T H E CLAIMS R E S E R V I N G P R O B L E M IN N O N - L I F E I N S U R A N C E :

SOME S T R U C T U R A L IDEAS

BY ELJA ARJAS

University of Oulu, Finland

ABSTRACT

We present some relatively simple structural ideas about how probabilistic

modeling, and in particular, the modern theory of point processes and
martingales, can be used in the estimation of claims reserves.

1. I N T R O D U C T I O N

The claims reserving problem, or the run off problem, has been studied rather
extensively. The monograph by TAYLOR (1986) covers most of the develop-
ments so far, and, interestingly enough, creates a taxonomy to the models
introduced. The booklet of VAN EEGHEN (1981) has a somewhat similar aim.
Because of these recent surveys we do not intend to describe " t h e state of the
a r t " in this area but confine ourselves to a few remarks.
There has been a clear tendency away from deterministic "accounting
m e t h o d s " into more descriptive probabilistic models. Early works in this
direction were BOHLMANN et al. (1980), HACHEMEISTER (1980), LINNEMANN
(1980) and REID (1981). Of more recent contributions we would like to
mention particularly PENTIKAINEN and RANTALA (1986), and three papers
dealing with unreported (IBNR) claims: NORBERG (1986), ROBBIN (1986) and
J E W E L L (1987).
Most authors today tend to agree that there are important benefits from
using structurally descriptive probabilistic models in insurance. However, there
appears to be a new problem : With the increased realism of such models, many
papers introduce, very early on, a long list of special assumptions and a
correspondingly complicated notation. A reader may then not be able to see
what ideas are really important and characteristic to the entire claims reserving
problem, and what are less so, only serving to make the calculations more
explicit. It would be more pleasant if the modeling could be started virtually
without any assumptions, and then only adding assumptions as it becomes
clear that advancing otherwise is difficult. We think that the modern theory of
stochastic processes comes here to aid, and try to illustrate this in the
following. We are mainly using " t h e martingale approach to point processes",
as discussed e.g. in BRfZMAUD (1981) and KARR (1986). However, apart from
some calculations towards the end, no previous knowledge of this theory is
really needed to understand the paper.
ASTIN BULLETIN, Vol. 19, No. 2
140 ELJA ARJAS

The emphasis of this paper is in the conceptual analysis of Section 2 and the
structural results of Section 3. Section 4 provides an illustration of how the
actual stochastic calculus, in a simple form, can be applied to obtain more
explicit results.
We want to stress that this paper contains very little that could be called
" n e w results": it is more important to us here how we arrive at them.

2. CLAIMS, INFORMATION AND SETTLEMENT

AS MARKED POINT PROCESSES
Considering a fixed accident year, say the unit interval (0, l], let the exact
occurrence times of the accidents be Ti* _< T~ _< ... An accident which occurs
at time T/* is reported to the company after a random delay D i so that its
reporting time is R~' = T i * + D i . We denote the ordered reporting times (order
statistics) by TI < T2 < ..., assuming for simplicity that they are all different.
We follow the convention that the accidents are indexed according to the
order in which they are reported to the company, i.e., the accident reported at
Ti is called " t h e i th accident". Because of the random delay in reporting this
indexing is often different from the one that refers to the occurrence times.
In practice the number of accidents in a given year of occurrence is of course
finite. We denote this (random) number by N. As a convention, we let the
sequence (T~) be infinite but define T,v+l = TN+2 = ... = oo.
Let us then assume that every time a new accident is reported to the
company, this will be followed by a sequence of " h a n d l i n g times". These
handlings could be times at which claim payments are paid, but also times at
which the file concerning the accident is updated because of some arriving new
information. Supposing that the i a' accident has altogether Ni handling times
following its reporting, we denote them by
(2.1) Ti = Tio < Til < Ti2 < ... < Ti, N,.

Again, we let Ti, N,+ t = Ti, u,+2 = ... = oo.

