North American Actuarial Journal

ISSN: 1092-0277 (Print) 2325-0453 (Online) Journal homepage: www.tandfonline.com/journals/uaaj20

Bayesian Estimation of Outstanding Claim

Reserves

Enrique de Alba

To cite this article: Enrique de Alba (2002) Bayesian Estimation of Outstanding Claim Reserves,
North American Actuarial Journal, 6:4, 1-20, DOI: 10.1080/10920277.2002.10596060

To link to this article: https://doi.org/10.1080/10920277.2002.10596060

Published online: 04 Jan 2013.

Submit your article to this journal

Article views: 219

View related articles

Citing articles: 3 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=uaaj20
 BAYESIAN ESTIMATION OF OUTSTANDING
CLAIM RESERVES
Enrique de Alba*

ABSTRACT
This paper presents a Bayesian approach to forecasting outstanding claims, either the total number
of claims or the total amount, that is used for claims reserving. The assumption is made that there
is complete information for one or two past years of origin and partial information for some
development years of other years of origin. It also assumes payments are made annually and that
the development of partial payments follows a stable payoff pattern from one year of origin to
another. Two different models are presented: one for the number of claims (intensity) and one for
claim amounts (severity). The advantage of using this procedure is that actuaries can derive the
complete predictive distribution of the reserve requirements, from which, in turn, it is possible to
obtain point estimates as well as probability intervals and other summary measures, such as mean,
variance, and quantiles.

1. INTRODUCTION periods) it takes for the claims to be completely

Because reserves constitute the largest liabilities paid is fixed and known, payments are made an-
carried by casualty insurers, one of the main nually, and the development of partial payments
activities of a practicing actuary is loss reserving. follows a stable payoff pattern. This is in agree-
The many uncertainties involved in the payment ment with many existing models for claims re-
of losses make the estimation of the required serving in nonlife (general) insurance that as-
reserves more difficult. Yet, some of the existing sume, explicitly or implicitly, that the proportion
methods, such as the popular chain-ladder, are of claim payments, payable in the j-th develop-
simple to apply and have been in use for many ment period, is the same for all periods of origin
years. However, it has become evident that there (see Taylor 1986). This assumption is also
is a need for better ways, not only to estimate the pointed out in the introduction to Hesselager and
reserves, but also to obtain some measures of Witting (1988).
their variability as well as information on their The results are applicable to any frequency of
overall probable future behavior. This has led to claim payments (years, quarters, and so forth)
the development of stochastic reserving models and length of payoff period. Presented here is a
developed by Taylor (2000) and England and Ver- Bayesian approach to forecasting total claims
rall (2002). (number or amount), given a few (one or two)
This paper presents an application of Bayesian years of past data and information on some de-
forecasting methods to insurance, specifically to velopment years for several occurrence years. Es-
the estimation of reserves for outstanding claims. sentially the data would correspond to a typical
Assumptions include that the time (number of runoff triangle used in loss reserving.
Goovaerts et. al. (1990) use the term “incurred
but not reported” (IBNR) reserves both “for claims
as yet unreported, and a reserve for claims known
* Enrique de Alba is Professor of Actuarial Science, Department of to the company but not completely paid” (p. 243).
Actuarial Science and Insurance at Instituto Tecnológico Autónomo
This is in agreement with what Brown (1993) calls
de México (ITAM), Rio Hondo No. 1, San Angel, Mexico, D.F. 01000,
Mexico, and an Adjunct Professor, Department of Statistics and
“gross IBNR.” Hence this paper refers indistinctly
Actuarial Science, University of Waterloo, Waterloo, Ontario N2L either to outstanding claims, IBNR reserves, or loss
3G1, Canada, e-mail: dealba@itam.mx. reserving. Formulation of the problem is similar to

