ⋃n=1T{u+n− ∑i=1nZi+j⩽0}),j∈N0, we have the following theorem.

Let us consider the inhomogeneous discrete-time risk model defined by Eq. (1) in which u∈N0 , zk(j)=P(Z1+j=k) , k,j∈N0 , and FZ(j)(x)=P(Z1+j⩽x) , x∈R . Then ψ(j)(u,1)=1−FZ(j)(u),ψ(j)(u,T)=ψ(j)(u,1)+ ∑k=0uψ(j+1)(u+1−k,T−1)zk(j) for all u∈N0 and T∈{2,3,…} .

According to this theorem, we can calculate the finite-time ruin probability ψ(0)(u,T) of the initial model for all u∈N0 and T∈N . Unfortunately, it is impossible to get formulas for ψ(u) similar to formulas (2) and (3) in the general case because in the case of nonidentically distributed claims, the future of model behavior at each time can be completely new. In paper [4], the general discrete-time risk model was restricted to the model with two kinds of claims. In this model, there are two differently distributed claim amounts that are changing periodically. We call such a model the bi-seasonal discrete-time risk model. In [4] (see Theorem 2.3), the following statement is proved for the calculation of the ultimate ruin probability.

Let us consider a bi-seasonal discrete-time risk model generated by independent random claim amounts X and Y, that is, Z2k+1=dX for k∈{0,1,…} and Z2k=dY for k∈{1,2,…} . Denote S=X+Y and xn=P(X=n) , yn=P(Y=n) , sn=P(S=n) for n∈N0={0,1,…} .

If EX+EY<2 , then limu→∞ψ(u)=0.

If s0=x0y0≠0 , then: ψ(0)=1−(2−ES)limn→∞bn+1−bnan−an+1,1−ψ(u)=αu(1−ψ(0))+βu(2−ES),u∈N, where {αn} , {βn} , n∈N0 , are two sequences of real numbers defined recursively by formulas: α0=1,α1=−1y0,αn=1s0(αn−2− ∑i=1n−1siαn−i−xn−1),n⩾2,β0=0,β1=1y0,βn=1s0(βn−2− ∑i=1n−1siβn−i+xn−1),n⩾2.

If s0≠0 , then ψ(1)=1−(1+ψ(0)−ES)/y0,ψ(u)=1+1s0(ψ(u−2)−1+ ∑k=1u−1sk(1−ψ(u−k)))−xu−1(1−ψ(1))x0,u∈{2,3,…}.

If x0=0 , y0≠0 , then s1≠0 and ψ(0)=1 .

If x0≠0 , y0=0 , then s1≠0 and ψ(0)=ES−1 .

If s0=0 , then, for u∈N , ψ(u)=1−1s1(1−ψ(u−1)− ∑k=2usk(1−ψ(u−k+1))).

In this paper, we consider the discrete-time risk model with three seasons. We obtain a list of formulas similar to those in Theorem 2 to calculate the ultimate ruin probability in such a model. In Section 2, we present a precise definition of the three-seasonal discrete-time risk model and our main statements, whereas in Sections 3 and 4, we give detailed proofs. Finally, Section 5 deals with some numerical examples.

 Main results

We now present the model under consideration.

We say that the insurer’s surplus Wu(n) follows the three-seasonal risk model if Wu(n) is given by Eq. (1) for each n∈N0 and the following assumptions hold:

the initial insurer’s surplus u∈N0 ,

the random claim amounts Z1,Z2,… are nonnegative integer-valued independent r.v.s,

for all k∈N0 , we have Z3k+1=dZ1 , Z3k+2=dZ2 , and Z3k+3=dZ3 .

Let us define p.m.f.s and p.d.f.s by the following equalities: ak=P(Z1=k),bk=P(Z2=k),ck=P(Z3=k),sk=P(S=k),k∈N0, where S=Z1+Z2+Z3 , A(x)= ∑k=0⌊x⌋ak,B(x)= ∑k=0⌊x⌋bk,C(x)= ∑k=0⌊x⌋ck,D(x)= ∑k=0⌊x⌋sk,x⩾0.

It is not difficult to see that the definitions of ruin time, finite-time ruin probability, ultimate ruin probability, and ultimate survival probability remain the same. All expressions of these quantities are the same as in the homogeneous discrete-time risk model. However, the procedures to calculate the finite-time or the ultimate probabilities are more complex than in the homogeneous or be-seasonal discrete-time risk models.

Our first result immediately follows from Theorem 1. The obtained formulas allow us to calculate the finite-time ruin probabilities ψ(u,T)=ψ(0)(u,T) in the three-seasonal risk model for all u∈N0 and all T∈N .

In the three-seasonal discrete-time risk model, for each u∈N0 , we have ψ(0)(u,1)=∑k>uak,ψ(1)(u,1)=∑k>ubk,ψ(2)(u,1)=∑k>uck, u}a_{k},\hspace{2em}{\psi }^{(1)}(u,1)=\sum \limits_{k>u}b_{k},\hspace{2em}{\psi }^{(2)}(u,1)=\sum \limits_{k>u}c_{k},\]]]> and for all u∈N0 and T∈{2,3,…} , we have the following recursive formulas: ψ(0)(u,T)=ψ(0)(u,1)+ ∑k=0uψ(1)(u+1−k,T−1)ak,ψ(1)(u,T)=ψ(1)(u,1)+ ∑k=0uψ(2)(u+1−k,T−1)bk,ψ(2)(u,T)=ψ(2)(u,1)+ ∑k=0uψ(0)(u+1−k,T−1)ck.

