Abstract
A multi-state life insurance model is described naturally in terms of the intensity matrix of an underlying (time-inhomogeneous) Markov process which specifies the dynamics for the states of an insured person. Between and at transitions, benefits and premiums are paid, defining a payment process, and the technical reserve is defined as the present value of all future payments of the contract. Classical methods for finding the reserve and higher order moments involve the solution of certain differential equations (Thiele and Hattendorff, respectively). In this paper we present an alternative matrix-oriented approach based on general reward considerations for Markov jump processes. The matrix approach provides a general framework for effortlessly setting up general and even complex multi-state models, where moments of all orders are then expressed explicitly in terms of so-called product integrals of certain matrices. Thiele and Hattendorff type of theorems may be retrieved immediately from the matrix formulae. As a main application, methods for obtaining distributions and related properties of interest (e.g. quantiles or survival functions) of the future payments are presented from both a theoretical and practical point of view, employing Laplace transforms and methods involving orthogonal polynomials.
Similar content being viewed by others
Notes
Here and in the following, a − subscript mimics ‘before S’ and a + ‘after S’.
References
Abate J, Whitt W (1995) Numerical inversion of Laplace transforms of probability distributions. INFORMS J Comput 7:36–43
Albrecher H, Teugels JL, Tichy RF (2001) On a gamma series expansion for the time-dependent probability of collective ruin. Insur Math Econ 29(3):345–355
Asmussen S (2003) Applied Probability and Queues, 2nd edn. Springer, New York
Asmussen S, Goffard PO, Laub PJ (2019) Orthonormal polynomial expansions and lognormal sum densities. In: Barrieu P (ed.) Ragnar Norberg. Risk and Stochastics, chap. 6, pp. 127–150
Asmussen S, Koole G (1993) Marked point processes as limits of Markovian arrival streams. J Appl Probab 30(2):365–372
Asmussen S, Steffensen M (2019) Risk and insurance. A graduate text. Probability theory and stochastic modelling. Springer, New York (to appear)
Barndorff-Nielsen O, Cox D (2013) Asymptotic techniques for use in statistics. Monographs on statistics and applied probability. Monographs on statistics and applied probability. Springer, New York
Bladt M, Meini B, Neuts M, Sericola B (2002) Distributions of reward functions on continuous-time Markov chains. In: 4th international conference on matrix analytic methods in stochastic models (MAM4). Adelaide, Australia
Bladt M, Nielsen BF (2017) Matrix-exponential distributions in applied probability. Springer, New York
Bowers N (1966) Expansion of probability density functions as a sum of gamma densities with applications in risk theory. Trans Soc Actuar 18(52):125–137
Buchardt K, Møller T (2015) Life insurance cash flows with policyholder behavior. Risks 3(3):290–317
Cramér H (1946) Mathematical methods of statistics (Almqvist & Wiksells Akademiska Handböcker). Princeton University Press, Princeton
Dobrushin RL (1953) Generalization of Kolmogorov’s equations for Markov processes with a finite number of possible states. Math Sb. (N.S.) 33(75):567–596
Gill RD, Johansen S (1990) A survey of product-integration with a view toward application in survival analysis. Ann Stat 18(4):1501–1555
Goffard PO, Laub PJ (2019) Two numerical methods to evaluate stop-loss premiums. J Comput Appl Math (to appear)
Goffard PO, Loisel S, Pommeret D (2016) A polynomial expansion to approximate the ultimate ruin probability in the compound poisson ruin model. J Comput Appl Math 296:499–511
Goffard PO, Loisel S, Pommeret D (2017) Polynomial approximations for bivariate aggregate claims amount probability distributions. Methodol Comput Appl Probab 19(1):151–174
