Ruin Probabilities in a Dependent Discrete-Time Risk Model With Gamma-Like Tailed Insurance Risks

This paper considered a dependent discrete-time risk model, in which the insurance risks are represented by a sequence of independent and identically distributed real-valued random variables with a common Gamma-like tailed distribution; the financial risks are denoted by another sequence of independent and identically distributed positive random variables with a finite upper endpoint, but a general dependence structure exists between each pair of the insurance risks and the financial risks. Following the works of Yang and Yuen in 2016, we derive some asymptotic relations for the finite-time and infinite-time ruin probabilities. As a complement, we demonstrate our obtained result through a Crude Monte Carlo (CMC) simulation with asymptotics.


Introduction
Consider a discrete-time risk model, where, for every i ≥ 1, the insurer's net loss (the aggregate claim amount minus the total premium income) within period i is represented by a real-valued random variable (r.v.) X i ; and the stochastic discount factor from time i to time i − 1 is denoted by a positive r.v.Y i .In the terminology in [1], {X i , i ≥ 1} and {Y i , i ≥ 1} are called the insurance risks and financial risks, respectively.Throughout this paper, we suppose that {X i , i ≥ 1} is a sequence of independent and identically distributed (i.i.d.) real-valued r.v.s with a common distribution F; {Y i , i ≥ 1} is another sequence of i.i.d.positive r.v.s with a common distribution G; and the random vectors {(X i , Y i ), i ≥ 1} are independent copies of (X, Y) following a certain dependence structure (see (7) below).In this framework, we are interested in the following quantities: with the maxima where S n denotes the stochastic discounted value of aggregate net losses within time n.
Then, the two-tail probabilities P(M n > x) and P(M ∞ > x) can be interpreted as the finite-time ruin probability within period n and the infinite-time ruin probability, respectively, where x ≥ 0 stands for the initial wealth of the insurer.Clearly, as n → ∞.The right-hand side converges almost surely (a.s.), if −∞ ≤ E ln Y < 0 and E ln max{X, 1} < ∞ (see Theorem 1.6 in [2] and Theorem 1 in [3]).Thus, M n converges a.s. to the limit M ∞ , which has a proper distribution on [0, ∞).In this paper, we aim to investigate the asymptotic behavior of the tail probabilities P(S n > x), P(M n > x) and P(M ∞ > x) as x → ∞.
In such a discrete-time risk model, under independence or some certain dependence assumptions imposed on X i s and Y i s, the asymptotic tail behavior of M n and M ∞ has been extensively studied by many researchers.Notice that the assumption of complete independence is for mathematical convenience, but appears unrealistic in most practical situations.A recent new trend of study is to introduce various dependence structures to describe the insurance and financial risks {X i , i ≥ 1} and {Y i , i ≥ 1}.One trend is to require the insurance risks {X i , i ≥ 1} to obey a certain dependence structure (see [4][5][6] among others).Another trend is to assume that {(X i , Y i ), i ≥ 1} form a sequence of i.i.d.random vectors, but for each i ≥ 1, some certain dependence structure exists between X i and Y i .Such a work was initially studied by [7].Chen considered the discrete-time risk model, in which each pair of the insurance and financial risks form the bivariate Farlie-Gumbel-Morgenstern (FGM) distribution.Later, Yang and Konstantinides extended Chen's results in [5], by considering a more general dependence structure than the FGM one.They derived the uniform estimates for the finite-time and infinite-time ruin probabilities, under the assumption that {X i , i ≥ 1} are of consistent variation.For more details, one can be refereed to [8][9][10][11] among others.We remark that all of the above works are studied in the heavy-tailed case, while, in this paper, we consider the light-tailed case, that is, the insurance risks are Gamma-like tailed, thus are light-tailed.
Throughout the paper, all limit relationships hold for x tending to ∞ unless stated otherwise.For two positive functions f (x) and g(x), we write f (x)∼g(x) if lim f (x)/g(x) = 1; write f (x) ≺ g(x) or g(x) f (x) if lim sup f (x)/g(x) ≤ 1; and write f (x) = o(g(x)) if lim f (x)/g(x) = 0.For two real-valued numbers x and y, denote by x ∨ y = max{x, y}, x ∧ y = min{x, y} and denote the positive part of x by x + = x ∨ 0. The indicator function of an event A is denoted by 1 A .
