We propose a general continuous-time risk model with a constant interest rate. In this model, claims arrive according to an arbitrary counting process, while their sizes have dominantly varying tails and fulfill an extended negative dependence structure. We obtain an asymptotic formula for the finite-time ruin probability, which extends a corresponding result of Wang (2008).

1. The Dependent General Risk Model

In this paper, we consider the finite-time ruin probability with constant interest rate in a dependent general risk model. In this model, the claim sizes {𝑋𝑛,𝑛≥1} form a sequence of identically distributed, not necessarily independent, and nonnegative random variables (r.v.s) with common distribution 𝐹 such that 𝐹(𝑥)=1−𝐹(𝑥)=P(𝑋1>𝑥)>0 for all 𝑥>0; the claim arrival process {𝑁(𝑡),𝑡≥0} is a general counting process, namely, a nonnegative, nondecreasing, right continuous, and integer-valued stochastic process with 0<E𝑁(𝑡)=𝜆(𝑡)<∞ for all large 𝑡>0. The times of the successive claims are denoted by {𝜏𝑛,𝑛≥1}. The total amount of premiums accumulated up to time 𝑡≥0, denoted by 𝐶(𝑡) with 𝐶(0)=0 and 𝐶(𝑡)<∞ almost surely for every 𝑡>0, is another nonnegative and nondecreasing stochastic process. Assume that {𝑋𝑛,𝑛≥1}, {𝑁(𝑡),𝑡≥0} and {𝐶(𝑡),𝑡≥0} are mutually independent. Let 𝛿>0 be the constant interest rate (i.e., after time 𝑡 one dollar becomes 𝑒𝛿𝑡 dollars), and let 𝑥≥0 be the initial capital reserve of an insurance company. Then, the total discounted reserve up to time 𝑡≥0, denoted by 𝐷(𝑡,𝑥), can be written as 𝐷(𝑡,𝑥)=𝑥+𝑡0𝑒−𝛿𝑠𝐶(d𝑠)−𝑁(𝑡)𝑛=1𝑋𝑛𝑒−𝛿𝜏𝑛.(1.1) For a finite time 𝑇>0, the finite-time ruin probability is defined by Ψ(𝑥,𝑇)=P(𝐷(𝑡,𝑥)<0,forsome0≤𝑡≤𝑇)=Psup𝑡∈[0,𝑇]𝑁(𝑡)𝑛=1𝑋𝑛𝑒−𝛿𝜏𝑛−𝑡0𝑒−𝛿𝑠,𝐶(d𝑠)>𝑥(1.2) while the ultimate ruin probability is defined by Ψ(𝑥)=Ψ(𝑥,∞)=P(𝐷(𝑡,𝑥)<0,forsome𝑡≥0).(1.3)

If the claim sizes {𝑋𝑛,𝑛≥1} are independent r.v.s, the model is called the independent general risk model, which was introduced by Wang [1]. In particular, if 𝐶(𝑡)=𝑐𝑡, 𝑡≥0, with 𝑐>0 a deterministic constant and {𝑁(𝑡),𝑡≥0} is a Poisson process, then the model reduces to the classical one.

2. Introduction and Main Result

Hereafter, all limit relationships hold for 𝑥 tending to ∞ unless otherwise stated. For two positive functions 𝑓(𝑥) and 𝑔(𝑥), we write 𝑓(𝑥)∼𝑔(𝑥) if lim𝑓(𝑥)/𝑔(𝑥)=1; write 𝑓(𝑥)≲𝑔(𝑥) if limsup𝑓(𝑥)/𝑔(𝑥)≤1 and 𝑓(𝑥)=𝑜(𝑔(𝑥)) if lim𝑓(𝑥)/𝑔(𝑥)=0. The indicator function of an event 𝐴 is denoted by 𝟏𝐴.