1
2 NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 6, NUMBER 4

Norberg (1986, 1993). Previous results are ex- tinue to hold, and then apply the pattern to esti-
tended using a full Bayesian model. In fact, two mate the claims that have been incurred but are
different models are presented: one to forecast the still outstanding.
number of outstanding claims and one for total However, mechanical application of any statis-
aggregate claims. tical method does not lead to a “correct” result,
As mentioned in England and Verrall (2002), and the result obtained will often need to be
“there is little in the actuarial literature which heavily qualified. There may be trends in the
considers the predictive distribution of reserve settlement pattern, for example, that are difficult
outcomes; to date the focus has been on estimat- or impossible to explain, or there may be some
ing variability using prediction errors” (p. 52). doubt about the likelihood of an existing trend
They found it difficult to obtain analytically these continuing in the future. In such circumstances, a
distributions, which take into account both the great deal of reliance cannot be placed on the
process variability and the estimation variability. projected future runoff. Or the settlement pat-
The models presented here allow the actuary to terns of past years may be so erratic that, al-
provide not only point estimates of the required though a result can be obtained by mechanically
reserves, but also some measures of dispersion, applying one of the available methods, confidence
such as the variance, as well as the complete in the final result will not be very great.
distribution for the reserves. This will make it Even if the past settlement pattern is reason-
feasible to compute other risk measures, for ex- ably stable, the future runoff may be quite uncer-
ample, the value at risk. tain because of doubts that the pattern will con-
The paper is structured as follows. Section 2 tinue, and because of claim inflation. The
gives a brief description of IBNR and previous provision to be held also will be affected by as-
results relevant to the approach presented here. sumed investment earnings. In practice, it is sel-
Section 3 introduces some Bayesian concepts and dom possible to do any better than suggest a fairly
their applications in actuarial science. Section 4 wide range of not-unreasonable provisions based
describes a Bayesian model for the number of on different assumptions about future claim infla-
claims in IBNR. Section 5 presents a Bayesian tion, investment earnings, and so forth, and pos-
model for claim amounts. Some examples are sibly on different statistical methodologies. For an
given in Section 6, and Section 7 offers some final extensive description of loss reserving concepts
remarks. All types of model are presented only in and methods see Goovaerts et al. (1990), Hossack
discrete time. et al. (1999), or Taylor (2000).
All IBNR methods, such as mechanical smooth-
ing, least squares, credibility, and regression can
2. LOSS RESERVING be applied to different claim figures (count,
The settlement of claims is usually subject to amount, and so forth), each of them leading to a
delay, and it is necessary for the insurer to set up specific forecasting method. Goovaerts et al.
reserve provisions for claims corresponding to (1990) describe some real and practical examples
losses that have been incurred by the insured of insurance problems that can be solved by IBNR
during the covered period but have not yet been techniques. In addition, they illustrate that not
settled. It is very important to estimate outstand- only loss reserving problems, but also any prob-
ing claims as accurately as possible to have a lem where a constant development pattern arises
correct view of an insurer’s financial situation. can be solved by these techniques. A broad range
Methods used to estimate the necessary reserve of potential applications can be found in de Alba
provisions are usually classified as deterministic and Mendoza (1996), Oliver (1987), and the ref-
or nonstochastic and statistical or stochastic erences contained therein.
(Hossack et al. 1999 and Taylor 2000). Verrall (1991) considers estimating claim re-
Different statistical approaches for dealing with serves and outstanding claims by means of loglin-
the problem have been developed in recent years. ear models. He compares these with maximum
These methods attempt to find a consistent claim likelihood estimates. Renshaw and Verrall (1994)
runoff pattern that has applied in the past, as- carry out the analysis assuming a Poisson model
sume that pattern is stable and that it will con- both for number of claims and claim amounts,
BAYESIAN ESTIMATION OF OUTSTANDING CLAIM RESERVES 3

while indicating that if a quasi-likelihood ap- known or observed variables (the upper left-hand
proach is taken, then its use will be justified for side) and a set of variables whose values are to be
data that are not positive integers. predicted (the lower right-hand side). Thus, we
In Verrall (2000), he further explores the rela- know the values Xit for i ⫹ t ⱕ k ⫹ 1, and the
tion between the chain-ladder technique and triangle of known values is the runoff triangle.
some stochastic models. Again he assumes a Pois- This formulation agrees closely with Norberg
son distribution for claim amounts and indicates (1986). It is the one used here and appears in
that this model should not necessarily be used for Table 2.
all data. Wright (1990) models claim number,
claim magnitude, and payment delay separately.
He claims that mean severity is not necessarily 3. BAYESIAN MODELS
the same for all development periods and pro- Bayesian methods will be described very briefly
poses a model in which expected payment is a here and discussed with respect to their applica-
function of the delay. Estimation is done by tions in actuarial science, specifically in loss re-
means of the Kalman filter. serving. For general discussions on Bayesian the-
For a comprehensive, although not exhaustive, ory and methods see Berger (1985), Bernardo and
review of existing stochastic methods, see En- Smith (1994), or Zellner (1971). For a discussion
gland and Verrall (2002). Although they provide of Bayesian methods in actuarial science see
some Bayesian results, most of the methods pre- Klugman (1992), Makov et al. (1996), Makov
sented there approach the problem from the (2001), and Scollnik (2001).
point of view of frequentist or classical statistics If the random variables Xit (i ⫽ 1, . . . , k; t ⫽
and in the framework of generalized linear mod- 1, . . . , k) denote claim figures (or loss ratios,
els. They provide predictions and prediction er- claim frequencies, or burning cost, i.e., the total
rors for the different methods discussed and show excess claims divided by the underlying premium
how the predictive distributions may be obtained volume), the (observed) runoff triangle has the
by bootstrapping and Monte Carlo methods. structure in Table 2. The unobserved variables
This paper uses standard notation, so that Xit ⫽ (the lower triangle) are to be estimated. Let
number (or amount) of events (claims) in the t-th f(xit兩␪) be the corresponding density function, so
development year corresponding to year of origin that L(␪兩x) ⫽ 兿i⫹tⱕk⫹1 f(xit兩␪) is the likelihood
(or accident year) i. Thus, {Xit; i ⫽ 1, . . . , k, t ⫽ function for the parameters given the data in the
1, . . . , s} where s ⫽ maximum number of years upper portion of the triangle. Available informa-
(subperiods) it takes to completely pay out the tion on the parameters is incorporated through a
total number (or amount) of claims correspond- prior density ␲(␪) that must be modeled by the
ing to a given exposure year. In this paper, we actuary.
assume Xit ⬎ 0 for all i ⫽ 1, . . . , k and t ⫽ This is then combined with the likelihood func-
1, . . . , s. These random variables can be ar- tion via Bayes’ Theorem to obtain a posterior dis-
ranged in a matrix, as in Table 1. tribution for the parameters: f(␪兩x) ⬀ L(␪兩x)␲(␪).
For most IBNR methods, it is usually assumed When interest centers on inference about the pa-
that s ⫽ k and the matrix is split into a set of rameters, it is carried out using f(␪兩x). When inter-