Hesselager O, Norberg R (1996) On probability distributions of present values in life insurance. Insur Math Econ 18(1):35–42
Hoem JM (1969) Markov chain models in life insurance. Blätter der DGVFM 9(2):91–107
Johansen S (1986) Product integrals and Markov processes. CWI Newsl 12:3–13
Loan Cv, Moler C (2003) Nineteen dubious ways to compute the exponential of a matrix, twenty-five years later. SIAM Rev 45:1–46
Moler C, Loan Cv (1978) Nineteen dubious ways to compute the exponential of a matrix. SIAM Rev 20(4):801–836
Norberg R (1991) Reserves in life and pension insurance. Scand Actuar J 1991(1):3–24
Norberg R (1995) Differential equations for moments of present values in life insurance. Insur Math Econ 17(2):171–180
Norberg R (2001) On bonus and bonus prognoses in life insurance. Scand Actuar J 2001(2):126–147
Protter M, Morrey C (1996) Intermediate calculus. Undergraduate texts in mathematics. Springer, New York
Ramlau-Hansen H (1988) Hattendorff’s theorem: a Markov chain and counting process approach. Scand Actuar J 1988(1–3):143–156
Szegö G (1939) Orthogonal Polynomials. American Mathematical Society Colloquium Publications, 23. American Mathematical Society, Providence
Van Loan C (1978) Computing integrals involving the matrix exponential. IEEE Trans Autom Control 23(3):395–404
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix: On the \(L_2\) condition without piecewise constancy
Appendix: On the \(L_2\) condition without piecewise constancy
Remark 7
Condition (iii) in Theorem 7 is not far from being necessary. Consider as an example the disability model in the time interval [0, 1] with states 0: active, 1: disabled, 2: dead and allowing recovery, with the same intensity \(\lambda >0\) for transitions from 0 to 1 as from 1 to 0 and (for simplicity) mortality rate 0 and discounting rate \(r=0\). The benefits are a lump sum \(b^{01}=1\) upon transition from 0 to 1 and the contributions a constant payment rate \(b^0<0\) when active.
When \(Z(0)=0\), the total number N of transitions in [0, 1] is Poisson\((\lambda )\), the total benefits \(U_1(0,1)\) are \(M=\lceil N/2\rceil\), and the total contributions \(U_0(0,1)\) equal \(|b^1|\) times the total time \(T_0\) spent in state 0. Thus \(U(0,1)=\)\(U_1(0,1)-U_0(0,1)\)\(=M-U_0(0,1)\) Obviously, \(U_0(0,1)\) is concentrated on the interval \([0,|b^1|]\) with a density g(x) which is bounded away from 0, say \(h(x)>c_1>0\).
Assuming, again for simplicity, that \(b^1>1\). Then the intervals \([m- b^1,m]\) overlap and so for a given x, at least one of them contribute to f(x). One candidate is the one with \(m=\lceil x\rceil\). This gives
But using Stirling’s approximation to estimate the Poisson probability, it follows after a little calculus that \(\mathbb {P}(N=2\lceil x\rceil )\ge\)\(c\mathrm {e}^{-4x\log x}\) for all large x, say \(x\ge x_0\), and some \(c>0\) (in fact the 4 can be replaced by any \(c_3>2\)). This gives for a normal\((\mu ,\sigma ^2)\)\(f_0\) that
That is, (8.1) fails and a similar calculation gives that (8.5) does so too.
The obvious way out is of course to take \(f_0\) with a heavier tail than the normal, say doubly exponential (Laplacian) with density \(\mathrm {e}^{-|x|}\) for \(-\infty<x<\infty\). Given that this example and other cases where condition (iii) is violated do not seem very realistic, we have not exploited this further.
Example 4
Consider again the disability model with states 0: active, 1: disabled, 2: dead and no recovery. As in [11], we assume that the payment stream has the form \(b^0(t)=-b_-^0\) for \(t\le S\) and \(b^0(t)=b_+^0\) for \(t>S\),Footnote 1\(b^1(t)\equiv b^1\) for all t. For example S could be the retirement age, say 65. Without recovery, the only non-zero transition rates are the \(\mu _{01}(t)\), \(\mu _{02}(t)\), \(\mu _{12}(t)\). With the values used in [11], these are bounded away from 0 and \(\infty\) on [0, T] for any T (say 75 or 80), and this innocent assumption is all that matters for the following.