In risk theory, heavy-tailed distributions are often used to model large claim amounts. They play a key role in insurance and finance. We will restrict the claim-size distribution 𝐹 to be heavy tailed. A distribution 𝑉 is said to be dominatedly varying tailed, denoted by 𝑉∈𝒟, if limsup𝑉(𝑥𝑦)/𝑉(𝑥)<∞ for any 𝑦>0. A distribution 𝑉 is said to be long tailed, denoted by 𝑉∈ℒ, if lim𝑉(𝑥+𝑦)/𝑉(𝑥)=1 for any 𝑦>0. A distribution 𝑉 is said to be subexponential, denoted by 𝑉∈𝒮, if 𝑉𝑛∗(𝑥)∼𝑛𝑉(𝑥) for any 𝑛≥2, where 𝑉𝑛∗ denotes the 𝑛-fold convolution of itself. A distribution 𝑉 is said to be regularly varying tailed, denoted by ℛ−𝛼,𝛼>0, if lim𝑉(𝑥𝑦)/𝑉(𝑥)=𝑦−𝛼 for any 𝑦≥1. A proper inclusion relationship holds that ℛ−𝛼⊂ℒ∩𝒟⊂𝒮⊂ℒ,(2.1) see, for example, Cline [2] or Embrechts and Omey [3]. For a distribution 𝑉, denote the upper Matuszewska index of the distribution 𝑉 by 𝐽+𝑉=−limğ‘¦â†’âˆžlog𝑉∗(𝑦)log𝑦with𝑉∗(𝑦)=liminfğ‘¥â†’âˆžğ‘‰(𝑥𝑦)𝑉(𝑥),𝑦>1.(2.2) In the terminology of Bingham et al. [4], the quantity 𝐽+𝑉 is actually the upper Matuszewska index of the function 1/𝑉(𝑥), 𝑥≥0, as also pointed out in Tang and Tsitsiashvili [5]. Additionally, denote 𝐿𝑉=lim𝑦↘1𝑉∗(𝑦) (clearly, 0≤𝐿𝑉≤1 ). The presented definitions yield that the following assertions are equivalent: (i)𝑉∈𝒟,(ii)𝑉∗(𝑦)>0forsome𝑦>1,(iii)𝐿𝑉>0,(iv)𝐽+𝑉<∞.(2.3)

The asymptotic behavior of the ruin probability in the classical risk model has been extensively investigated in the literature. Klüppelberg and Stadtmüller [6] considered the ultimate ruin probability for the case of regularly-varying-tailed claim sizes. Using the reflected random walk theory, Asmussen [7] extended the study to a larger class of heavy-tailed distributions; see Corollary  4.1(ii) of his paper. Complementary discussions on the ultimate ruin probability can be found in Kalashnikov and Konstantinides [8], Konstantinides et al. [9], Tang [10], among others.

In this paper, we are interested in the finite-time ruin probability. In this aspect, Tang [11] established an asymptotic result in the classical risk model: under the condition 𝐹∈𝒮, he obtained that for every 𝑇>0 for which 𝜆(𝑇)>0, Ψ(𝑥,𝑇)∼𝑇0−𝐹𝑥𝑒𝛿𝑡𝜆(d𝑡).(2.4) Recently, Wang [1] derived some important and interesting results in two independent risk models. One is the delayed renewal risk model, in which (2.4) holds if 𝐹∈𝒮; another is the general risk model, in which (2.4) also holds if 𝐹∈ℒ∩𝒟. We are interested in the latter, for example, the general risk model, and restate Theorem   2.2 of Wang [1] here.

Theorem 2.1. In the independent general risk model introduced in Section 1, assume that the claim sizes {𝑋𝑛,𝑛≥1} are independent and identically distributed nonnegative r.v.s with common distribution 𝐹∈ℒ∩𝒟. Assume that for any 𝑇>0 with 𝜆(𝑇)−𝜆(0)>0, there exists some constant 𝜂=𝜂(𝑇)>0 such that E(1+𝜂)𝑁(𝑇)<∞.(2.5) Then, (2.4) holds.

In the present paper, we aim to deal with the extended negatively dependent general risk model to get a similar result under 𝐹∈𝒟. Simultaneously, the condition (2.5) can be weakened to (2.8) below.

We call r.v.s {𝜉𝑛,𝑛≥1} are extended negatively dependent (END) if there exists some positive constant 𝑀 such that both P𝑛𝑘=1𝜉𝑘>𝑦𝑘≤𝑀𝑛𝑘=1P𝜉𝑘>𝑦𝑘P,(2.6)𝑛𝑘=1𝜉𝑘≤𝑦𝑘≤𝑀𝑛𝑘=1P𝜉𝑘≤𝑦𝑘(2.7) hold for each 𝑛≥1 and all 𝑦1,…,𝑦𝑛. This dependence structure was introduced by Liu [12]. Recall that r.v.s {𝜉𝑛,𝑛≥1} are called upper negatively dependent (UND) if (2.6) holds with 𝑀=1, they are called lower negatively dependent (LND) if (2.7) holds with 𝑀=1, and they are called negatively dependent (ND) if both (2.6) and (2.7) hold with 𝑀=1. These negative dependence structures were introduced by Ebrahimi and Ghosh [13] and Block et al. [14]. Clearly, ND r.v.s must be END r.v.s., and Example  4.1 of Liu [12] shows that the END structure also includes some other dependence structures.

Motivated by the work of Wang [1], under the END structure, we formulate our main result as follows.