A distribution F on R is said to be Gamma-like tailed with shape parameter α > 0 and scale parameter γ > 0 if there exists a slow function l(•) : (0, ∞) → (0, ∞) such that (see [12,13]).A larger distribution class is that of generalized exponential distributions.A distribution F on R is said to belong to the class L(γ) with γ ≥ 0, if for any y ∈ R, Clearly, if γ = 0, the class L(0) consists of all long-tailed distributions, which are heavy-tailed.If γ > 0, then all distributions in the class L(γ) are light-tailed.A class larger than the generalized exponential distribution class L(γ) is that of rapidly varying tailed distributions.A distribution F on R is said to be rapidly varying tailed, denoted by F ∈ R −∞ , if F(xy) = o(F(x)) holds for all y > 1.For a distribution F ∈ R −∞ , from Theorem 1.2.2 of [14], it can be seen that for any > 0 and δ > 0, there exists a sufficiently large constant D > 0 such that for all y ≥ x ≥ D, We remark that if a distribution F is Gamma-like tailed with α > 0 and γ > 0, then F ∈ L(γ) ⊂ R −∞ holds.
In the case of light-tailed insurance risks, in [15,16] Tang and his coauthor first established some asymptotic formulas for the finite-time ruin probability P(M n > x) under the independence structure, and the conditions that the insurance risks X i s have a common convolution-equivalent or a rapidly varying tail, and the financial risks Y i s have a common distribution G with a finite upper endpoint Precisely speaking, consider a discrete-time risk model, in which {X i , i ≥ 1} and {Y i , i ≥ 1} are two sequences of i.i.d.r.v.s with common distributions F and G, respectively, and these two sequences are mutually independent.If F ∈ R −∞ and G have a finite upper endpoint y * > 1 with p * = P(Y 1 = y * ) > 0, then, for each fixed n ≥ 1, Recently, in [17] Yang and Yuen derived some more precise results than relation (6) in the presence of Gamma-like tailed insurance risks, under the independence structure or a certain dependence structure, where each pair of the insurance risks and the financial risks follow a bivariate Sarmanov distribution (see the definition below).They investigated the asymptotic tail behavior of S n , M n and M ∞ in three cases of 0 < y * < 1, y * = 1 and y * > 1, respectively, and dropped the condition p * > 0.
In this paper, we restrict ourselves to the framework in which a more general dependence structure exists between each pair of the insurance risks and the financial risks.Precisely speaking, the random vectors of the insurance and financial risks {(X i , Y i ), i ≥ 1} are assumed to be i.i.d.copies of a generic pair (X, Y) with the dependent components X and Y fulfilling the relation holding uniformly for all y ∈ (0, y * ] as x → ∞, i.e., lim Here, h(•) : [0, ∞) → (0, ∞) is a positive measurable function, and if y is not a possible value of Y, then the left side of relation ( 7) consists of the unconditional probability; thus, h(y) equals to 1.Such a dependence structure (7) was introduced by [18], which contains many commonly-used bivariate copulas, such as the Ali-Mikhail-Haq copula, the Farlie-Gumbel-Morgenstern copula, and the Frank copula among others, and allows both positive and negative dependence structures.We remark that if P(X > x) > 0 for all x ∈ R, then relation (7) leads to Eh(Y) = 1.See [19][20][21] for more details on such a dependence structure.In particular, if X and Y follow a bivariate Sarmanov distribution defined by and assume that the limit lim x→∞ φ 1 (x) = d 1 exists, then it can be directly verified that relation ( 7) is satisfied with h(y) = 1 + θd 1 φ 2 (y).
Motivated by [17], in this paper, we aim to study the asymptotic tail relations of S n , M n and M ∞ in two cases, i.e., 0 < y * < 1 and y * = 1, under the assumption that {(X i , Y i ), i ≥ 1} are i.i.d.random vectors with dependent components fulfilling relation (7).In the case 0 < y * < 1, we also obtain a uniform result for both finite-time and infinite-time ruin probabilities, by considering the asymptotic formulas for P(M n > x) and P(M ∞ > x).We still restrict the insurance risks to be Gamma-like tailed.Our obtained results essentially extend the corresponding ones in [17].
The rest of this paper is organized as follows.In Section 2, the main results of the present paper are provided, and Section 3 displays their proofs.In Section 4, we perform a simulation to verify the approximate relationships in the main results by using the Crude Monte Carlo (CMC) method.

