Professional Documents
Culture Documents
1 Simple examples
We start by benefits.
∗ Departmentof Theoretical Statistics. Department of Mathematical Sciences,
Aarhus University, Ny Munkegade, 8000 Aarhus C, Denmark; asmus@imf.au.dk;
http://home.imf.au.dk/asmus
1
Life insurance This pays b at the time of death so the present value is be−rTx .
The payment may only take place in a certain age span (x+m, x+n] where
0 ≤ m < n ≤ ∞, in which case the present value is be−rTx I(m < Tx ≤ n).
Life annuities This pays b at a continuous rate until Tx so that the present
value is Z Tx
b
e−rt b dt = 1 − e−rTx .
0 r
The payments may be delayed until time m > 0, as would typically be the
case in pension arrangements, giving a present value of
Z Tx
b −rm
I(Tx > m) e−rt b dt = e − e−rTx I(Tx > m) .
m r
The simplest forms of contributions are (a) a lump sum c paid at t = 0 (the
present value is just c), (b) a continuos payment at rate c up to n contingent
upon survival. This has the same form as a life annuity with m = 0 so by similar
reasoning the present value is
Z Tx ∧n
c
e−rt c dt = 1 − e−r(Tx ∧n) .
0 r
2
• The future interest rate and mortality rates are not completely known
• Even if the previous points are dealt with, one is running a business where
expenses and incomes just balance. Thus, there is no profit which in turn
(in the presence of randomness) can be proved to imply that the company
will be eventually ruined.
In practice, the equivalence principle is nevertheless the one used for premium
calculation, but with a twist in the implementation, to use an interest rate and
a set of mortality rates which are not the ones one expects to be in force but
biased in favour of the company. Such a set of rates in referred to as the first
order basis.
In which direction the rates of interest and mortality should be biased in
order to be in favour of the company depends crucially on the type of contract.
RT
Consider for example a life annuity with present value 0 x e−rt b dt. This is a
decreasing function of r so it is clear that the r of the first order basis should be
smaller than the interest rate one expects to be in force. Similarly, the present
value is increasing in Tx so that the first order basis would have a distribution
of Tx which is in some sense larger than the true one. This is certainly the
case if one takes the λ(x) to be smaller than the true ones for all x. For a life
insurance, there is also an easy answer, see the Exercises. However, for many
types of contracts the situation is less clear cut. For example, it is not clear
whether (1) is increasing or decreasing in r.
3 Regulation
In vitually any country, the life insurance business is heavily regulated by the
government in order to protect the insurance buyers. There are (at least) two
basic concerns:
• To ensure that the company will be able to meet its obligations. This is
done by demanding that the reserve (the bank account, investments etc.)
is at least the expected future payments to the policy holders, and by re-
quiring the premiums to be sufficiently large to ensure that the probability
of the company making a loss is very small.
• To put limits on the companys profit. This may appear to be in conflict
with the requirements on not too small premiums. The most common
way to overcome this is via bonus, a payment going back to the insured
which is calculated using the second order basis, that is, the interest and
mortality rates which were actually in force during the running time of the
contract. For a simple example of a possible implementation, consider a
life insurance with a lump premium of c = bEe−rTx calculated using a first
order basis which at time Tx turned out to have been correct for the λ(y)
but not for the r used, say the value at time t ≤ Tx was r(t). Assuming
the company is allowed to make 15% profit, the premium should then have
RT
been c∗ = 1.15b exp − 0 x r(t) dt rather than c and the payment is then
raised from b to bc/c∗ .
3
4 Mortality modeling
Recall that x is the age of the insured and Tx ≥ x his/her remaining life time.
If T0 is the life length of a newborn baby and F0 its distribution, then
F0 (x + y)
Fx (y) = P(Tx ≤ y) = P T0 ≤ x + y T0 > x =
1 − F0 (x)
The hazard rate and the density are given in the following figures.
The shape of the density is obviously close to what one expects for human
mortality. Whether the G82 parameters exactly match the mortality in the
population is not that essential because of later regulation by bonus.
For simulation, the easiest way to generate a GM r.v. is acceptance rejection.
It is seen from the plot of the density that a majorizing density is the uniform
density u(x) ≡ u = 0.0085 on [0, 120] for which f (x) ≤ 4u(x), x ≤ 120. Thus,
the algorithm for generating T0 is
1. Generate two independent uniforms U, V on (0, 1).
2. Let Y = 120U (trial for T0 )
3. if V ≤ f (Y )/4u, output T0 = Y ; otherwise return to 1.