Theorem 2.2. In the dependent general risk model introduced in Section 1, assume that the claim sizes {𝑋𝑛,𝑛≥1} are END nonnegative r.v.s with common distribution 𝐹∈𝒟 and finite mean 𝜇. Assume that for any 𝑇>0 with 𝜆(𝑇)−𝜆(0)>0, there exists some constant 𝑝>𝐽+𝐹 such that E(𝑁(𝑇))𝑝<∞.(2.8) Then, it holds that 𝐿𝐹𝑇0−𝐹𝑥𝑒𝛿𝑡𝜆(d𝑡)≲Ψ(𝑥,𝑇)≲𝐿𝐹−1𝑇0−𝐹𝑥𝑒𝛿𝑡𝜆(d𝑡).(2.9) Furthermore, if 𝐹∈ℒ∩𝒟, then (2.4) holds.

The rest of the present paper consists of two sections. We give some lemmas and the proof of Theorem 2.2 in Section 3. In Section 4, we perform some numerical calculations to verify the approximate relationship in our main result.

3. Proof of Main Result and Some Lemmas

In the sequel, 𝑀 and ğ‘Ž always represent some finite and positive constants whose values may vary in different places. In this section, we start by giving some lemmas to show some properties of the class 𝒟 and the END structure. The first lemma is a combination of Proposition  2.2.1 of Bingham et al. [4] and Lemma  3.5 of Tang and Tsitsiashvili [15].

Lemma 3.1. If a distribution 𝑉∈𝒟, then (i)for any 𝛾>𝐽+𝑉, there exist positive constants ğ‘Ž and 𝑏 such that 𝑉(𝑦)/𝑉(𝑥)â‰¤ğ‘Ž(𝑦/𝑥)−𝛾 holds for all 𝑥≥𝑦≥𝑏 and(ii) it holds for every 𝛾>𝐽+𝑉 that 𝑥−𝛾=𝑜(𝑉(𝑥)).

By direct verification, END r.v.s have the following properties similar to those of ND r.v.s; see Lemma  3.1 of Liu [12]. For some refined properties of END r.v.s, one can refer to Chen et al. [16]. The following lemma can also be found in Lemma  2.2 of Chen et al. [16].

Lemma 3.2. (i) If r.v.s {𝜉𝑛,𝑛≥1} are nonnegative and END, then for any 𝑛≥1, there exists some positive constant 𝑀 such that ∏E(𝑛𝑘=1𝜉𝑘∏)≤𝑀𝑛𝑘=1E𝜉𝑘.
(ii) If r.v.s {𝜉𝑛,𝑛≥1} are END and {𝑓𝑛(⋅),𝑛≥1} are either all monotone increasing or all monotone decreasing, then {𝑓𝑛(𝜉𝑛),𝑛≥1} are still END.

The following two lemmas play important roles in the proof of our main result.

Lemma 3.3. Let {𝜉𝑛,𝑛≥1} be identically distributed and END r.v.s with common distribution 𝑉 and 𝜇+𝑉=E𝜉1𝟏{𝜉1≥0}<∞. Then, for any 𝜃>0, 𝑥>0 and 𝑛≥1, there exists some positive constant 𝑀 such that P𝑛𝑘=1𝜉𝑘>𝑥≤𝑛𝑉(𝜃𝑥)+𝑀𝑒𝜇+𝑉𝑛𝑥𝜃−1.(3.1)