Our second result describes the meaning of the net profit condition in the three-seasonal discrete-time risk model. The proof of the theorem is presented in Section 3.

Consider the three-seasonal discrete time risk model generated by independent random claim amounts Z1 , Z2 , and Z3 . If   ES>3 3$]]>, then ψ(u)=1 for each initial surplus u∈N0 . If   ES=3 , then we have the following possible cases:

ψ(0)=ψ(1)=ψ(2)=1 and ψ(u)=0 for u∈{3,4,…} if {a3=b0=c0=1} ;

ψ(0)=ψ(1)=1 and ψ(u)=0 for u∈{2,3,…} if {a0=b3=c0=1} , {a2=b1=c0=1} , {a1=b2=c0=1}, or {a2=b0=c1=1} ;

ψ(0)=1 and ψ(u)=0 for all u∈N if {a0=b0=c3=1} , {a0=b2=c1=1} , {a0=b1=c2=1} , {a1=b0=c2=1} , or {a1=b1=c1=1} ;

ψ(u)=1 for all u∈N0={0,1,2,…} if s3<1 .

Our last statement proposes a recursive procedure for calculation of the ultimate survival probabilities φ(u)=1−ψ(u),u∈N0 . The proof of the formulas is given in Section 4.

Consider the three-seasonal discrete-time risk model generated by independent random claim amounts Z1 , Z2 , and Z3 . Denote S=Z1+Z2+Z3 , sn=P(S=n) for n∈N0 , and suppose that ES<3 . Then the following statements hold.

limu→∞φ(u)=1 .

If s0≠0 , then φ(n)=αnφ(0)+βnφ(1)+γn(3−ES),n∈N0, where α0=1,α1=0,α2=−1b0c0,αn=1s0(αn−3−∑k=1n−1skαn−k−an−2),n⩾3; β0=0,β1=1,β2=−c1c0−1b0,βn=1s0(βn−3−∑k=1n−1skβn−k−an−2c0+c0∑k=0n−1akbn−1−k),n⩾3; γ0=0,γ1=0,γ2=1b0c0,γn=1s0(γn−3−∑k=1n−1skγn−k+an−2),n⩾3.

If {a0=0,b0≠0,c0≠0,a1≠0} , then φ(0)=0,φ(n)=βˆnφ(1)+γˆn(3−ES),n∈N, where βˆ1=β1,βˆ2=β2,γˆ1=γ1,γˆ2=γ2,βˆn=1s1(βˆn−2−∑k=2nskβˆn−k+1−an−1c0−c0φ(1)∑k=1nakbn−k),n⩾3,γˆn=1s1(γˆn−2−∑k=2nskγˆn−k+1+an−1),n⩾3.

If {a0≠0,b0=0,c0≠0,b1≠0} then φ(n)=α˜nφ(0)+γ˜n(3−ES),n∈N, where α˜1=−1/c0,α˜2=c1/c02+1/(a0b1c0),γ˜1=1/c0,γ˜2=−c1/c02,α˜n=1s1(α˜n−2−∑k=2nskα˜n−k+1−∑k=0n−1akbn−k),n⩾3,γ˜n=1s1(γ˜n−2−∑k=2nskγ˜n−k+1+∑k=0n−1akbn−k),n⩾3.

If {a0≠0,b0≠0,c0=0,c1≠0} , then φ(n)=α˘nφ(0)+γ˘n(3−ES),n∈N0, where α˘0=1,α˘1=−1/(b0c1),γ˘0=0,γ˘1=1/(b0c1),α˘n=1s1(α˘n−2−∑k=2nskα˘n−k+1−∑k=0n−1akbn−k),n⩾2,γ˘n=1s1(γ˘n−2−∑k=2nskγ˘n−k+1+∑k=0n−1akbn−k),n⩾2.

If {a0=0,b0=0,c0≠0} , then φ(0)=0 , φ(1)=(3−ES)/c0 , and φ(u+1)=1s2((1−s3)φ(u)− ∑k=1u−1φ(k)su+3−k+c0φ(1) ∑k=0u+2akbu+2−k),u∈N.

If {a0=0,b0≠0,c0=0} , then φ(0)=s2φ(1) , φ(1)=(3−ES)/(s2+b0c1) , and φ(u+1)=1s2((1−s3)φ(u)− ∑k=1u−1φ(k)su+3−k+au+1b0c1φ(1)),u∈N.

If {a0≠0,b0=0,c0=0} , then φ(0)=3−ES , φ(1)=(3−ES)/s2 , and φ(u+1)=1s2((1−s3)φ(u)+ ∑k=1u−1φ(k)su+3−k),u∈N.

If {a0=a1=0,b0≠0,c0≠0} , then φ(0)=0 , φ(1)=(3−ES)/(1/a2+c0) , and the recursion formula (8) is satisfied.

If {a0≠0,b0=b1=0,c0≠0} , then φ(0)=0 , φ(1)=(3−ES)/c0 , and the same recursion formula (8) holds.

If {a0≠0,b0≠0,c0=c1=0} , then φ(0)=3−ES , φ(1)=(3−ES)/s2 , and the recursion formula (10) is satisfied.

