The moments of the time of ruin in Sparre Andersen risk models

Abstract We derive formulae for the moments of the time of ruin in both ordinary and modified Sparre Andersen risk models without specifying either the inter-claim time distribution or the individual claim amount distribution. We illustrate the application of our results in the special case of exponentially distributed claims, as well as for the following ordinary models: the classical risk model, phase-type(2) risk models, and the Erlang( 
$\mathscr{n}$
 ) risk model. We also show how the key quantities for modified models can be found.


Introduction
We start with a description of an ordinary Sparre Andersen (SA) model, sometimes called a renewal risk model, which was introduced by Sparre Andersen (1957). Claims occur according to a renewal process {N(t)} t≥0 , with inter-claim times being a sequence of independent and identically distributed (i.i.d.) random variables. Specifically, let V 1 be the time until the first claim, and for j = 2, 3, 4, . . . let V j be the time between claims j − 1 and j. We assume that {V j } ∞ j=1 are positive random variables with distribution function K = 1 −K, density function k, and a finite mean μ V . Individual claim amounts are modelled as a sequence of i.i.d. random variables {X j } ∞ j=1 which are also positive, with a density function which we denote f , and a finite mean μ X .
See, for example, Willmot & Woo (2017). The solutions of this equation are key in most of our applications. The modified SA model differs in only one respect from the ordinary model -the distribution of V 1 is different to that of {V j } ∞ j=2 with density function k 1 = k. In the special case when the distribution of V 1 is the equilibrium distribution of K, so that k 1 = k e =K/μ V , we have the equilibrium SA model. For the modified and equilibrium models, we introduce superscripts m and e, respectively, to functions defined for the ordinary model, so, for example, ψ e (u) is the probability of ultimate ruin for the equilibrium model.
Moments of the time of ruin have been studied for a large number of years. Gerber (1979) gives the first moment of the time of ruin in the classical risk model with exponential claims, while Drekic & Willmot (2003) found an easily implemented explicit formula for any moment of the time of ruin in the same setting. A recursion formula for moments of the time of ruin in the classical risk model was first given by Lin & Willmot (2000) who derived their result through analysis of a defective renewal equation; a much simpler proof of their result was later given by Albrecher & Boxma (2005). Dickson & Hipp (2001) considered moments of the time of ruin in an ordinary SA model with Erlang(2) inter-claim times. They showed how moments of T u could be found from explicit solutions for a Gerber-Shiu function defined as the Laplace transform of T u , and they also derived general expressions for the mean and variance of T 0 . In a pair of papers, Lee & Willmot (2014, 2016 find expressions satisfied by the moments of the time of ruin in dependent SA models with particular assumptions. Independent models can be retrieved as special cases, and their results provide an alternative method of obtaining some of the explicit results obtained below for ordinary SA models, but not for modified models. Yu et al. (2010) obtain results for the moments of the time of ruin for a Markovian risk model. Kim & Willmot (2016) show that the Laplace transform of the time of ruin in a modified SA model satisfies a defective renewal equation, but they do not give any results on moments of the time of ruin.
In this contribution, we take a completely different approach to the works mentioned above. Without making any distributional assumptions, we obtain new general expressions for the moments of the time of ruin in ordinary and modified SA models, and we apply these results in a number of settings. The results presented for modified SA models cannot be obtained from existing results in the literature, and although the results for ordinary SA models can, we believe that our approach is more straightforward to apply than that of Lee & Willmot (2016). Our proofs are certainly simpler. We also identify and implement an approach to finding moments of the time of ruin in an equilibrium SA model for which k is the Erlang(n) density.
In the next section, we outline some preliminaries and then give the main results in section 3. We then consider the application of these results, starting with the special case of exponentially distributed individual claim amounts in section 4. The classical risk model is considered in section 5, the phase-type(2) risk model in section 6, and the Erlang(n) risk model in section 8. In section 7, we illustrate an approach to modified SA models.

Preliminaries
In this section, we obtain some results for Laplace transforms which we will apply in the proofs of Theorems 3.1 and 3.2, and it is convenient to deal with the ordinary and modified models separately. Throughout, we use the notationã to denote the Laplace transform of a function a, with a second tilde when we consider a bivariate Laplace transform.