Proof. Following the proof of Lemma  2.3 of Tang [17], we employ a standard truncation argument to prove this lemma. For simplicity, we write 𝑆𝜉𝑛=∑𝑛𝑘=1𝜉𝑘, 𝑛≥1. If 𝜇+𝑉=0, then 𝜉𝑛 is almost surely nonpositive for each 𝑛≥1, implying P(𝑆𝜉𝑛>𝑥)=0 for any positive 𝑥, and thus (3.1) holds.
Let, in the following, 𝜇+𝑉>0. For any fixed 𝜃>0 and positive integer 𝑛, define ̃𝜉𝑛𝜉=min𝑛,̃𝜉,𝜃𝑥+𝑛̃𝜉=max𝑛,0=𝜉𝑛𝟏{0≤𝜉𝑛≤𝜃𝑥}+𝜃𝑥𝟏{𝜉𝑛>𝜃𝑥}.(3.2) According to Lemma 3.2(ii), {̃𝜉𝑛,𝑛≥1} and {̃𝜉+𝑛,𝑛≥1} are still END r.v.s, respectively. Denote 𝑆𝜉𝑛=∑𝑛𝑘=1̃𝜉𝑘, 𝑛≥1. Clearly, P𝑆𝜉𝑛𝑆>𝑥=P𝜉𝑛>𝑥,max1≤𝑘≤𝑛𝜉𝑘𝑆>𝜃𝑥+P𝜉𝑛>𝑥,max1≤𝑘≤𝑛𝜉𝑘≤𝜃𝑥≤𝑛𝑆𝑉(𝜃𝑥)+P𝜉𝑛.>𝑥(3.3) It remains to estimate the second summand in (3.3). For a positive ℎ, by Lemma 3.2(ii), {ğ‘’â„ŽÌƒğœ‰+𝑛,𝑛≥1} are END nonnegative r.v.s. Hence, using identity Eğ‘’â„ŽÌƒğœ‰+1=0ğœƒğ‘¥î€·ğ‘’â„Žğ‘¢î€¸î€·ğ‘’âˆ’1𝑉(d𝑢)+â„Žğœƒğ‘¥î€¸âˆ’1𝑉(𝜃𝑥)+1,(3.4) by Markov inequality and Lemma 3.2(i) we have P𝑆𝜉𝑛>ğ‘¥â‰¤ğ‘’âˆ’â„Žğ‘¥Eğ‘’â„Žî‚ğ‘†ğœ‰ğ‘›â‰¤ğ‘’âˆ’â„Žğ‘¥Eğ‘’â„Žâˆ‘ğ‘›ğ‘˜=1̃𝜉+ğ‘˜â‰¤ğ‘’âˆ’â„Žğ‘¥ğ‘€(Eğ‘’â„ŽÌƒğœ‰+1)𝑛=ğ‘€ğ‘’âˆ’â„Žğ‘¥î‚µî€œ0ğœƒğ‘¥î€·ğ‘’â„Žğ‘¢î€¸î€·ğ‘’âˆ’1𝑉(d𝑢)+â„Žğœƒğ‘¥î€¸âˆ’1𝑉(𝜃𝑥)+1𝑛.(3.5) Since 1+𝑢≤𝑒𝑢 for all 𝑢∈ℝ and (ğ‘’â„Žğ‘¢âˆ’1)/𝑢 is strictly increasing in 𝑢>0, from (3.5), we obtain P𝑆𝜉𝑛𝑛>𝑥≤𝑀exp0ğœƒğ‘¥ğ‘’â„Žğ‘¢âˆ’1𝑢𝑒𝑢𝑉(d𝑢)+ğ‘›â„Žğœƒğ‘¥î€¸âˆ’1𝑛𝑒𝑉(𝜃𝑥)âˆ’â„Žğ‘¥â‰¤ğ‘€expâ„Žğœƒğ‘¥âˆ’1𝜃𝑥0𝜃𝑥𝑢𝑉(d𝑢)+𝜃𝑥𝑛𝑒𝑉(𝜃𝑥)âˆ’â„Žğ‘¥â‰¤ğ‘€expâ„Žğœƒğ‘¥âˆ’1𝜇𝜃𝑥+𝑉.âˆ’â„Žğ‘¥(3.6) Choose ℎ=(𝜃𝑥)−1log(𝑥(𝜇+𝑉𝑛)−1+1), which is positive. For such ℎ, by (3.6), we have P𝑆𝜉𝑛1>𝑥≤𝑀exp𝜃−1𝜃𝑥log𝜇+𝑉𝑛1+1≤𝑀exp𝜃log𝑒𝜇+𝑉𝑛𝑥.(3.7) The last estimate and (3.3) imply the desired estimate (3.1). The lemma is proved.

Lemma 3.4. In the dependent general risk model introduced in Section 1, assume that the claim sizes {𝑋𝑛,𝑛≥1} are END nonnegative r.v.s with common distribution 𝐹∈𝒟. Let 𝑍 be an arbitrary nonnegative r.v. and assume that {𝑋𝑛,𝑛≥1}, {𝑁(𝑡),𝑡≥0} and 𝑍 are mutually independent. Then, for any 𝑇>0 and any positive integer 𝑛0, 𝐿𝐹𝑛0𝑘𝑘=1𝑗=1P𝑋𝑗𝑒−𝛿𝜏𝑗≲>𝑥,𝑁(𝑇)=𝑘𝑛0𝑘=1P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝜏𝑗>𝑥+𝑍,𝑁(𝑇)=𝑘≲𝐿𝐹𝑛−10𝑘𝑘=1𝑗=1P𝑋𝑗𝑒−𝛿𝜏𝑗.>𝑥,𝑁(𝑇)=𝑘(3.8) Furthermore, if 𝐹∈ℒ∩𝒟, then 𝑛0𝑘=1P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝜏𝑗∼>𝑥+𝑍,𝑁(𝑇)=𝑘𝑛0𝑘𝑘=1𝑗=1P𝑋𝑗𝑒−𝛿𝜏𝑗>𝑥,𝑁(𝑇)=𝑘.(3.9)

We remark that if 𝐹 is consistently varying tailed (see the definition in Chen and Yuen [18]), then by conditioning (3.9) easily follows from Theorem  3.2 of Chen and Yuen [18]. Note that this case is in a broader scope, since there is no need to assume independence between (𝜏1,…,𝜏𝑛0) and 𝑍.