We observe that all formulas presented in Theorem 5 can be used to calculate numerical values of survival or ruin probabilities for an arbitrary three-seasonal risk model and for an arbitrary initial surplus value u. The algorithms based on the derived relations work quite quickly and accurately. A few numerical examples for calculating ruin probability in the various versions of the three-seasonal risk model are presented in Section 5.

 Net profit condition

In this section, we present a proof of Theorem 4. We recall that we denote the ultimate survival probability by φ(u) .

For an arbitrary u∈N0 , we have φ(u)=P( ⋂n=1∞{u+n− ∑i=1nZi>0})=P( ⋂n=3∞{u+n− ∑i=1nZi>0}∩{Z1⩾u+1}∩{Z1+Z2⩾u+2})−P( ⋂n=3∞{u+n− ∑i=1nZi>0}∩{Z1⩾u+1})−P( ⋂n=3∞{u+n− ∑i=1nZi>0}∩{Z1+Z2⩾u+2})+P( ⋂n=3∞{u+n− ∑i=1nZi>0}). 0\Bigg\}\Bigg)\\{} & \displaystyle =\mathbb{P}\Bigg({\bigcap \limits_{n=3}^{\infty }}\Bigg\{u+n-{\sum \limits_{i=1}^{n}}Z_{i}>0\Bigg\}\cap \{Z_{1}\geqslant u+1\}\cap \{Z_{1}+Z_{2}\geqslant u+2\}\Bigg)\\{} & \displaystyle \hspace{1em}-\mathbb{P}\Bigg({\bigcap \limits_{n=3}^{\infty }}\Bigg\{u+n-{\sum \limits_{i=1}^{n}}Z_{i}>0\Bigg\}\cap \{Z_{1}\geqslant u+1\}\Bigg)\\{} & \displaystyle \hspace{1em}-\mathbb{P}\Bigg({\bigcap \limits_{n=3}^{\infty }}\Bigg\{u+n-{\sum \limits_{i=1}^{n}}Z_{i}>0\Bigg\}\cap \{Z_{1}+Z_{2}\geqslant u+2\}\Bigg)\\{} & \displaystyle \hspace{1em}+\mathbb{P}\Bigg({\bigcap \limits_{n=3}^{\infty }}\Bigg\{u+n-{\sum \limits_{i=1}^{n}}Z_{i}>0\Bigg\}\Bigg).\end{array}\]]]> Since the model is three-seasonal, the last probability in (11) can be expressed by the sum ∑k=0u+2skP( ⋂n=1∞{u+n+3−k− ∑i=1nZi>0})= ∑k=0u+2skφ(u+3−k), 0\Bigg\}\Bigg)={\sum \limits_{k=0}^{u+2}}s_{k}\hspace{2.5pt}\varphi (u+3-k),\]]]> where, as before, sk=P(Z1+Z2+Z3=k) for k∈N0 .

The second probability in (11) is equals ∑k=u+1∞akP(⋂n⩾3{u+n−k−Z2−Z3− ∑i=4nZi>0})=au+2P(Z2+Z3=0)P( ⋂n=4∞{n−2− ∑i=4nZi>0})+au+1P(Z2+Z3=0)P( ⋂n=4∞{n−1− ∑i=4nZi>0})+au+1P(Z2+Z3=1)P( ⋂n=4∞{n−2− ∑i=4nZi>0})=au+2b0c0φ(1)+au+1b0c0φ(2)+au+1b0c1φ(1)+au+1b1c0φ(1). 0\Bigg\}\Bigg)\\{} & \displaystyle \hspace{1em}=a_{u+2}\hspace{2.5pt}\mathbb{P}(Z_{2}+Z_{3}=0)\hspace{2.5pt}\mathbb{P}\Bigg({\bigcap \limits_{n=4}^{\infty }}\Bigg\{n-2-{\sum \limits_{i=4}^{n}}Z_{i}>0\Bigg\}\Bigg)\\{} & \displaystyle \hspace{2em}+a_{u+1}\hspace{2.5pt}\mathbb{P}(Z_{2}+Z_{3}=0)\hspace{2.5pt}\mathbb{P}\Bigg({\bigcap \limits_{n=4}^{\infty }}\Bigg\{n-1-{\sum \limits_{i=4}^{n}}Z_{i}>0\Bigg\}\Bigg)\\{} & \displaystyle \hspace{2em}+a_{u+1}\hspace{2.5pt}\mathbb{P}(Z_{2}+Z_{3}=1)\hspace{2.5pt}\mathbb{P}\Bigg({\bigcap \limits_{n=4}^{\infty }}\Bigg\{n-2-{\sum \limits_{i=4}^{n}}Z_{i}>0\Bigg\}\Bigg)\\{} & \displaystyle \hspace{1em}=a_{u+2}b_{0}c_{0}\varphi (1)+a_{u+1}b_{0}c_{0}\varphi (2)+a_{u+1}b_{0}c_{1}\varphi (1)+a_{u+1}b_{1}c_{0}\varphi (1).\end{array}\]]]> Similarly, the third probability in (11) is P(Z1+Z2=u+2)P(Z3=0)P( ⋂n=4∞{n−2− ∑i=4nZi>0})=c0φ(1) ∑k=0u+2akbu+2−k, 0\Bigg\}\Bigg)\\{} & \displaystyle \hspace{1em}=c_{0}\varphi (1){\sum \limits_{k=0}^{u+2}}a_{k}b_{u+2-k},\end{array}\]]]> and, finally, the first probability in (11) is P(Z1=u+1)P(Z2=1)P(Z3=0)P( ⋂n=4∞{n−2− ∑i=4nZi>0})+P(Z1=u+2)P(Z2=0)P(Z3=0)P( ⋂n=4∞{n−2− ∑i=4nZi>0})=au+1b1c0φ(1)+au+2b0c0φ(1). 0\Bigg\}\Bigg)\\{} & \displaystyle \hspace{2em}+\mathbb{P}(Z_{1}=u+2)\hspace{2.5pt}\mathbb{P}(Z_{2}=0)\hspace{2.5pt}\mathbb{P}(Z_{3}=0)\hspace{2.5pt}\mathbb{P}\Bigg({\bigcap \limits_{n=4}^{\infty }}\Bigg\{n-2-{\sum \limits_{i=4}^{n}}Z_{i}>0\Bigg\}\Bigg)\\{} & \displaystyle \hspace{1em}=a_{u+1}b_{1}c_{0}\varphi (1)+a_{u+2}b_{0}c_{0}\varphi (1).\end{array}\]]]> Substituting (12)–(15) into (11), we get that φ(u)= ∑k=0u+2skφ(u+3−k)−au+1b0c0φ(2)−au+1b0c1φ(1)−c0φ(1) ∑k=0u+2akbu+2−k for all u∈N0 .