Main Results
Denote by p * = P(Y 1 = y * ) ≥ 0, which can be equal to 0. The first result investigates the asymptotics for the finite-time ruin probability in two cases of 0 < y * < 1 and y * = 1.Theorem 1.Consider the above-mentioned discrete-time risk model, where {(X i , Y i ), i ≥ 1} are a sequence of i.i.d.random vectors with a generic pair (X, Y) satisfying relation (7) uniformly for all y ∈ (0, y * ].Assume that F is Gamma-like tailed with shape parameter α > 0 and scale parameter γ > 0 defined in relation (3), G has a finite upper endpoint y * defined in relation (5), and inf y∈(0,y * ] h(y) > 0.
The second result gives formula (9) with n = ∞, which implies the asymptotic relation for the infinite-time ruin probability.Combining relation (9), a uniform result for both finite-time and infinite-time ruin probabilities is also derived in the case 0 < y * < 1.
Theorem 2. Under the conditions of Theorem 1, if 0 < y * < 1, then it holds that Furthermore, relation (9) holds uniformly for all n ≥ 1, that is,

Proofs of Main Results
We start this section by the following lemma, which is initiated by Lemma 2 in [22], where they considered that X and Y are two independent r.v.s and Y is supported on (0, 1], indicating y * = 1.Lemma 1.Let X and Y be two dependent r.v.s with distributions F on R and G on (0, y * ], respectively.If F ∈ R −∞ and relation (7) holds uniformly for all y ∈ (0, y * ], then it holds that Proof of Lemma 1.As done in the proof of Lemma 2 in [22], for any x > 0 and any z ∈ (0, y * ), we have that For K 3 , by relation (7), we have that ∼ P(X > xy −1 * )h(y * )P(Y = y * ).
By Theorem 1.5.6 (i) in [23], for any δ > 0 and sufficiently large x, we have that is a slowly varying function, according to the dominated convergence theorem, we have that Again by the slow variety of l(•) and the the induction assumption, it holds that Plugging relations (21) and (22) into relation (21), by 0 < (1 + )y * < 1, we obtain that We next deal with I 3 .According to relation (3), F ∈ L(γ) and the dominated convergence theorem, we have that As for I 2 , by the induction assumption we have that Thus, by noting 0 < y * < 1, we derive from relations ( 19) and ( 23)-(25) that Now, we shall show that if X m+1 and Y m+1 are dependent according to relation (7) holding uniformly for all y ∈ (0, y * ], then T m + X m+1 and Y m+1 follow the same dependence structure.Clearly, for all y ∈ (0, y * ], since (X m+1 , Y m+1 ) is independent of T m , we have that where a(x) < x is any infinitely increasing function.For L 1 , by relations ( 7) and (26), it holds uniformly for all y ∈ (0, y * ] that As for L 2 , for the above sufficiently small 0 < < 1 satisfying 0 < (1 + )y * < 1, choose a(x) = (1 + ) −1 x.Then, by the induction assumption and the slow variety of l(•), we have that Plugging relations ( 29) and (29) into relation ( 27), together with inf y∈(0,y * ] h(y) > 0, yields that uniformly for all y ∈ (0, y * ], For the lower bound, by relations ( 26)-( 29), we have that uniformly for all y ∈ (0, y * ], (31) Thus, by relations (26), ( 30) and (31), we derive that uniformly, for all y ∈ (0, y * ], which means that T m + X m+1 and Y m+1 follow the dependence structure defined in relation (7).Therefore, according to Lemma 1, we can obtain from relation (26) that which also implies that E e γT m+1 < ∞ and F T m+1 ∈ L(γ/y * ).
We next prove relation (9).Introduce a Markov chain {W n , n ≥ 1} satisfying By using the identity and by Theorem 2.1 in [15], similarly to relation (18), we have that Obviously, the relation (34) holds for n = 1.Following the same line of the proof of relation ( 8), we obtain that which coincides with relation (9).Therefore, we complete the proof of claim (1).
For claim (2), according to the above proof, it suffices to show We proceed again by induction on n.Trivially, relation (35) holds for n = 1 by Lemma 1. Assume that relation (35) holds for n = m for some integer m ≥ 1, which implies that F T m ∈ L(γ).We aim to prove that relation (35) holds for n = m + 1.As done in relation (19), we still divide the tail probability P(T m + X m+1 > x) into three parts, denoted by I 1 , I 2 , and I 3 , respectively.Starting from I 1 , we construct two independent positive conditional r.v.s X c m+1 = (X m+1 | X m+1 > 0) and T c m = (T m | T m > 0), whose tail distributions, by relation (3) and the induction assumption satisfy and Similarly to the proof of relation (4.13) in [17], we have that As for I 2 , by the induction assumption, we have that Proof of Theorem 2. We firstly prove the asymptotic relation (11).For each n ≥ 0, define nonnegative r.v.s Thus, by relation ( 2), for all n ≥ 1, we have that By ln y * < 0, E ln(X 1 ∨ 1) < ∞ and Theorem 1.6 in [2], we obtain that M n converges a.s. to a limit M ∞ as n → ∞.Then, in order to verify relation (11), we need to prove that E e γM ∞ < ∞.As done in [17], we introduce a nonnegative r.v.Z, which is independent of {(X i , Y i ), i ≥ 1}, with the tail distribution We further construct a nonnegative conditional r.v.Z c = (Z | Z > x 0 ).It has been proved by [17] that, for every , where the symbol d ≤ denotes 'stochastically not larger than', that is, for all x ≥ 0, Letting n → ∞, we have that According to the same method of [17], by relation (46), the dominated convergence theorem and the Jensen's inequality, for any > 0 with 0 < (1 + )y * < 1 and arbitrarily fixed ȳ ∈ ( 3 √ y * , 1) (implying y * < ȳ3 < 1), we can choose a sufficiently large integer n 0 ≥ 3 such that and For the upper bound of relation (11), By relation (9) in Theorem 1, relations (47), (48) and the dominated convergence theorem, we have that (51) As for J 2 , it is dealt with along the same line of that in [17], we have that Plugging relations (51) and (52) into relation (51), we obtain that For the lower bound of relation (11), we derive from relation (9) in Theorem 1 and relation (47 Therefore, the desired relation (11) can be obtained by the arbitrariness of > 0. The second part of Theorem 2 can be dealt with the standard argument, which was also shown in [17].This ends the proof of Theorem 2.