Proof. We follow the line of the proof of Lemma  3.6 of Wang [1] with some modifications in relation to the properties of the class 𝒟 and the END structure. Clearly, for each 𝑘=1,…,𝑛0, P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝜏𝑗=>𝑥+𝑍,𝑁(𝑇)=𝑘{0≤𝑡1≤⋯≤𝑡𝑘≤𝑇,𝑡𝑘+1>𝑇}∞0−P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝑡𝑗>𝑥+𝑧×P𝑍∈d𝑧,𝜏1∈d𝑡1,…,𝜏𝑘+1∈d𝑡𝑘+1.(3.10) We first show the upper bound. For any fixed 𝑙>0, P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝑡𝑗>𝑥+𝑧≤P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝑡𝑗>𝑥+𝑧−𝑙+P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝑡𝑗>𝑥+𝑧,max1≤𝑗≤𝑘𝑋𝑗𝑒−𝛿𝑡𝑗≤𝑥+𝑧−𝑙∶=𝐼1+𝐼2.(3.11) By 𝐹∈𝒟, for any 0<𝜃<1 and each 𝑘=1,…,𝑛0, we have uniformly for all 𝑡1,…,𝑡𝑘∈[0,𝑇] and 𝑧∈[0,∞), 𝐼1≤𝑘𝑗=1𝐹𝜃(𝑥+𝑧)𝑒𝛿𝑡𝑗≲𝐿𝐹𝑘−1𝑗=1𝐹(𝑥+𝑧)𝑒𝛿𝑡𝑗,(3.12) by firstly letting ğ‘¥â†’âˆž then 𝜃↗1. We note that {𝑋𝑛,𝑛≥1} are END r.v.s. Then, by 𝐹∈𝒟, there exists some positive constant 𝑀=𝑀(𝑛0) such that for sufficiently large 𝑥, each 𝑘=1,…,𝑛0, all 𝑡1,…,𝑡𝑘∈[0,𝑇] and 𝑧∈[0,∞), 𝐼2=P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝑡𝑗>𝑥+𝑧,𝑥+𝑧𝑘<max1≤𝑗≤𝑘𝑋𝑗𝑒−𝛿𝑡𝑗≤𝑥+𝑧−𝑙≤P𝑘𝑖=1𝑗≠𝑖𝑋𝑗𝑒−𝛿𝑡𝑗>𝑙,𝑋𝑖𝑒−𝛿𝑡𝑖>𝑥+𝑧𝑘≤𝑘𝑖=1𝑗≠𝑖P𝑋𝑗𝑒−𝛿𝑡𝑗>𝑙𝑘−1,𝑋𝑖𝑒−𝛿𝑡𝑖>𝑥+𝑧𝑘≤𝑀𝑘𝑖=1𝑗≠𝑖𝐹𝑙𝑒𝛿𝑡𝑗𝑘−1𝐹(𝑥+𝑧)𝑒𝛿𝑡𝑖𝑘≤𝑀𝐹𝑙𝑛0−1𝑘𝑗=1𝐹(𝑥+𝑧)𝑒𝛿𝑡𝑗.(3.13) Since 𝑙 can be arbitrarily large, it follows that limsupğ‘™â†’âˆžlimsupğ‘¥â†’âˆžsup𝑡1,…,𝑡𝑘∈[][0,𝑇,𝑧∈0,∞)𝐼2∑𝑘𝑗=1𝐹(𝑥+𝑧)𝑒𝛿𝑡𝑗=0.