If instead one wants to generate Tx , say for x = 55, the modification is
4
1.4
1.2
0.8
0.6
0.4
0.2
0
0 20 40 60 80 100 120
0.035
0.03
0.025
0.02
0.015
0.005
0
0 20 40 60 80 100 120
5
3. If Y < 55, return to 1
4. if V ≤ f (Y )/4u, output T0 = Y − 55; otherwise return to 1.
The algorithms are easily programmed, but not very effficient; for example, the
acceptance probability for T0 is just 1/4 and for T55 even lower. However, for
the average user this is of little concern compared to the effort in speeding up
the algorithms by factors of 2 or 3.
5 Exercises
1. If T0 has cdf F (x), density f (x) and hazard rate λ(x) = f (x)/F (x) where
F (x) = 1 − F (x), what is then the hazard rate λ(x) of Tx ?
2. Assume that the time unit is years. The yearly interest rate i is then defined by
one monetary unit increasing to 1 + i in a year. Find the formula connecting i
and the continuous interest rate r.
3. Assume insted that interest is added every half year, increasing the account by
a factor of 1 + i2 /2. If i2 = 5%, what is then i?
4. How should the parameters of the first order basis compare with the actual ones
for a life insurance with n = ∞?
5. Explain what is the relevance for the Assignment of the following material on
the delta method.
R R
1 X 1 X
X = Xr , Y = Yr ,
R r=1 R r=1
but how should one estimate the error and thereby give a confidence interval?
This is what the delta method is about.
The idea is just to use a Taylor expansion around (x, y) which with fx =
∂f /∂x, fy = ∂f /∂y the partial derivatives means
fb − f (x, y) = f X, Y − f (x, y)
R
1 X
≈ fx (x, y) X − x + fy (x, y) Y − y = (Zr − z)
R r=1
where Zr = fx (x, y)Xr + fy (x, y)Yr , z = EZ = fx (x, y)x + fy (x, y)y. Since
Z1 , . . . , ZR are i.i.d., we have
R
1 X
√ (Zr − z) → N (0, ω 2 )
R r=1
6
√
where ω 2 = V ar(Z). This suggests R(fb −√ f (x, y)) → N (0, ω 2 ) so that the
b
95% confidence interval should be f ± 1.96ω/ R.
In practice, we have the complication that not only ω 2 but also fx (x, y), fy (x, y)
are not known so that we cannot observe the Zr . The obvious procedure is to
approximate Zr by
let
R
1 X ∗ ∗ 2
b2 =
ω Zr − Z
R − 1 r=1
√
and use the confidence interval fb ± 1.96b
ω/ R.
NB The above introduction to the delta method is written with undergradu-
ates in mind. After later courses, you will know the multivariate central limit
theorem which enables a somewhat more streamlined presentation.
7 Assignment
Consider a contract with benefits Y and premium payable as a lump sum c
at t = 0. When calculating c via the equivalence principle, the size of safety
margin (loading) will usually depend on how risky the contract is in the sense
of variability of Y around its expected value. Any measure of variability would
involve the distribution of Y which is in general difficult to calculation so we
will use simulation to compare the benefits
b1
Y1 = 1 − Ee−rTx , Y2 = b2 Ee−rTx
r
of two types of contracts, a life annuity and a life insurance. Take x = 55 and
use r = 5% and the G82 Gompertz-Makeham parameters; as help, you will be
given a S function generating Tx with this distribution.
As first measure of variability, we take the variances EY12 /EY1 −1, EY22 /EY2 −
1. Give simulation estimates for both r.v.’s and, if time permits, a confidence
interval using the delta method. You will be helped by be given a S function in
the setting of the delta method with f (x, y) = x/y
As second measure of variability, take if time permits the VaR (value at
risk) y0.99 defined as the 99% quantile in the distribution of a r.v. Y , that is,
the number satisfying P(Y ≥ y0.99 ) = 0.01. The use of the VaR comes from
mathematical finance where one thinks of Y as a potential loss. A simulation
estimate is obtained by simulating (say) R = 105 replications Y1 , . . . , YR and
the empirical 99% quantile (details later).
References
[1] H.U. Gerber (199x) Life Insurance. Springer.
[2] R. Norberg (2002) Basic Life Insurance Mathematics. Lecture Notes used by the
actuarial studies at the Univ. Copenhagen.