Simulation Study
In this section, we conduct a simulation study through the software MATLAB R2014a (The MathWorks, Inc., Natick, MA, USA) to verify the asymptotic relation (10) for the finite-time ruin probability in the main theoretical result Theorem 1.
Assume that {(X i , Y i ), i ≥ 1} is a sequence of i.i.d.random vectors with a generic pair (X, Y), where the insurance risk X follows a common exponential distribution with parameters λ > 0, µ ∈ R We assume X and Y follow a bivariate FGM distribution with parameter |δ| ≤ 1.Note that the bivariate FGM distribution is included by the dependence structure defined in relation (7).
The following algorithm is used to generate the component r.v.s X and Y of i.i.d.random vectors {(X i , Y i ), i ≥ 1}, fulfilling the bivariate FGM distribution: Step a: Generate two i.i.d.r.v.s u and v following the uniform distribution on (0, 1); Step b: Set X = − log(1 − u)/λ + µ; Step Thus, the generated (X, Y) returns the outcome of two dependent r.v.s.fulling the bivariate FGM distribution.
We use the CMC method to perform the simulation.The computation procedure of the estimation of the theoretical finite-time ruin probability ψ(x; n) is listed as the following: Step 1: Assign a value for the variate x and set m = 0; Step 2: Generate the i.i.d.r.v.s random vectors {(X i , Y i ), i ≥ 1} satisfying the certain dependence structure according to the above algorithm; Step 3: Calculate the vector (Y 1 , Y 1 Y 2 , . . ., ∏ n j=1 Y j ) T .Then, S n is equal to the product of the two vectors (X 1 , X 2 , . . ., X n ) and (Y 1 , Y 1 Y 2 , . . ., ∏ n j=1 Y j ) T ; Step 4: Select the maximum value from S 1 , . . ., S n and denote it by M n , and compare M n with x: if M n > x; then, m = m + 1; Step 5: Repeat step 2 to step 4 for N times; Step 6: Calculate ψ 1 (x; n) = m/N as the estimate of ψ(x; n).
Step 7: Repeat step 1 to step 6 for l times, and set the mean of all ψ 1 (x; n) as the final estimate.
All parameters are set as: λ = 0.1, µ = −16, δ = 1, n = 8, N = 10, 000, 000 and l = 20.We set the initial asset x from 100 to 150 and compare the simulated estimate values with the asymptotic values of the finite-time ruin probability in the following Table 1.The standard error of each estimate computed via the CMC method is presented in the bracket behind the estimate.Without surprise, the larger the initial wealth x is, the smaller both the simulated estimate values and the asymptotic values of the ruin probability become, but the more fluctuation their ratio exhibits, the less effective the estimates are.In fact, this is due to the poor performance of the CMC method, which requires a sufficiently large sample size to meet the demands of high accuracy.In order to eliminate the influence of large initial wealth, we repeat the simulation with the sample size N increasing from 10, 000, 000 up to 15, 000, 000.A significant improvement is observed.See Table 2 below.

Conclusions
In this paper, we study the asymptotics for finite-time and infinite-time ruin probabilities.We conduct our study in a discrete-time risk model, in which the insurance risks have a common Gamma-like tail the financial risks all have a finite upper endpoint, but a conditionally tailed asymptotic dependence structure exists between each pair of them.We demonstrate through simulations that the approximate relationships obtained in our main results are reasonable.

=
where d represents equality in distribution.Clearly, T n d = S n , n ≥ 1; thus, in order to prove relation (8), we only need to verify the relation