Therefore, for v∈N0 , we have ∑u=0vφ(u)= ∑u=0v ∑k=0u+2skφ(u+3−k)−b0c0φ(2)(A(v+1)−a0)−b0c1φ(1)(A(v+1)−a0)−c0φ(1) ∑u=0v ∑k=0u+2akbu+2−k. We observe that the sum ∑u=0v ∑k=0u+2akbu+2−k can be rewritten in the form a0 ∑u=0vbu+2+a1 ∑u=0vbu+1+a2 ∑u=0vbu+ ∑k=3v+2ak ∑u=k−2vbu+2−k=a0(B(v+2)−b0−b1)+a1(B(v+1)−b0)+a2B(v)+ ∑k=3v+2akB(v+2−k)= ∑k=0v+2akB(v+2−k)−a0b0−a0b1−a1b0. Similarly, the sum ∑u=0v ∑k=0u+2skφ(u+3−k) equals ∑k=1v+3φ(k)D(v+3−k)−s0φ(1)−s1φ(1)−s0φ(2), where D(x)= ∑k=0⌊x⌋sk= ∑k=0⌊x⌋P(Z1+Z2+Z3=k). Relations (17), (18), and (19) imply that ∑k=0vφ(k)= ∑k=1v+3φ(k)D(v+3−k)−b0c0φ(2)A(v+1)−b0c1φ(1)A(v+1)−c0φ(1) ∑k=0v+2akB(v+2−k) or, equivalently, ∑k=0v+3φ(k)(1−D(v+3−k))=φ(v+1)+φ(v+2)+φ(v+3)−φ(0)D(v+3)−b0c0φ(2)A(v+1)−b0c1φ(1)A(v+1)+c0φ(1)( ∑k=0v+2akB‾(v+2−k)+A(v+2)). For each K∈[1,v+2), we have ∑k=0v+2akB‾(v+2−k)= ∑k=0KakB‾(v+2−k)+ ∑k=K+1v+2akB‾(v+2−k). Therefore, lim supv→∞ ∑k=0v+2akB‾(v+2−k)⩽ ∑k=K+1∞ak for each K⩾1 , and so limv→∞ ∑k=0v+2akB‾(v+2−k)=0.

The sequence φ(u),u∈N0 , is bounded and nondecreasing. So, the limit φ(∞):=limu→∞φ(u). Similarly to the derivation of (21), we can get that lim supv→∞ ∑k=0v+3(φ(∞)−φ(k))(1−D(v+3−k))⩽supk⩾K+1(φ(∞)−φ(k))ES for each fixed K⩾1 . Therefore, limv→∞ ∑k=0v+3φ(k)(1−D(v+3−k))=φ(∞)limv→∞ ∑k=0v+3(1−D(k))=φ(∞)ES. Relations (20), (21), and (22) imply that φ(∞)(3−ES)=φ(0)+b0c0φ(2)+b0c1φ(1)+c0φ(1). Now we consider the last equality and examine all possible cases.

If ES>3 3$]]>, then (23) implies that φ(∞)=0 because the left side of (23) is nonnegative in all cases. So, in this case, ψ(u)=1 for all u∈{0,1,2,…} .

If ES=3 , then (23) implies that φ(0)+b0c0φ(2)+b0c1φ(1)+c0φ(1)=0. Additionally, in this situation we have that s3=1 or s3<1 .

If s3=1 , then we have a0b0c0=0,a1b0c0+a0b1c0+a0b0c1=0,a0b0c2+a0b2c0+a2b0c0+a1b1c0+a0b1c1+a1b0c1=0,a0b0c3+a0b3c0+a3b0c0+a1b2c0+a2b1c0+a0b1c2+a0b2c1+a1b0c2+a2b0c1+a1b1c1=1,φ(0)+b0c0φ(2)+b0c1φ(1)+c0φ(1)=0.

Taking into account that all numbers ak , bk , and ck are local probabilities for all k∈N0 , the last system implies the following possible cases.

{a3=b0=c0=1} and φ(0)=φ(1)=φ(2)=0 . In this case, ψ(0)=ψ(1)=ψ(2)=1 and ψ(u)=0,u∈{3,4,…} because Wu(n)=u−2ifn≡1mod3,u−1ifn≡2mod3,uifn≡0mod3.