Next we need to specify the event that takes place at To.. If a payment is
made then, we denote the amount paid by X 0. If nothing is paid at T 0 we
simply let X 0 = 0. Similarly, it is convenient to have a notation for the
information which is used for updating the accident file. Let I~0 be the
information which becomes available when the accident is reported, and let I o
be the new information which arrives at handling time T 0. If there is no such
information, we set I o = 0, signalling " n o new information". In particular we
set X 0 = 0 and I 0.= 0 whenever T O.= oo.
Our analysis will not depend on what explicit form the variables I 0 are
thought to have. They could well be strings of letters and numbers, reflecting,
for example, how the accident is classified by the company at time T O. Iio will
often determine what was the delay in reporting the i rk accident. If further
payments are made after the case was thought in the c o m p a n y to be closed, it is
probably convenient to consider the arrival of the first such claim as a new
reporting time, also initiating a new sequence of handlings.
 THE CLAIMS RESERVING PROBLEM IN NON-LIFE INSURANCE 141

The above definitions give rise to a number of stochastic processes which are
of interest in the claims reserving problem. The first definitions will be accident
specific, after which we obtain the corresponding collective processes by simple
summation.
We start by assigning the payments X~j to the handling times T U. In this way
we arrive, for each i, at a sequence (To., Xo)j~o, where T~ = Ti0 < Ta < ...
(with strict inequalities if the variables are finite) and X o > O. Thus
(To., Xu)j~ o can be viewed as a marked point process (MPP) on the real line,
with non-negative real " m a r k s " X O. We call it the payment process. Equiva-
lently, of course, we can consider the cumulative payment process ( X i ( t ) )
defined by

(2.2) Xi (t) = 2 XU"

{j:Tq~ t}

Clearly, X i ( t ) represents the total amount of payments (arising from accident

i) made before time t. The function t~-~Xj(t) is an increasing step function,
with Xi(t ) = 0 for t < T i ( = reporting time) and X i ( t ) approaching, as
t ~ oo, the limit

(2.3) X~(oo) = ~ Xu,

jmo

which is the total compensation paid for the i u' accident. Similarly,

u~(t) = x , ( o o ) - x~(t)

(2.4) = 2 X,j
{2: To > t}

represents the total liability at t coming from future payments, with

U i ( t ) = Xi(oo) for t < Ti and U i ( t ) decreasing stepwise to 0 as t--, oo.
We remark here that, in order to keep this simple structure, we do not
consider explicitly the effects of interest rate or inflation. This means, among
other things, that the future claims must be expressed in standardized (deflated)
currency.
Second, we can consider the sequence (Tu, Io.)2~ o and call it the information
process for the i u' accident. This, too, is an MPP, with mark IL/taking values in
some conveniently defined set. As mentioned earlier the form of the marks is
not restricted in any real way" It will suffice, for example, that there is a
countable number of possible marks.
Our third MPP is obtained by combining the marks of the other two, into
pairs (Xij,lij). We call (To,(Xij, lu))j> o the settlement process of the i th
accident.
142 ELJA ARJAS

Considering finally all accidents collectively, we obtain the corresponding

collective payment process, information process and settlement process by a
simple summation (superposition) over the index i. However, we do not need a
separate notation for these MPP's and will therefore confine ourselves to the
cumulative payment process

(2.5) X . ( t ) = ~ X,(t)
i
and the liability process

(2.6) U. (t) 2 U, (t).

i

Observe that it is not necessary to restrict the summation to indices i satisfying

i < N because, unless this is satisfied, X , ( t ) = Ui(t) -- 0 for all t.

3. C L A I M S RESERVES AS A P R E D I C T | O N PROBLEM

The estimation of the claims reserves can now be viewed as a prediction

problem where, at a given time t representing " t h e present", an assessment of
the future payments is made on the basis of the available information. Most of
our mathematical considerations do not depend on whether the assessment
concerns the payments from an individual i th accident, or all accidents during
the considered year of occurrence. Because of this we will often simply drop the
subscript ( " i " or " - " ) from the notation. Thus, for example, U ( t ) can be
taken to be either the accident specific liability Ui(t) or their sum U. (t).
The role of the information process above is to provide a formal basis for the
assessments made. This is done most conveniently in terms of histories, i.e.,
families of a-fields in the considered probability space, which correspond to the
knowledge of the values of the random variables generating them. In particu-
lar, we let the a-field

(3.1) "~-t N = tY{(Tij,lo,Xo)i>_ ,,j~0 : T~/< t}

respresent the information carried by the pre-t settlement process arising from
all claims. (For background, see e.g. KARR (1986), Section 2.1). For complete-
ness, we also allow for the possibility of having information which is exogenous
to the settlements. Writing G for such pre-t information, we shall base the
estimation of the future payments on the history ( 3 , ) , with
(3.2) ~r = .:ft N V ~'t'