Define the stopping times
with the usual convention that the stopping time is \(\infty\) if there is no t meeting the requirement in the definition. One then easily checks that the sets \(F_0,\ldots ,F_{1_+2_+}\) defined in Fig. 5 defines a partition of the sample space. Here \(F_0\) (corresponding to zero transitions in [S, T]) contributes with an atom at
and the remaining 7 events with absolutely continuous parts, say with (defective) densities \(g_{2_-},\ldots ,g_{1_+2_+}\). It is therefore sufficient to show that each of these is bounded. In obvious notation, the contribution to U(S, T) of the first 4 of these 7 events (corresponding to precisely one transition in [S, T]) are
The desired boundedness of \(g_{2-},g_{2+},g_{1-},g_{1_+}\) therefore follows from the following lemma, where \(\tau\) may be improper so that \(\int h<1\):
Lemma 3
If a r.v. \(\tau\) has a bounded density h and the function \(\varphi\) is monotone and differentiable with \(\varphi '\) bounded away from 0, then the density of \(\varphi (\tau )\) is bounded as well.
Proof
The assumptions imply the existence of \(\psi =\varphi ^{-1}\). Now just note that the density of \(\varphi (\tau )\) is \(h(\psi (x))/\varphi '(\psi (x))\) if \(\varphi\) is increasing and \(h(\psi (x))/\left| \varphi '(\psi (x))\right|\) if it is decreasing.\(\square\)
The cases of \(g_{1_-2_-},g_{1_-2_+},g_{1_+2_+}\) (corresponding to precisely two transitions in (s, T]) is slightly more intricate. Consider first \(g_{1_-2_-}\). The contribution to U(0, T) is here
Now the joint density \(h(t_1,t_2)\) of \((\tau _0^1,\tau _0^2)\) at a point \((t_1,t_2)\) with \(0<t_1<t_2\le R\) is
so that \(h(t_1,t_2)\) is bounded. Consider now the transformation taking \((\tau _0^1,\tau _0^2)\) into \((\tau _0^1,A_{12_-})\). The inverse of the Jacobian J is
which is uniformly bounded away from 0 when \(\tau _1^2\le S\). Therefore also the joint density \(k(z_1,z_2)\) of \((\tau _0^1,A_{12-})\) is bounded. Integrating \(k(z_1,z_2)\)w.r.t. \(z_1\) over the finite region \(0\le z_1\le S\) finally gives that \(g_{1_-2_-}\) is bounded.
The argument did not use that \(\tau _1^2\le S\) and hence also applies to \(g_{1_-2_+}\). Finally note that the contribution to \(g_{1_+2_+}\) from [0, S] is just a constant, whereas the one from (S, T] has the same structure as used for \(g_{1_-2_-}\). Hence also \(g_{1_+2_+}\) is bounded.
Remark 8
The calculations show that \(F_{1+}\) also is an atom if \(b_+^0=b^1\). However, since the contribution rate \(b_-^0\) is calculated via the equivalence principle given the benefits \(b^0_+,b^1\) and the transition intensities, this would be a very special case. The somewhat less special situation \(b^0_+=b^1\) (same annuity to an disabled as to someone retired as active) also gives an atom, now at \(F_{1+}\). In fact, \(b_-^0=b^1\) is assumed in [11].
Remark 9
For a counterexample to (8.1), consider again the disability model with the only benefit being a lump sum of size \(b^{01}(t)=\)\(\mathrm {e}^{rt}\varphi (t)\) being paid out at \(\tau _0^1\) where \(\varphi (t)=\)\((t-a)^21_{t\le a}+b\), cf. Fig. 6.
Then \(U(s,T]=\varphi (\tau _0^1)\) has an atom at b (corresponding to \(\tau _0^1>a\)) and an absolutely continuous part on \((b,b+a^2]\). Letting \(\psi :\,[b,b+a^2]\mapsto [0,a]\) be the inverse of \(\varphi\), \(t=\psi (y)\) satisfies
where the minus sign after the first \(\Rightarrow\) follows since \(\psi\) is decreasing. With h the density of \(\tau _0^1\), we get as in Lemma 3 that the conditional density f of the absolutely continuous part is given by
But there are \(c_1,c_2>0\) such that \(f_0(y)\le c_1\) on \([b,b+a^2]\) and \(h(t)\ge c_2\) on \([b,b+a^2]\), and so we get
meaning that (8.1) does not hold.
Rights and permissions
About this article
Cite this article
Bladt, M., Asmussen, S. & Steffensen, M. Matrix representations of life insurance payments. Eur. Actuar. J. 10, 29–67 (2020). https://doi.org/10.1007/s13385-019-00222-0
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s13385-019-00222-0