{a0=b3=c0=1} and φ(0)=φ(1)=0 . In this case, ψ(0)=ψ(1)=1 and ψ(u)=0 for u∈{2,3,…} because Wu(n)=u+1ifn≡1mod3,u−1ifn≡2mod3,uifn≡0mod3.

{a0=b0=c3=1} and φ(0)=0 . In this case, ψ(0)=1 and ψ(u)=0 for u∈{1,2,…} because Wu(n)=u+1ifn≡1mod3,u+2ifn≡2mod3,uifn≡0mod3.

{a2=b1=c0=1} and φ(0)=φ(1)=0 . In this case, ψ(0)=ψ(1)=1 and ψ(u)=0 for u∈{2,3,…} .

{a1=b2=c0=1} and φ(0)=φ(1)=0 . In this case, ψ(0)=ψ(1)=1 and ψ(u)=0 for u∈{2,3,…} .

{a0=b2=c1=1} and φ(0)=0 . In this case, ψ(0)=1 and ψ(u)=0 for u∈{1,2,…} .

{a0=b1=c2=1} and φ(0)=0 . In this case, ψ(0)=1 and ψ(u)=0 for u∈{1,2,…} .

{a2=b0=c1=1} and φ(0)=φ(1)=0 . In this case, ψ(0)=ψ(1)=1 and ψ(u)=0 for u∈{2,3,…} .

{a1=b0=c2=1} and φ(0)=0 . In this case, ψ(0)=1 and ψ(u)=0 for u∈{1,2,…} .

{a1=b1=c1=1} and φ(0)=0 . In this case, ψ(0)=1 and ψ(u)=0 for u∈{1,2,…} .

If s3<1 and ES=3 , then it is necessary that s0≠0 or s1≠0 or s2≠0 because, on the contrary, ES=3s3+4s4+5s5+⋯>3(s3+s4+⋯)=3 3(s_{3}+s_{4}+\cdots \hspace{0.1667em})=3$]]>. In this situation, it suffices to consider the following cases: {s0≠0},{s0=0,s1≠0},{s0=0,s1=0,s2≠0}.

If s0=a0b0c0≠0 , then (23) implies that φ(0)=φ(1)=φ(2)=0 , and from (16) we obtain φ(u)=0 for all u∈{3,4,…} . So, ψ(u)=1 if u∈{0,1,…} in this case.

If s0=a0b0c0=0 and s1=a0b0c1+a0b1c0+a1b0c0≠0, then we have the following possible cases.

{a0=0 , a1≠0 , b0≠0 , c0≠0} . In this situation, Eq. (23) implies that φ(0)=φ(1)=φ(2)=0 , whereas (16) implies that φ(u)=0 for all u∈{3,4,…} . So, ψ(u)=1 for all u∈{0,1,…} in this case.

{a0≠0 , b0=0 , b1≠0 , c0≠0} . In this situation, Eq. (23) implies that φ(0)=φ(1)=0 , and (16) implies that φ(u)=0 for all u∈{2,3,…} . Therefore, ψ(u)=1 for all u∈{0,1,…} again.

{a0≠0 , b0≠0 , c0=0 , c1≠0} .

Equality (23) implies that φ(0)=φ(1)=0 , and (16) implies that φ(u)=0 for all u∈{2,3,…} . So, in this case, ψ(u)=1 for all u∈{0,1,…} .

If s0=a0b0c0=0 , s1=a0b0c1+a0b1c0+a1b0c0=0 and s2=a0b0c2+a0b2c0+a2b0c0+a1b1c0+a1b0c1+a0b1c1≠0 , then there exist the following possible cases.

{a0=0 , a1=0 , a2≠0 , b0≠0 , c0≠0} ;

{a0≠0 , b0=0 , b1=0 , b2≠0 , c0≠0} ;

{a0≠0 , b0≠0 , c0=0 , c1=0 , c2≠0} ;

{a0=0 , a1≠0 , b0=0 , b1≠0 , c0≠0} ;

{a0=0 , a1≠0 , b0≠0 , c0=0 , c1≠0} ;

{a0≠0 , b0=0 , b1≠0 , c0=0 , c1≠0} .

 In all cases, Eqs. (23) and (16) imply that φ(u)=0, and so ψ(u)=1 for all u∈N0 . Theorem 4 is proved.  □

 Recursive formulas

In this section, we prove Theorem 5. Equality (23) from the previous section plays a crucial role.

Let we consider the case ES<3 . First, we prove that φ(∞)=1 . According to the definition φ(∞)=limu→∞P( ⋂n=1∞{ ∑i=1n(Zi−1)<u})=limu→∞P(supu⩾1ηn<u), where ηn= ∑i=1n(Zi−1),n∈N. If n=3N , N∈N , then ηnn=η3N3N=13(1N ∑i=1N(Z3i−2−1)+1N ∑i=1N(Z3i−1−1)+1N ∑i=1N(Z3i−1)). If n=3N+1 , N∈N , then ηnn=η3N+13N+1=N+13N+11N+1 ∑i=1N+1(Z3i−2−1)+N3N+1(1N ∑i=1N(Z3i−1−1)+1N ∑i=1N(Z3i−1)). If n=3N+2 , N∈N , then ηnn=η3N+23N+2=N3N+21N ∑i=1N(Z3i−1)+N+13N+2(1N+1 ∑i=1N+1(Z3i−2−1)+1N+1 ∑i=1N+1(Z3i−1−1)). Hence, the strong law of large numbers implies that ηnn→n→∞13(EZ1−1+EZ2−1+EZ3−1)=ES−33 almost surely.