In an obvious sense, the most complete assessment at time t concerning X(ov),

the total of paid claims, is provided by the conditional distribution

= I .7,).
 THE CLAIMS RESERVING PROBLEM IN NON-LIFE INSURANCE 143

When t varies, these conditional distributions form a so called prediction

process (~i) (see e.g. NORROS (1985)). Here, however, we restrict our attention
to the first two moments of u~. Assuming square integrability throughout this
paper, we write

(3.3) M, = E",(X(oo)) (= f xlt,(dx) )

and

(3.4) V, = Var :'(X(oo))

We now derive some fundamental properties of (Mi) and (VI). From now on
we also write Xt and Ut instead of X(t) and U(t).
Having introduced the idea that ,7, represents "information which the
company has at time t " , it is of course the case that the payments already
made are, at least in principle, included in such knowledge. Formally this
corresponds to the decomposition of X~o into X, and Ut (see (2.4)), i.e.,
(3.5) Xo~ = X,+ U,,
where X, is determined from ...~ (i.e., ,,a-measurable). Therefore, the (.F~)-
based prediction of Xo~ is equivalent to predicting Ut.

CONDITIONAL EXPECTATIONS. Let us first consider the expected values M t. As

a stochastic process, (M, is easily seen to have the martingale-property: For
any t < u,
(3.6) E:'(Mu) = M,.
Thus, since M, is an estimate of X~o at time t and M u is a corresponding
updated estimate at a later time u, (3.6) expresses the simple consistency
principle :
(PI) " C u r r e n t estimate of a later estimate, which is based on more
information, is the same as the current estimate".
Another way to express the martingale property is to say that the estimates
(M,) have no trend with respect to t.
Since Xt is determined from : , we clearly have

Mt = Xt+E:'(Ut) = Xt+ml.
def

Here, the estimated liability at t,

(3.7) m, = E:'(U,),
144 ELJA ARJAS

is a supermartingale, with the "decreasing trend p r o p e r t y "

EJ'(mu) < m, for t < u.

This follows readily from the fact that the true liability Ut is decreasing in time,
as more and more of the claims are paid. Unfortunately such a monotonicity
property is of little direct practical use because the process (U~) is unobserva-
ble : Only the differences U , - Ut = X , - X,, can be observed, but not the actual
values of U, or U,.
The trend properties of (Mr) and (mr) lead to a crude idea about how the
reserve estimates should behave as functions of time. Considering them as a
time series may therefore be useful. On the other hand, one has to remember
that the (super)martingale property is quite weak and only concerns the
(.Tt)-conditional expected values. Thus an apparently downward trend in an
observed time series could be balanced by a rare but big jump upwards.
For a more refined analysis, it would be interesting to study (Mt) in terms of
its martingale integral representation (see e.g. BREMAUD (1981)). The key
ingredient in that representation is the innovation gains process which deter-
mines how (M~) is updated in time when ( J , ) is observed. This theory is well
understood. Unfortunately, however, actuaries seem to have very little idea
about what properties the updating mechanism should realistically possess, and
presently there is no detailed enough data to study the question statistically.
Therefore, a more systematic research effort must wait.
It is instructive to still consider the differences

(3.8) M(t,u) = M,,-M,, t<u.

By the martingale property (3.6) we clearly have EJ'(M(t, u)) = 0. Now, using
the analogous notation X(t, u ) = X , , - X t for the cumulative payments we
easily find that

M(t, u) = [X(t, u) - E ; , ( X ( i , u))] + [EJo(U,,) - E ; ' ( U , ) ] .

The first term on the right is the error in the estimate concerning payments in
the time interval (t, u]. The second term, then, is the updating correction which
is made to the estimated liability when the time of estimation changes from t to u.
Both terms have .7,-conditional expected value 0. This suggests that it might
be beneficial in practice to split the estmate into two parts: one that covers the
time interval to the next update (typically a year) and another for times
thereafter.