(3.14) Hence, from (3.10)–(3.14), we obtain for each 𝑘=1,…,𝑛0, P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝜏𝑗>𝑥+𝑍,𝑁(𝑇)=𝑘≲𝐿𝐹𝑘−1𝑗=1{0≤𝑡1≤⋯≤𝑡𝑘≤𝑇,𝑡𝑘+1>𝑇}∞0−𝐹(𝑥+𝑧)𝑒𝛿𝑡𝑗×P𝑍∈d𝑧,𝜏1∈d𝑡1,…,𝜏𝑘+1∈d𝑡𝑘+1=𝐿𝐹𝑘−1𝑗=1P𝑋𝑗𝑒−𝛿𝜏𝑗>𝑥+𝑍,𝑁(𝑇)=𝑘≤𝐿𝐹𝑘−1𝑗=1P𝑋𝑗𝑒−𝛿𝜏𝑗.>𝑥,𝑁(𝑇)=𝑘(3.15) As for the lower bound for (3.10), since {𝑋𝑛,𝑛≥1} are END r.v.s, we have for sufficiently large 𝑥 and each 𝑘=1,…,𝑛0, P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝑡𝑗>𝑥+𝑧≥P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝑡𝑗≥>𝑥+𝑧𝑘𝑗=1𝐹(𝑥+𝑧)𝑒𝛿𝑡𝑗−1≤𝑖<𝑗≤𝑘P𝑋𝑖𝑒−𝛿𝑡𝑖>𝑥+𝑧,𝑋𝑗𝑒−𝛿𝑡𝑗≥>𝑥+𝑧𝑘𝑗=1𝐹(𝑥+𝑧)𝑒𝛿𝑡𝑗−𝑀1≤𝑖<𝑗≤𝑘𝐹(𝑥+𝑧)𝑒𝛿𝑡𝑖𝐹(𝑥+𝑧)𝑒𝛿𝑡𝑗=(1−𝑜(1))𝑘𝑗=1𝐹(𝑥+𝑧)𝑒𝑟𝑡𝑗(3.16) holds uniformly for all 𝑡1,…,𝑡𝑘∈[0,𝑇] and 𝑧∈[0,∞). By 𝐹∈𝒟 and Fatou's lemma, we have for any ̃𝜃>1 and all 𝑗=1,2,…, 1liminfP𝑋𝐹(𝑥)𝑗>𝑥+𝑍𝑒𝛿𝑇=liminf∞0−𝐹𝑥+𝑧𝑒𝛿𝑇≥𝐹(𝑥)P(𝑍∈d𝑧)∞0−liminf𝐹̃𝜃𝑥𝐹=(𝑥)P(𝑍∈d𝑧)𝐹∗̃𝜃⟶𝐿𝐹,̃𝜃↘1,(3.17) which means P𝑋𝑗>𝑥+𝑍𝑒𝛿𝑇≳𝐿𝐹𝐹(𝑥).(3.18) Similar to (3.15), from (3.10), (3.16), and (3.18), we obtain for each 𝑘=1,…,𝑛0, P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝜏𝑗≳>𝑥+𝑍,𝑁(𝑇)=𝑘𝑘𝑗=1P𝑋𝑗𝑒−𝛿𝜏𝑗≥>𝑥+𝑍,𝑁(𝑇)=𝑘𝑘𝑗=1{0≤𝑡1≤…≤𝑡𝑘≤𝑇,𝑡𝑘+1>𝑇}P𝑋𝑗>𝑥𝑒𝛿𝑡𝑗+𝑍𝑒𝛿𝑇P𝜏1∈d𝑡1,…,𝜏𝑘+1∈d𝑡𝑘+1≳𝐿𝐹𝑘𝑗=1{0≤𝑡1≤…≤𝑡𝑘≤𝑇,𝑡𝑘+1>𝑇}𝐹𝑥𝑒𝛿𝑡𝑗P𝜏1∈d𝑡1,…,𝜏𝑘+1∈d𝑡𝑘+1=𝐿𝐹𝑘𝑗=1P𝑋𝑗𝑒−𝛿𝜏𝑗.>𝑥,𝑁(𝑇)=𝑘(3.19) The desired relation (3.8) follows now from (3.15) and (3.19).
If 𝐹∈ℒ∩𝒟, (3.9) follows by using the properties of the class ℒ to establish analogies of relations (3.12) and (3.17). This ends the proof of the lemma.