The ordinary SA model
By considering whether or not the surplus level ever falls below u, we have Willmot & Lin., 2001), so that giving (see Dickson, 2008), and we define the bivariate Laplace transform of w(u, t) as to be the bivariate Laplace transform of w(0, y, t) and defining We next consider derivatives of these transforms with respect to δ as well as the transforms evaluated at δ = 0. First, We note that ν 0 (y) = g(0, y).

The modified SA model
For the modified SA model, we need a smaller number of preliminary results. Using the arguments that underpin equation (3), we find that We then have

The Main Results
We now derive expressions for moments of the time of ruin for ordinary and modified SA models in Theorems 3.1 and 3.2, respectively. Theorem 3.1. For n = 1, 2, 3, . . ., the nth moment of the time of ruin for the ordinary SA risk model is given by Proof. Differentiate equation (4) n times with respect to δ using Liebniz rule (e.g. Abramowitz & Stegun, 1965) where A n (u) is as defined in the statement of the theorem. Then, using (2) to replace 1/(1 −g(0, s)) in (10) we have for n = 1, 2, 3, . . .
Corollary 3.1. When the initial surplus is 0, for n = 0, 1, 2, . . . we have Theorem 3.2. For n = 1, 2, 3, . . ., the nth moment of the time of ruin for the modified SA risk model is given by Proof. Differentiate equation (7) n times with respect to δ using Liebniz rule to get and setting δ = 0 we get To prove Theorems 3.1 and 3.2, we did not need to make any distributional assumptions. However, our third result applies only when the equilibrium distribution of K is of a particular form.
denote the nth moment of the time of ruin for a modified SA process which has κ i as the density function of the time to the first claim for i = 1, 2, 3, . . . , r. Then for n = 0, 1, 2, . . .
If there is value of i such that κ i (t) = k(t), then we replace ψ m i n (u) with ψ n (u).
We remark that we do not require positive values for each p i . Proof. By conditioning on the time and the amount of the first claim, we have and hence by (6) Thus, for n = 0, 1, 2, . . .
In order to obtain explicit solutions from these results for particular distributions for interclaim times and individual claim amounts, it is clear that we need to know something about w(0, y, t) and w m (0, y, t), but as we shall see in the following sections, we do not need explicit formulae.

SA Models with Exponential Claims
We first consider any ordinary SA model with exponential individual claims, that is, with f (x) = αe −αx , x > 0, since in this case we have the same general form of both the Laplace transform of T u and ψ(u) regardless of the form of k whenever the negative solution of Lundberg's fundamental equation exists. Specifically, where −R(δ) is the negative solution of Lundberg's fundamental equation. See Willmot & Woo (2017). The special case δ = 0 gives ψ(u) = ψ(0)e −Ru where R( = R(0)) is the adjustment coefficient.
One way to find the moments of T 0 is to set u = 0 in (12) and to note that and we can find the derivatives of R(δ) from Lundberg's fundamental equation given the distribution of V 1 . For example, using formulae given by Dickson & Hipp (2001) for the first two derivatives of R(δ) evaluated when δ = 0 for the Erlang(2) risk model, we find that ψ 1 (0) = 4.0744 and ψ 2 (0) = 187.4743 when c = 1.2, α = 1 and the Erlang scale parameter is β = 2. It is easily verified that when n = 2, equations (9) and (8) give We remark that Borovkov & Dickson (2008) obtained a general expression for w (u, t) in terms of the convolutions of the density k, and that in some cases, for example Erlang inter-claim times, an explicit solution for w(u, t) exists. It is not an easy task to find the moments of the time of ruin directly from such solutions when u > 0; this point is amply demonstrated by Drekic & Willmot (2003) who consider the case of the classical risk model.
For a modified SA model, we have w m (0, Setting n = 1 in equation (11) gives Similarly, setting n = 2 in equation (11) gives the final result as We discuss an approach to finding ψ m n (0) in section 7, and in section 8, we show how ψ m n (0) can be obtained for some modified Erlang risk models.

The Classical Risk Model
We now consider the classical risk model, so that {N(t)} t≥0 is a Poisson process with parameter λ. For this model, we can apply formula (8) for individual claim amount distributions for which the factorisation of f introduced by Willmot (2007) holds, namely for functions {η j , τ j } m j=1 . Willmot (2007) shows that if the individual claim amount distribution is an infinite mixture of Erlang distributions with the same scale parameter then this factorisation applies. See also Willmot & Woo (2007) for further applications.