CONDITIONAL VARIANCES. The variances Vt give rise to somewhat similar

considerations. First observe that, since X, is determined by ._.~'~,the variance V,
defined in (3.4) satisfies

(3.9) Vr = Var~'-~(U,) = Var"-'(M(t, ~ ) ) .

 THE CLAIMS RESERVING PROBLEM IN N O N - L I F E INSURANCE 145

Thus, if the used estimation method produces also estimates of Vt, the
observed oscillations in (U~) can be compared with the square root of V,.
(Warning: Do not expect normality in short time series!) Second, it is
interesting to note that (Vt) is a supermartingale as well, i.e.,
(3.10) E'"-'(V~,) < Vt for t < u.
This expresses the following intuitively plausible principle:
(P2) "Measured by the conditional variance, the estimates M t tend to
become more accurate as time increases and more information becomes
available".
To show that (3.10) holds, we first find that
E"-'(M(t, u) X(u, oo)) = E"-'(M(t, u) E'Z°M(u, oo)) = 0
so that M(t, u) and M(u, oo) are uncorrelated. This implies the well known
additivity property (" Hattendorf's formula", e.g. GERBER (1979))
(3.11) Var'T'(M(t, or)) = VarJ'(M(t, u)) +Var';'(M(u, co)).
On the other hand,
(3.12) VarJ'(M(u, co)) = EJ'(Var~o(X~)) = EJ,(V,),
so that (3.10) follows by combining (3.9), (3.11) and (3.12).

REMARK. Recall the following well-known result which complements this

picture: with respect to a quadratic loss function, the conditional expectation
M t is the optimal estimate of X(ov). More precisely, for any estimate ,Q, of Xoo
which can be determined from .Y-t(i.e., IQt is .95-measurable), the following
inequality is satisfied:
(3.13) E.7,((X _ ~ , ) 2 ) > E z,((X _ M , ) 2 ) (= V,).

KNOWN AND UNKNOWN ACCIDENTS. Finally in this section we divide the

collective estimate m . t - - E~'(U t) into two parts depending on whether the
considered accidents are at time t known (= reported, IBNER) or unknown
(= not reported, IBNR).
Let the number of known (= reported) accidents at time t be

(3.14) N~ = Z llr,~l"
i

The corresponding liability from future payments is then 2 Uit. Since the
i < Nr
events {T,.< t} are determined by .~, the corresponding .7,-conditional
estimate is simply given by
146 ELJA ARJAS

This formula expresses the intuitively obvious fact that the reserves correspon-
ding to reported accidents could, at least in principle, be assessed individually.
If we are willing to make the assumption, which may not be completely
realistic, that the liabilities Ui, are uncorrelated across accidents given .y,, we
also have a corresponding equality for variances:

(3.16) VarY' ( 2 i<_N, Uu) =2Var'/U' it=Zi<

Nt i X N, Vit"

Note that although the processes (m~,) and (V,) were above found to be
supermartingales, the processes defined by (3.15) and (3.16) do not have this
property. This is because Ni is increasing.
Considering then the unknown (IBNR) accidents, it is obvious that also their
number N - N , is unknown (i.e., not determined by .~) and therefore the

liability estimate E'~' ( ~i>u, U ~ , ) c a n n o t be determined "termwise" a s w a s

done in (3.15). Therefore the estimate needs to be determined collectively

for all IBNR-accidents, a task which we consider in the next section. The only

qualitative pr°perty which we n°te here is that the pr°cess ( E " ( Z,>N, Uil))

is again a supermartingale. This is an easy consequence of the supermartingale

property of (mil), which was established above, and the fact that N, is
increasing.

4. AN I L L U S T R A T I O N : THE ESTIMATION OF IBNR CLAIMS RESERVES

We now illustrate, considering the IBNR claims reserves, how the mathemati-
cal apparatus of the stochastic calculus can be used to derive explicit estimates.
But we are also forced to introduce some more assumptions in order to reach
this goal.
For known accidents, the delays in the reporting times T~ are only important
in so far as they are thought to influence the distribution of the corresponding
payment process. For unknown accidents the situation is completely different:
For unknown accidents the only thing which is known is that if an i '/' accident
occurred during the considered year and it is still unknown at time t, its
reporting time T, exceeds t. (Recall the convention that T,. = oo for i > N).
Therefore, it is impossible to estimate the IBNR reserves individually. A
natural idea in this situation is to use the information which has been collected
about other (i.e., known) accidents and hope that they would have enough in
common with those still unknown. The problem resembles closely those in
software reliability, where the aim is to estimate the unknown number of
" b u g s " remaining in the program. More generally, it is a state estimation or
filtering problem.
 T H E C L A I M S R E S E R V I N G P R O B L E M IN N O N - L I F E I N S U R A N C E 147