Proof of Theorem 2.2. We use the idea in the proof of Theorem  2.2 of Wang [1] (e.g., Theorem 2.1 of this paper) to prove this result. Clearly, 𝐹∈𝒟 and 𝜇<∞ imply 𝐽+𝐹≥1. By (2.8), we have for any 𝜖>0, there exists some positive integer 𝑛1=𝑛1(𝑇,𝜖) such that E(𝑁(𝑇))𝑝𝟏{𝑁(𝑇)>𝑛1}≤𝜖.(3.20)
To estimate the upper bound of Ψ(𝑥,𝑇), we split it into two parts as Ψ(𝑥,𝑇)≤P𝑁(𝑇)𝑗=1𝑋𝑗𝑒−𝛿𝜏𝑗=⎛⎜⎜⎝>𝑥𝑛1𝑘=1+âˆžî“ğ‘˜=𝑛1+1⎞⎟⎟⎠P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝜏𝑗>𝑥,𝑁(𝑇)=𝑘∶=𝐼3+𝐼4.(3.21) According to Lemma 3.4 of this paper and Lemma  3.5 of Wang [1], we have for sufficiently large 𝑥, 𝐼3≤(1+𝜖)𝐿𝐹𝑛−11𝑘𝑘=1𝑗=1P𝑋𝑗𝑒−𝛿𝜏𝑗>𝑥,𝑁(𝑇)=𝑘≤(1+𝜖)ğ¿ğ¹âˆžâˆ’1𝑗=1P𝑋𝑗𝑒−𝛿𝜏𝑗>𝑥,𝑁(𝑇)≥𝑗=(1+𝜖)ğ¿ğ¹âˆžâˆ’1𝑗=1P𝑋𝑗𝑒−𝛿𝜏𝑗>𝑥,𝜏𝑗=≤𝑇(1+𝜖)𝐿𝐹−1𝑇0−𝐹𝑥𝑒𝛿𝑡𝜆(d𝑡).(3.22) By Lemma 3.3, 𝐹∈𝒟, Lemma 3.1(ii), (3.20), and 𝑝>𝐽+𝐹≥1, there exists some positive constant 𝑀 such that for sufficiently large 𝑥, 𝐼4â‰¤âˆžî“ğ‘˜=𝑛1+1P𝑘𝑗=1𝑋𝑗≤>𝑥P(𝑁(𝑇)=𝑘)𝐹𝑝−1ğ‘¥î€¸âˆžî“ğ‘˜=𝑛1+1𝑘P(𝑁(𝑇)=𝑘)+𝑀(𝑒𝜇)ğ‘ğ‘¥âˆžâˆ’ğ‘î“ğ‘˜=𝑛1+1𝑘𝑝P(𝑁(𝑇)=𝑘)≤𝑀𝐹(𝑥)E𝑁(𝑇)𝟏{𝑁(𝑇)>𝑛1}+E(𝑁(𝑡))𝑝𝟏{𝑁(𝑇)>𝑛1}=𝑀𝜖𝐹(𝑥).(3.23) By Lemma 3.1(i), for any 𝛾>𝐽+𝐹, there exists some positive constant ğ‘Ž such that for sufficiently large 𝑥, 𝑇0−𝐹𝑥𝑒𝛿𝑡𝜆(d𝑡)â‰¥ğ‘Žâˆ’1𝐹(𝑥)𝑇0−𝑒−𝛾𝛿𝑡𝜆(d𝑡)â‰¥ğ‘Žâˆ’1𝑒−𝛾𝛿𝑇(𝜆(𝑇)−𝜆(0))𝐹(𝑥),(3.24) which, combining (3.23) and 𝜆(𝑇)−𝜆(0)>0, implies 𝐼4≤𝑀𝜖𝑇0−𝐹𝑥𝑒𝛿𝑡𝜆(d𝑡).(3.25) From (3.21), (3.22), and (3.25), we derive the right-hand side of (2.9).
As for the lower bound of Ψ(𝑥,𝑇), by Lemma 3.4, we have for the above given 𝜖>0 and sufficiently large 𝑥, Ψ(𝑥,𝑇)≥P𝑁(𝑇)𝑗=1𝑋𝑗𝑒−𝛿𝜏𝑗>𝑥+𝑇0𝑒−𝛿𝑠≥𝐶(d𝑠)𝑛1𝑘=1P𝑘𝑗=1𝑋𝑗𝑒−𝛿𝜏𝑗>𝑥+𝑇0𝑒−𝛿𝑠𝐶(d𝑠),𝑁(𝑇)=𝑘≥(1−𝜖)𝐿𝐹𝑛1𝑘𝑘=1𝑗=1P𝑋𝑗𝑒−𝛿𝜏𝑗>𝑥,𝑁(𝑇)=𝑘=(1−𝜖)ğ¿ğ¹âŽ›âŽœâŽœâŽâˆžî“ğ‘—=1P𝑋𝑗𝑒−𝛿𝜏𝑗>𝑥,ğœğ‘—î€¸âˆ’â‰¤ğ‘‡âˆžî“ğ‘˜=𝑛1𝑘+1𝑗=1Pî€·ğ‘‹ğ‘—ğ‘’âˆ’ğ›¿ğœğ‘—î€¸âŽžâŽŸâŽŸâŽ >𝑥,𝑁(𝑇)=𝑘∶=(1−𝜖)𝐿𝐹𝑇0−𝐹𝑥𝑒𝛿𝑡𝜆(d𝑡)−𝐼5.(3.26) Analogously to the estimate for 𝐼4, we have for sufficiently large 𝑥, 𝐼5≤𝐹(𝑥)E𝑁(𝑇)𝟏{𝑁(𝑇)>𝑛1}≤𝑀𝜖𝑇0−𝐹𝑥𝑒𝛿𝑡𝜆(d𝑡).(3.27) From (3.26) and (3.27), we obtain the left-hand side of (2.9).
If 𝐹∈ℒ∩𝒟, then (2.4) follows by using (3.9) in the proof of (3.22) and (3.26).