Phase-Type(2) Risk Models
We can follow a process very similar to that in the previous section when the distribution of times between claims is phase-type(2) (also called Coxian (2)). However, we must now impose extra conditions on the individual claim amount distribution beyond the factorisation (13), but these conditions are not particularly restrictive. Following Dickson & Li (2010), we consider the situation when the functions {η j } m j=1 satisfy the factorisation In the previous section, we made use of the unique positive solution ρ of Lundberg's fundamental equation for the classical risk model. We know from Ji & Zhang (2012) that in the phase-type(2) risk model, there are two distinct positive solutions to Lundberg's fundamental equation, and we can use these in a similar way.
To find ψ e n (u), we apply results from Dickson & Li (2012) who consider modified Erlang(n) risk models under which the distribution of the time to the first claim is Erlang(j) for j = 1, 2, 3, . . . , n − 1. In the case of the modified Erlang(2) risk model under which the distribution of the time to the first claim is Erlang(1) with the same scale parameter β, we find from formula (4.2) of Dickson & Li (2012) 1+2r 1 ) (1 + 2r 1 )(1 + 2r 2 ) ⎞ ⎠ .
The arguments that allow us to obtain the functionsη m i (δ) for i = 1, 2 in the above example are given in section 4 of Dickson & Li (2012). Unfortunately, similar arguments do not seem to apply for other phase-type(2) risk models. However, in the next section we show how the negative solutions of Lundberg's fundamental equation can be used to obtain the bivariate Laplace transform of w m (0, y, t), giving us what we need to find ψ m n (u).

Modified SA Models
We now illustrate how we can find the Laplace transform of w m (0, y, t) (leading to the functions ν m n (u)) using the negative solutions of Lundberg's fundamental equation. We can do this by considering the Gerber-Shiu function with initial surplus 0 for a modified SA model of the form is the equivalent Gerber-Shiu function for the ordinary SA model. The objective is to first solve for ϕ(u) and then use this to express ϕ m (0) as a sum of products of Laplace transforms in δ and s. We illustrate the idea in the next example. It seems difficult to give general results as we require an assumption for k 1 in order to solve for ϕ m (0).
Example 7.1. Suppose that in an ordinary SA model, the claim inter-arrival times are distributed as the sum of two independent exponential random variables with parameters λ 1 and λ 2 respectively where λ 1 < λ 2 , and f (x) = 2 i=1 q i α i e −α i x where q 1 + q 2 = 1. Gerber & Shiu (2005) show that for this model where −R i ( = −R i (δ)) for i = 1, 2 are the negative solutions of Lundberg's fundamental equation for i = 1, 2. Equations (18) are exactly the same equations that are solved in the context of the classical risk model by Dickson & Drekic (2006), so we have Suppose that we want to find the moments ψ e n (u). As both k and k e are combinations of exponentials, it is sufficient to consider the case when k 1 (t) = λe −λt for t > 0, as we can then find the required quantities using Theorems 3.2 and 3.3. We get We can evaluateη l (0) and by differentiating Lundberg's fundamental equation (17) we can find the derivatives of R i (δ) when δ = 0, and we can insert these into the derivatives ofη l (δ) to evaluate the quantities H i,m l for l = 1, 2 and i = 1, 2, 3, . . .. All other details required to find ψ n (u) for this example can be found in Gerber & Shiu (2005).
We remark that this approach can also be used to find the bivariate transform of w(0, y, t). For this, we simply require to find ϕ(u). However, as in the above example, the transforms with respect to δ may be complicated expressions in terms of the negative solutions of Lundberg's fundamental equation.
It is also possible to use this approach to find an expression for the Laplace transform with respect to δ of the time of ruin for the modified process, and to then find the moments of the time of ruin for the modified process by differentiating this Laplace transform. However, this can be a very cumbersome process; interested readers can try to reproduce the results for ψ m n (u) in Example 6.1 using this approach.