It is most convenient to formulate the " c o m m o n elements" in terms of

unobservable (latent) variables whose distribution is updated according to the
information 9-~...~, has thereby an indirect effect on the behaviour of IBNR
claims. In the following we study the expected value and the variance of the
IBNR liability. The presentation has much in common with JEWELL (1980,
1987), and ROBBIN (1986), and in particular NORBERG (1988).
Since the marked points belonging to the settlement process of an unknown
accident are all " i n the future", most considerations concerning the reserves
will not change if the payments are assigned directly to the reporting time T,..
This is possible because we, as stated before, don't consider the effects of
interest rate or inflation. This will simplify the notation to some extent. We
therefore consider the MPP (T,., Xi) , where X; = Xi(ov ) is the size of the claim
caused by the i r;' accident. The corresponding counting process is
{Nt(A); t > 0, A c Rt}, where

(4.1) Ni(A) = Z I{T~t,X,~AI

i

counts the number of accidents reported before t and such that their liability X;
is in the set A. (Note that N r (A) cannot in general be determined from .y-~since
the Xi's counted before t may also include payments made after time t. Also
observe the connection to (3.14): Nt = N~(RI)).
For the purpose of using the apparatus of the stochastic calculus we start by
writing the total liability from IBNR claims as an integral (pathwise):

(4.2) 2 Ui, = ~ Xi = x dN=(dx) .

i>N~ {i:Ti>t} I= I=
s=t .r=0

We also let

U(t, u; A ) = 2
{i:t<Tt~u, XieA}
X; = II
s=t xEA
x dN=(dx),

so that ~2. U , = 0 ( t , co; RI).

i> N t

Adapting the idea from NORBERG (1986) we now suppose that the above
mentioned latent variables form a pair (~, O) and are such that q~ can be
viewed as a parameter of the distribution of the process (Nt), formed by the
reporting times, whereas O parametrizes the distribution of the claim sizes (X/).
(Note that this simple model is " s t a t i c " in the sense that the latent variables do
not depend on time. This assumption could be relaxed, for example, by
introducing an autoregressive scheme of state equations, as in the Kalman
filter). There are no restrictions on the dimension of (~, O). On the other hand,
148 ELJA ARJAS

these parameters are assumed to be sufficient in the sense that if q~ and O,

together with some initial information J 0 , were known, no information from
.Y-t would change the prediction concerning the IBNR claims after t. Thus the
estimates of q~ and O which are obtained from .9",, or more exactly, their
conditional distribution given.Y-t, can be said to include " t h a t part of
J r i n f o r m a t i o n which is relevant in the IBNR-problem"
The formal expression of this idea is as follows. Fixing t (" the present") we
consider times u > t and define

(4.3) .Y,, = ~o v o-{~, O} v a { ( T , , X 3 ; t < 7;,. _< u}.

Thus .-~oorepresents the information contained collectively in .~0, the parame-

ters • and O, and all post-t payments, cf. KARR (1986), Section 2.1. We then
assume the conditional independence property

(4.4) --~o~ U J~t,

Jov a;q~, O}

stating that 3, is irrelevent for predicting the post-t payments provided that
J-0 and (q~, O) are known.
Let the (ff,)-intensity of counting process (N,,(A)), ~ t, be (~,(A)),,>t, with
A c R t. The probabilistic interpretation of 2,,(A) is that

(4.5) i , ( A ) d u = P(dN,,(A) = !1 ...~,,_) = P ( T i e du, X, eA [ ~._)

on the interval ~ _ l < u < Ti. On the other hand, ~,(A) can obviously be
expressed as the product