4. Numerical Calculations

In this section, we perform some numerical calculations to check the accuracy of the asymptotic relations obtained in Theorem 2.2. The main work is to estimate the finite-time ruin probability defined in (1.2).

We assume that the claim sizes {𝑋𝑛,𝑛≥1} come from the common Pareto distribution with parameter 𝜅=1, 𝛽=2, 𝜅𝐹(𝑥;𝜅,𝛽)=1−𝜅+𝑥𝛽,𝑥≥0,(4.1) which belongs to the class ℒ∩𝒟, and {(𝑋2𝑛−1,𝑋2𝑛),𝑛≥1} are independent replications of (𝑋1,𝑋2) with the joint distribution 𝐹𝑋1,𝑋21(𝑥,𝑦)=−𝛼𝑒ln1+−𝛼𝐹(𝑥)𝑒−1−𝛼𝐹(𝑦)−1𝑒−𝛼−1,(4.2) with parameter 𝛼=1, where the joint distribution 𝐹𝑋1,𝑋2(𝑥,𝑦) is constructed according to the Frank Copula. It has been proved in Example  4.2 of Liu [12] that 𝑋1 and 𝑋2 are END r.v.s. Since {(𝑋2𝑛−1,𝑋2𝑛),𝑛≥1} are independent copies of (𝑋1,𝑋2), the r.v.s {𝑋𝑛,𝑛≥1} are END as well.

Assume that the claim arrival process 𝑁(𝑡) is the homogeneous Poisson process with intensity parameter 𝜆. Clearly, such an integer-valued process 𝑁(𝑡) satisfies the condition (2.8). Choose 𝜆=0.1. The total amount of premiums is simplified as 𝐶(𝑡)=𝑐𝑡 with the premium rate 𝑐=500, and the constant interest rate 𝛿=0.02. Here, we set the time 𝑇 as 𝑇=10 and the initial capital reserve 𝑥=500,103,2×103,5×103, respectively. We aim to verify the accuracy of relation (2.4). The procedure of the computation of the finite-time ruin probability Ψ(𝑥,𝑇) in Theorem 2.2 is listed here.

Step 1. Assign a value for the variable 𝑥 and set 𝑙=0.

Step 2. Divide the close interval [0,𝑇] into 𝑚=1000 pieces, and denote each time point as 𝑡𝑖, 𝑖=1,…,𝑚.

Step 3. For each 𝑡𝑖, generate a random number 𝑛𝑖 from the Poisson distribution 𝑃(𝜆𝑡𝑖), and set 𝑛𝑖 as the sample size of the claims.

Step 4. Generate the accident arrival time {𝜏𝑖𝑘,𝑘=1,…,𝑛𝑖} from the uniform distribution 𝑈(0,𝑡𝑖) and the claim sizes {𝑋𝑖𝑘,𝑘=1,…,𝑛𝑖} from (4.1) and (4.2).

Step 5. Calculate the expression 𝐷 below for each 𝑡𝑖 and denote them as {𝐷𝑖,𝑖=1,…,𝑚}: 𝐷𝑖=𝑛𝑖𝑘=1𝑋𝑖𝑘𝑒−𝑟𝜏𝑖𝑘−𝑡𝑖0𝑒−𝑟𝑠𝐶(d𝑠),𝑖=1,…,𝑚,(4.3) where 𝑟 and 𝐶(𝑡) have been defined and their values have also been assigned.

Step 6. Select the maximum value from {𝐷𝑖,𝑖=1,…,𝑚}, and denote it as 𝐷∗, compare 𝐷∗ with 𝑥; if 𝐷∗>𝑥, then the value of 𝑙 increases 1.

Step 7. Repeat Step 2 through Step 6, 𝑁=109 times.

Step 8. Calculate the moment estimate of the finite-time ruin probability, 𝑙/𝑁.

Step 9. Repeat Step 1 through Step 8 ten times and get ten estimates. Then, choose the median of the ten estimates as the analog value of the finite-time ruin probability.

For different value of 𝑥, the analog value and the theoretical result of the finite-time ruin probability are presented in Table 1, and the percentage of the error relative to the theoretical result is also presented in the bracket behind the analog value. It can be found that from Table 1, the larger 𝑥 becomes, the smaller the difference between the analog value and the theoretical result is. Therefore, the approximate relationship in Theorem 2.2 is reasonable.


The authors would like to thank the two referees for their useful comments on an earlier version of this paper. The revision of this work was finished during a research visit of the first author to Vilnius University. He would like to thank the Faculty of Mathematics and Informatics for its excellent hospitality.  Research supported by National Natural Science Foundation of China (no. 11001052), China Postdoctoral Science Foundation (20100471365), National Science Foundation of Jiangsu Province of China (no. BK2010480), Natural Science Foundation of the Jiangsu Higher Education Institutions of China (no. 09KJD110003), Postdoctoral Research Program of Jiangsu Province of China (no. 0901029C), and Jiangsu Government Scholarship for Overseas Studies, Qing Lan Project.