From this it follows that P(supm⩾n|ηmm+μ|<μ2)→n→∞1 with μ:=(ES−3)/3>0 0$]]>.

For arbitrary positive u and N∈N , we have P(supn⩾1ηn<u)⩾P(( ⋂n=1N{ηn⩽u2})∩( ⋂n=N+1∞{ηn⩽u2}))⩾P( ⋂n=1N{ηn⩽u2})+P( ⋂n=N+1∞{ηn⩽0})−1⩾P( ⋂n=1N{ηn⩽u2})+P(supm⩾N+1|ηmm+μ|<μ2)−1. The last inequality implies that limu→∞P(supn⩾1ηn<u)⩾P(supm⩾N+1|ηmm+μ|<μ2) for arbitrary N∈N .

Hence, according to (24), we have that φ(∞)=1 .

Substituting this into (23), we get 3−ES=φ(0)+b0c0φ(2)+b0c1φ(1)+c0φ(1). In addition, Eq. (16) can be rewritten as follows: φ(u)= ∑k=0u+2su+2−kφ(k+1)−au+1b0c0φ(2)−au+1b0c1φ(1)−c0φ(1) ∑k=0u+2akbu+2−k,u∈N0. Now we consider the last two formulas to get a suitable recursion procedure described in Theorem 5.

First, let s0=a0b0c0≠0 , and let the sequences αn , βn , γn be defined in the statement of Theorem 5.

We prove (4) by induction. We observe that relation (25) implies immediately: φ(0)=α0φ(0)+β0φ(1)+γ0(3−ES),φ(1)=α1φ(0)+β1φ(1)+γ1(3−ES),φ(2)=α2φ(0)+β2φ(1)+γ2(3−ES). Now suppose that Eq. (4) holds for all n=0,1,…,N−1 , and we will prove that (4) holds for n=N . By (26) we have φ(N−3)= ∑k=0N−1sN−1−kφ(k+1)−aN−2b0c0φ(2)−aN−2b0c1φ(1)−c0φ(1) ∑k=0N−1akbN−1−k. Therefore, using the induction hypothesis, we get s0φ(N)=φ(N−3)− ∑k=1N−1skφ(N−k)+aN−2b0c0φ(2)+aN−2b0c1φ(1)+c0φ(1) ∑k=0N−1akbN−1−k=αN−3φ(0)+βN−3φ(1)+γN−3(3−ES)− ∑k=1N−1sk(αN−kφ(0)+βN−kφ(1)+γN−k(3−ES))+aN−2b0c0φ(2)+aN−2b0c1φ(1)+c0φ(1) ∑k=0N−1akbN−1−k. Since φ(2)=−1b0c0φ(0)−c1c0φ(1)−1b0φ(1)+3−ESb0c0 due to (25), we obtain from (27) that φ(N)=φ(0)1s0(αN−3− ∑k=1N−1skαN−k−aN−2)+φ(1)1s0(βN−3− ∑k=1N−1skβN−k−aN−2c0+c0 ∑k=0N−1akbN−1−k)+(3−ES)1s0(γN−3− ∑k=1N−1skγN−k+aN−2)=αNφ(0)+βNφ(1)+γN(3−ES). Hence, the desired relation (4) holds for all n∈N0 by induction.

If {a0=0,b0≠0,c0≠0,a1≠0}, then s0=0 and s1≠0 . Equality (26) with u=0 implies that φ(0)=0 . The recursive relation (5) can be derived from the basic equalities (25) and (26) in the same manner as relation (4).

If {a0≠0,b0=0,c0≠0,b1≠0}, then it follows from Eq. (25) that 3−ES=φ(0)+c0φ(1). Hence, φ(1)=α˜1φ(0)+γ˜1(3−ES) . This is Eq. (6) for n=1 . The validity of (6) for the other n can be derived from (26) using the induction arguments.

In the case {a0≠0,b0≠0,c0=0,c1≠0} , formula (7) follows from (25) if n=1 . For the other n, formula (7) follows from (26) again by using the induction arguments.

In the case {a0=0,b0=0,c0≠0} , we have that s0=s1=0 and s2≠0 because of ES<3 . It follows immediately from (26) that φ(0)=0 , whereas from (25) it follows that φ(1)=(3−ES)/s2 . Finally, we can get the recursive formula (8) from (26) using the same induction procedure.

In the case {a0=0,b0≠0,c0=0} , similarly as in the previous one, we derive that φ(0)=s2φ(1) from (26), we derive that 3−ES=φ(0)+b0c1φ(1) from (25), and we derive the desired formula (9) again from (26).

The case {a0≠0,b0=0,c0=0} is considered completely analogously as both previous cases. Here we omit details.

We have that ES<3 . So, it remains to study the following possible cases: {a0=a1=0,b0≠0,c0≠0},{a0≠0,b0=b1=0,c0≠0},{a0≠0,b0≠0,c0=c1=0}. In all these cases, the presented recursion relations follow from Eq. (26), and the initial values of survival probability φ(0) and φ(1) can be obtained using Eq. (25) together with Eq. (26) with u=0 or u=1 . Theorem 5 is proved.  □

 Numerical examples

In this section, we present three examples of computing numerical values of the finite-time ruin probability and the ultimate ruin probability. All calculations are carried out using software MATHEMATICA. In all presented tables, the numbers are rounded up to three decimal places.