(4.6) ~,(A) = 7.,, ~0,,(A),

where 2, = ).,,(R I) and ~o~(A)/2,, (cf. KARR (1986), Example 2.24). Here (2,)
is the (ff,)-intensity of the counting process (N,), i.e., 2,,du =
P ( d N , = 1 [ 57,_) = P ( T i e d u [ . ~ _ ) for T~_, < u _ < Ti, whereas ~0,,(A) can
be interpreted as the conditional probability of { X i e A } given ...~,_ and that
TieR.
It follows from (4.4) that the intensity (7~,(.)),>, can be chosen to be
.y-0-measurable and parametrized by (~, O). According to " t h e division of
roles of • and O " we now assume that in fact X, in (4.6) is parametrized by ~,
and ~0,(.) by O..~, can then be expressed in the form 2, = h (u; ~),where, for
fixed ~, u~----,h(u; q~) is .g0-measurable. This is only another way of saying that
the reporting process (N,) is assumed to be a doubly stochastic (non-
homogeneous) Poisson process (or Cox process) with random parameter ~.
 T H E C L A I M S R E S E R V I N G P R O B L E M IN N O N - L I F E I N S U R A N C E 149

Similarly, we assume that the claim size distributions tp,(.) can be written as
~o,(A) = F,,(A; 0), where, for fixed u and O, F , ( . ; O) is a distribution function
on R~_. This, then, amounts to saying that, given 0 and the (unobserved)
IBNR reporting times, the claim sizes X i are independent.
We n°w derive an expressi°n f°r the expected lBNR-liability E : ' ( >~N, Ui' )

First note that fit = ,70 v a ( ~ , 0). By a straightforward calculation we get

that

E J' (z)
i>Nr
Uil = E:' (I I
u=t x=O
xdN~(dx) )
= E" x (dx)du
(*) u=, x=o

= h(u; ~) x F~(dx; O) du
U=I .x'~O

O9
=
i h ( u ; q~) m u ( 0 ) du,

where rnu(O) is the mean

(4.7) mu(O) =
Sx=O
x F~(dx; 0 ) .

(The equality (*) here is a simple consequence of the definition of (~,); for a
general result see e.g. KARR (1986, Theorem 2.22). On the other hand, because
of the conditional independence (4.4), we have that

and therefore finaly

(4.8) E:' (z)I i>Nr

Ui, =
u=t
e:, (h (u; ~) me(O)) du.

We consider some special cases at the end of this section.

Let us then go over to calculating the corresponding conditional variance

expression V a r : ' ( ~ , . , Ui, ) . The calculation goes as follows.

150 ELJA ARJAS

= Var Z x dN,,(dx) = E'L(( U(t, "; R ~) )o~)

u = t x = 0 (*)

= E', ( O(t,.;ax))oo = E, xZa(Nc.)(dx)),

x=O ~'*s x=O u=t

= E 7' x21,,(dx) du = h(u, qb) x2F.(dx;19) du

X=0 11=1 //=[ X=0

oo

=
i /d:f
h (u; 4) m~2) (19) du,

where m(u2)(19) is the second moment

oo

(4.9) rn<,2)(19)=
sx=O
x2 Fu(dx; 19).

(Here ( ( . ) , , ) is the predictable variation process, see e.g. KARR (1986),

Appendix B, (*) is a direct consequence of the definition of this process, and
(**) follows from Theorem B.12 in KARR (1986)). Therefore, and again using
the conditional independence (4.4),

(4.10) Var~'(~i , Ui') =EJ'Var)'( 2i>N,Ui') +Var';'E'L ( 2i>N,Ui,)

=
i ll=l
EJ,(h(u; qs) m~2)(O)) d u + V a r J,
U=I )
h(u; q~)rn.(O) du .

The formulas (4.8) and (4.10) can be briefly summarized by saying that the
conditional expectation and the conditional variance of the IBNR liability
Ui, can be obtained if the following are known:
i> N r

(i) the intensities h(.; 4);

(ii) the first two moments of the distributions F(.; 19), and
(iii) the conditional distribution of the latent variables (4, O) given ~ .

Concerning (i), the common expression for h(.; 4) (e.g. RANTALA (1984)) is
obtained by assuming that during the considered year ( = unit interval (0, 1])
 THE CLAIMS RESERVING PROBLEM IN NON-LIFE INSURANCE 151

accidents occur according to the Poisson(~)-process, and that the reporting

delays D~ are i.i.d, and distributed according to some k n o w n distribution G ( ' ) .
Then it is easily seen that

(4.11) h(u; ~) = ~[G(u) - G((u- 1)÷)].