The Erlang(n) Risk Model
In this section, we consider the Erlang(n) risk model and take a different approach to finding the quantities for this model that are equivalent to H n j in section 5 and H n,m j in section 6. We extend ideas from Dickson & Li (2012), and we start with some definitions from that paper. For m = 0, 1, 2, . . . let where β is the scale parameter of the Erlang(n) distribution, and for j = 0, 1, 2, . . . , n − 1 define where f is still the density function for individual claims and f r * is the r-fold convolution of f with itself. Dickson & Li (2012) show that when the factorisation (13) holds, for a modified Erlang(n) risk model under which the distribution of the time to the first claim is Erlang(j) with scale parameter β, then, with m j denoting this modification, the joint density of the time of ruin and deficit at ruin when the initial surplus is zero is of the form and this holds for j = 1, 2, 3, . . . , n − 1. They give an expression forh j,l (δ) (the final formula in their section 3.1), but this expression does not appear helpful for our purpose. It is convenient to adopt the notation w m n (0, y, t) for the joint density of T 0 and Y 0 for the ordinary Erlang(n) risk model. Using the arguments in Dickson & Li (2012), for this modified Erlang(n) risk model claims occur at occurrences j, n + j, 2n + j, . . . of an underlying Poisson process, and we find that The first two terms in the above consider the situation when a claim occurs in the interval (t, t + dt), while the third term adjusts for upcrossings of the surplus process through level 0 prior to time t. Suppose the last upcrossing through 0 occurs with n − i Poisson events until the next claim, for i = 0, 1, 2, . . . , n − 1. Then, we have a contribution w m n−i (0, y, t − s). This means there have been i Poisson events from the time of the previous claim to time s, so the number of Poisson events to time s is rn + j + i for some r, r = 0, 1, 2, . . .. Using (19) and (13) we get In order to find moments of the time of ruin for the modified processes, we require quantities we denote by H r j,l given by for r = 0, 1, 2, . . .. For brevity, we now write (20) as where a j,l (t) = β p j−1 (t) η l (ct) + β Then from (21), we get Thus, we have a system of equations that we can use to find the quantities H r j,l . We note that for n = 0, 1, 2, . . . Example 8.1. We consider the situation when the claim inter-arrival time distribution is Erlang(3,β), and the individual claim amount distribution is Erlang(2,α), and we find the moments ψ n (u) and ψ e n (u) for n = 1, 2. As where κ i (t) is the Erlang(i,β) density, we can find the moments ψ e n (u) using Theorem 3.3 by finding ψ m j n (u), j = 1, 2, 3. The factorisation (13) gives η 1 (x) = αxe −αx , η 2 (x) = e −αx , τ 1 (y) = αe −αy , τ 2 (y) = α 2 ye −αy , so that for j = 1, 2, 3 w m j (0, y, t) = h j,1 (t) τ 1 (y) + h j,2 (t) τ 2 (y) and for the ordinary Erlang(3) risk model g(0, y) = H 0 3,1 τ 1 (y) + H 0 3,2 τ 2 (y) .
Figures 1 and 2 respectively show the mean and standard deviation of the time of ruin, given that ruin occurs, for both the ordinary and the equilibrium models. It is not surprising that the plots are close in each figure, given that the only difference for these risk processes is the distribution of the time to the first claim.

Concluding Remarks
We have presented general formulae for moments of the time of ruin in ordinary and modified Sparre Andersen risk models, and we have shown how they can be applied to find exact results. Our approach does not offer any advantage over the existing result by Lin & Willmot (2000) for the classical risk model. However, it does provide a fairly simple way of finding solutions for ψ n (u) in phase-type(2) risk models. In the case of Erlang risk models, we are able to find moments in the ordinary, equilibrium, and some modified cases. Application of our approach in sections 5 and 6 hinges on some knowledge of the functions w(0, y, t) and w m (0, y, t), as well as the factorisations we have imposed. In practice, these factorisations are not serious impositions as many individual claim amount distributions satisfy them. The approach in section 7 is more generally applicable, but is likely to require more tedious differentiation than in sections 5 and 6. To apply the approach in section 8, we must be able to compute convolutions of the individual claim amount distribution. This is not a great restriction as we can do this when the individual claim amount distribution is an infinite mixture of Erlang distributions with the same scale parameter, and Willmot & Woo (2007) show that a wide variety of distributions are of this type.
An open question is how to find the quantity ν n (y) given by (5), either analytically or numerically (which is perhaps more likely), for forms of f for which the factorisation (13) does not apply. If we can do this, then it should be possible to use numerical approaches to find ψ n (u) for u > 0. An approach for the case u = 0 is presented in Dickson & Hipp (2001) where the inter-claim times are Erlang(2) distributed and the claim amount distribution is Pareto.