First example. Suppose that the three-seasonal discrete-time risk model is generated by the r.v.s Z1012P0.50.250.25,Z2012P0.40.30.3,Z3012P0.30.350.35.

In Table 1, we give the finite-time ruin probabilities for initial surpluses u∈{0,1,…10,20} and times T∈{1,2,…,10,20} together with the ultimate ruin probabilities for the same u.

Ruin probabilities for the first model

T u
0 1 2 3 4 5 6 7 8 9 10 20
1 0.5 0.25 0 0 0 0 0 0 0 0 0 0
2 0.65 0.325 0.075 0 0 0 0 0 0 0 0 0
3 0.703 0.404 0.128 0.026 0 0 0 0 0 0 0 0
4 0.733 0.445 0.169 0.046 0.007 0 0 0 0 0 0 0
5 0.751 0.475 0.2 0.066 0.014 0.002 0 0 0 0 0 0
6 0.768 0.503 0.233 0.089 0.026 0.005 0.001 0 0 0 0 0
7 0.779 0.523 0.256 0.106 0.035 0.009 0.002 0 0 0 0 0
8 0.788 0.538 0.275 0.122 0.045 0.014 0.003 0.001 0 0 0 0
9 0.796 0.554 0.295 0.139 0.056 0.019 0.006 0.001 0 0 0 0
10 0.802 0.566 0.310 0.152 0.065 0.024 0.008 0.002 0 0 0 0
20 0.836 0.632 0.402 0.243 0.138 0.075 0.038 0.018 0.008 0.003 0.001 0
∞ 0.877 0.722 0.541 0.404 0.301 0.224 0.167 0.125 0.093 0.069 0.052 0.003

Ruin probabilities for the Poison model

T u
0 1 2 3 4 5 6 7 8 9 10 20
1 0.393 0.09 0.014 0.002 0 0 0 0 0 0 0 0
2 0.481 0.152 0.037 0.008 0.001 0 0 0 0 0 0 0
3 0.535 0.205 0.066 0.019 0.005 0.001 0 0 0 0 0 0
4 0.549 0.221 0.075 0.023 0.006 0.002 0 0 0 0 0 0
5 0.562 0.236 0.086 0.028 0.009 0.002 0.001 0 0 0 0 0
6 0.576 0.254 0.099 0.036 0.012 0.004 0.001 0 0 0 0 0
7 0.581 0.26 0.103 0.038 0.013 0.004 0.001 0 0 0 0 0
8 0.585 0.266 0.109 0.042 0.015 0.005 0.002 0.001 0 0 0 0
9 0.591 0.274 0.115 0.046 0.017 0.006 0.002 0.001 0 0 0 0
10 0.593 0.277 0.118 0.048 0.018 0.007 0.002 0.001 0 0 0 0
20 0.605 0.295 0.134 0.059 0.026 0.011 0.005 0.002 0.001 0.0003 0.0001 0
∞ 0.609 0.3 0.139 0.064 0.029 0.013 0.006 0.003 0.001 0.001 0.0002 0

Ruin probabilities for the geometric model

T u
0 1 2 3 4 5 6 7 8 9 10 20
1 0.25 0.063 0.016 0.004 0.001 0 0 0 0 0 0 0
2 0.333 0.111 0.037 0.012 0.004 0.001 0 0 0 0 0 0
3 0.556 0.34 0.218 0.143 0.094 0.063 0.042 0.028 0.019 0.012 0.008 0
4 0.566 0.35 0.226 0.149 0.099 0.066 0.044 0.029 0.019 0.013 0.009 0
5 0.576 0.362 0.236 0.156 0.104 0.069 0.046 0.031 0.021 0.014 0.009 0
6 0.653 0.461 0.334 0.243 0.176 0.127 0.091 0.065 0.046 0.033 0.023 0.001
7 0.657 0.466 0.338 0.247 0.18 0.13 0.093 0.067 0.048 0.034 0.024 0.001
8 0.661 0.471 0.344 0.252 0.184 0.133 0.096 0.069 0.049 0.035 0.025 0.001
9 0.703 0.529 0.406 0.312 0.239 0.181 0.137 0.102 0.076 0.056 0.042 0.002
10 0.705 0.532 0.409 0.315 0.241 0.184 0.139 0.104 0.078 0.057 0.042 0.002
20 0.774 0.635 0.528 0.438 0.363 0.298 0.243 0.197 0.159 0.127 0.101 0.008
∞ 0.927 0.879 0.84 0.803 0.769 0.736 0.705 0.675 0.647 0.619 0.593 0.385

Numerical values of the finite-time ruin probabilities are calculated using the algorithm presented in Theorem 3, whereas the values of the ultimate ruin probabilities are obtained using the formulas of Theorem 5. Namely, first, we observe that ES=2.7 and s0≠0 in this case. So, Eq. (4) holds for an arbitrary n∈N0 . In particular, φ(250)=α250φ(0)+β250φ(1)+0.3γ250,φ(251)=α251φ(0)+β251φ(1)+0.3γ251, According to the first statement of Theorem 5, we can suppose that φ(250)=φ(251)=1 . So, we get φ(0) and φ(1) from this system after calculating the values of {α0,α1,…α251} , {β0,β1,…,β251}, and {γ0,γ1,…,γ251} . Now it remains to use Eq. (4) again to obtain the values φ(u)=1−ψ(u) for initial surplus values u∈{2,3,…,10,20} .