M o r e generally, • can parametrize both the occurrence process and the

distribution o f the delays in the reporting, cf. JEWELL (1987).
The simplest case in (ii) is o f course when only the n u m b e r N - N t o f future
claims is considered, instead o f the liability they cause. Then we can make the
obvious convention that every X i = I, giving m s ( O ) = m~2)(O) = 1.
Requirement (iii), finally, strongly supports the use o f the Bayesian para-
digm. It is particularly appealing to use the P o i s s o n - g a m m a conjugate distribu-
tions for the pair (Nt, # ) since this makes the updating extremely simple (see
GERBER (1979) and NORBERG (1986)). Since deciding on claims reserves is a
m a n a g e m e n t decision, rather than a problem in science in which some physical
constant needs to be determined, Bayesian arguments should not be a great
deterrent to a practitioner. C h o o s i n g a reasonable prior for (~, O) could be
viewed as a g o o d o p p o r t u n i t y for an actuary to use, in a quantitative fashion,
his experience and best hunches.

ACKNOWLEDGEMENT

M o s t o f what I today know a b o u t insurance mathematics l have learned,

during m a n y long conversations, from JARMO JACOBSSON, of the Statistical
Center o f the Finnish Insurance Companies. I also appreciate the referees'
careful reading o f the paper.
This paper was read at the A S T I N Annual Meeting which was held in
Helsinki in July 1988, in connection with the 23 rd International Congress o f
Actuaries.

REFERENCES
BR~MAUD, P. (1981) PohU Processes and Queues: Marthlgale Dynamics. Springer-Verlag, Berlin.
B~HLMANN,H., SCHNtEPER,R. and STRAUS,E. (1980) Claims reserves in casualty insurance based
on a probabilistic model. Bullet#l of the Association of Swiss Actuaries, 21--45.
GERBER, H.U. (1979) An huroduction to Mathematical Risk Theory. S.S. Huebner Foundation of
Insurance Education, University of Pennsylvania, Philadelphia.
HACHEMEISTER, C.A. (1980) A stochastic model for loss reserving. Transactions of the 21 ~'
International Congress of Actuaries I, 185-194.
JEWELL, W.S. (1980) Generalized models of insurance business. Report of Introduction. Transac-
tions of the 21 s' International Congress of Actuaries S, 87-141.
JEWELL, W.S. (1987) Predicting IBNYR events and delays I. Continuous time. Department of
Industrial Engineering and Operations Research, University of California at Berkeley, Research
Report.
KARR, A. F. (1986) Point Processes and Their Statistical Inference. Dekker, New York.
LINNEMANN,P. (1980) A multiplicative model for loss reserves: A stochastic process approach.
University of Copenhagen, Laboratory of Actuarial Mathematics, Working Paper No. 32.
NORBERG, R. (1986) A contribution to modelling of IBNR claims. Scandinavian Actuarial Journal
1986, 155-203.
152 ELJA ARJAS

NORROS, I. (1985) Systems weakened by failures. Stochastic Processes Appl. 20, 181-196.
PENTIK.~INEN, T. and RANTALA, J. (1986) Run-off risk as a part of claims fluctuation. ASTIN
Bulletin 16, 113-147.
RANTALA, J. (1984) An application of stochastic control theory to insurance business. Acta
Universitatis Tamperensis A, 164 (Academic dissertation).
REID, D. H. (1981) A method of estimating outstanding claims in motor insurance with applications
to experience rating. Colloque: Les math6matiques en Sciences Actuarielles-1981. Institut des
Hautes Etudes de Belgique, Brussels, Belgium, 275-289.
Roaam, 1. (1986) A Bayesian credibility formula for IBNR counts. Proceedingsof the Cos. Actuarial
Society Vol. LXXlll (1986), No. 139-140, 129-167 (with discussion).
TAYLOR, G.C. (1986) Claims Reserv#lg in Non-Life Insurance. North-Holland, Amsterdam.
VAN EEGHEN, J. (1981) Loss Reserving Methods. Surveys of Actuarial Studies No. 1, Nationale
Nederlanden N.V.

ELJA ARJAS
Department o f Applied Mathematics and Statistics, University o f Oulu,
90570 Oulu, Finland.