Second example. Suppose now that the three-seasonal discrete-time risk model is generated by three Poison distributions: Z1 with parameter 1/2 , Z2 with parameter 2/3, and Z3 with parameter 4/5 . In Table 2, we present the finite-time ruin probabilities for u∈{0,1,…10,20} , T∈{1,2,…,10,20} and the ultimate ruin probabilities for u∈{0,1,…10,20} . All calculations are made similarly as in the first example.

Third example. We write ξ∼G(p) if ξ is a r.v. having the geometric distribution with parameter p∈(0,1) , that is, P(ξ=k)=p(1−p)k,k∈N0 . Suppose that the three-seasonal risk model is generated by r.v.s Z1∼G(3/4) , Z2∼G(2/3), and Z3∼G(1/3) . In Table 3, we present the finite-time and infinite-time ruin probabilities for this geometric model. The values of initial surpluses u and times T are the same as in the previous examples.

The presented tables show that the behavior of ruin probabilities is closely related to the structure of generating r.v.s and not only to the global model characteristic ES .

 Acknowledgement

We would like to thank the anonymous referee for the detailed and helpful comments on the first version of the manuscript.

 References Asmussen, S., Albrecher, H.: Ruin Probabilities. World Scientific, Singapore (2010). MR2766220. doi:10.1142/9789814282536 Bieliauskienė, E., Šiaulys, J.: Gerber–Shiu function for the discrete inhomogeneous claim case. Int. J. Comput. Math. 89, 1617–1630 (2012). MR2949469. doi:10.1080/00207160.2012.693607 Blaževičius, K., Bieliauskienė, E., Šiaulys, J.: Finite-time ruin probability in the inhomogeneous claim case. Lith. Math. J. 50, 260–270 (2010). MR2719562. doi:10.1007/s10986-010-9084-2 Damarackas, J., Šiaulys, J.: Bi-seasonal discrete-time risk model. Appl. Math. Comput. 247, 930–940 (2014). MR3270895. doi:10.1016/j.amc.2014.09.040 De Vylder, F.E., Goovaerts, M.J.: Recursive calculation of finite-time ruin probabilities. Insur. Math. Econ. 7, 1–8 (1988). MR0971858. doi:10.1016/0167-6687(88)90089-3 De Vylder, F.E., Goovaerts, M.J.: Explicit finite-time and infinite-time ruin probabilities in the continuous case. Insur. Math. Econ. 24, 155–172 (1999). MR1704808. doi:10.1016/S0167-6687(98)00049-3 Dickson, D.C.M.: On numerical evaluation of finite time survival probabilities. Br. Actuar. J 5, 575–584 (1999). doi:10.1017/S135732170000057X Dickson, D.C.M.: Insurance Risk and Ruin. Cambridge University Press, Cambridge (2005). MR2160707. doi:10.1017/CBO9780511624155 Dickson, D.C.M., Waters, R.: Recursive calculation of survival probabilities. ASTIN Bull. 21, 199–221 (1991). doi:10.2143/AST.21.2.2005364 Gerber, H.U.: Mathematical fun with the compound binomial process. ASTIN Bull. 18, 161–168 (1988). doi:10.2143/AST.18.2.2014949 Lefèvre, Cl.: Loisel St.: On finite-time ruin probabilities for classical risk models. Scand. Actuar. J. 2008, 41–60 (2008). MR2414622. doi:10.1080/03461230701766882 Leipus, R., Šiaulys, J.: Finite-horizon ruin probability asymptotics in the compound discrete-time risk model. Lith. Math. J. 51(2), 207–219 (2011). MR2805739. doi:10.1007/s10986-011-9120-x Li, S., Lu, Y., Garrido, J.: A review of discrete-time risk models. Rev. R. Acad. Cienc. Exactas Fís. Nat., Ser. A Mat. 103, 321–337 (2009). MR2582636. doi:10.1007/BF03191910 Mikosch, T.: Non-Life Insurance Mathematics. Springer, Berlin (2009). MR2503328. doi:10.1007/978-3-540-88233-6 Picard, Ph.: Lefèvre Cl.: The probability of ruin in finite time with discrete claim size distribution. Scandinavian Actuarial journal, 58–69 (1997). 1997. MR1440825. doi:10.1080/03461238.1997.10413978 Picard, Ph.: Lefèvre Cl.: Probabilité de ruine éventuelle dans un modèle de risque à temps discret. J. Appl. Probab. 40, 543–556 (2003). MR1993252. doi:10.1239/jap/1059060887 Seal, H.L.: The Stochastic Theory of a Risk Business. Wiley, New York (1969) Shiu, E.S.W.: Calculation of the probability of eventual ruin by Beckman’s convolution series. Insur. Math. Econ. 7, 41–47 (1989). MR0971864. doi:10.1016/0167-6687(88)90095-9 Shiu, E.S.W.: The probability of eventual ruin in the compound binomial model. ASTIN Bull. 19, 179–190 (1989). doi:10.2143/AST.19.2.2014907 Tang, Q.: The ruin probability of a discrete-time risk model under constant interest with heavy tails. Scand. Actuar. J. 2004, 229–240 (2004).