Table 1
Matrix of Claims by Year of Origin and Development Year

Year of Origin Development Year

1 2 ... t ... s
1 X 11 X 12 ... X 1t ... X 1s
2 X 21 X 22 ... X 2t ... X 2s
3 X 31 X 32 ... X 3t ... X 3s
⯗
k X k1 X k2 ... X kt ... X ks
4 NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 6, NUMBER 4

Table 2
Observed Claims Data Summarized in a Runoff Triangle

Year of Origin Development Year

1 2 ... t ... k⫺1 k
1 X 11 X 12 ... X 1t ... X1,k⫺1 X 1t
2 X 21 X 22 ... X 2t X2,k⫺1 —
3 X 31 X 32 ... X 3t —
⯗ — —
k ⫺ 1 Xk⫺1,1 Xk⫺1,2 — —
k X k1 — —

est is on prediction, as in loss reserving, the past ried out directly from this distribution and any of
(known) data in the upper portion of the triangle, x its characteristics and properties, such as quan-
⫽ (x11, . . . , xkk), are used to predict the observa- tiles, can be used for this purpose. However, if the
tions in the lower triangle zit by means of the pos- predictive distribution is not of a known type, or
terior predictive distribution if it does not have a closed form, then it is possible

冕
to derive approximations by means of Monte
Carlo simulation methods.
f共 z it 兩x兲 ⫽ f共zit兩␪兲 f共␪兩x兲 d␪,
One alternative is the application of direct
Monte Carlo, where the random values are gen-
兵i ⫽ 1, . . . , k, t ⫽ 1, . . . , k, with i ⫹ t ⬎ k ⫹ 1其. erated directly from their known distribution,
The Bayesian approach constitutes a powerful which is assumed to be available in an explicit
formal alternative to deterministic and classical form (see Tanner 1996, Sect. 3.3.2, p. 54). In
statistical methods when prior information is recent years, Monte Carlo methods have devel-
available. But they can also be used when there is oped extensively, in particular, Markov chain
no agreement on the prior information or when Monte Carlo (MCMC) methods. For their use in
there is a total lack of it. In this last situation use actuarial science, see Scollnik (2001).
can made of so-called noninformative or refer- Hence, in loss reserving, the benefit of the
ence priors. Inference under these circumstances Bayesian approach is in providing the decision
is known as objective Bayesian inference (Berger maker with a posterior predictive distribution for
1985). every entry in the lower triangle of the runoff
The use of Bayesian ideas and techniques in triangle and, consequently, for functions of them,
actuarial science dates back to the late 1960s like their sum, that is, the required reserves. Ad-
when the papers of Bühlmann (1967) and others equate understanding and knowledge of the com-
laid the foundation for empirical Bayes credibil- plete distribution is essential. Such a distribution
ity, although the Bayesian contents of this ap- allows the assessment of the required reserves in
proach have been greatly minimized. To date, terms not only of point estimators. That is, in
Bayesian methodology is used in various areas addition to expected values, we are looking at
within actuarial science, among them loss reserv- certain characteristics of probability distribu-
ing. However, in this case, relatively little has tions, for example, probability of ruin, value at
been done from the Bayesian perspective. Its in- risk, and so on.
frequent use in the past is probably attributable As mentioned, actuarial science is a field where
to the fact that application has been, up until very frequently one has considerable prior infor-
recently, computationally cumbersome or down- mation, be it in the form of global or industrywide
right impossible. information (experience) or in the form of tables.
In many cases, Bayesian methods can provide Bayesian methods provide a natural way to in-
analytic closed forms for the predictive distribu- corporate this prior information, whether statis-
tion of the variables involved, for example, out- tical or not, in the form of tables or as expert
standing claims; predictive inference is then car- judgment, through the prior distribution of the
BAYESIAN ESTIMATION OF OUTSTANDING CLAIM RESERVES 5

parameter, ␲(␪). In this respect, it is indeed sur- are given a credibility interpretation. Two alter-
prising that Bayesian methods have not been native formulations are considered, one with no
used more extensively up to now. There is a prior information and another where he uses a
wealth of “objective” prior information available specific prior, ␲(␪). In a more recent paper, Haas-
to the actuary that can be used to specify the trup and Arjas (1996) model the occurrence and
prior distribution ␲(␪). development of claims as marked point process. A
A standard measure of variability is prediction claim is described by an occurrence time, two
error, defined as the standard deviation of the covariates, and a development. The latter con-
distribution of possible reserves. In the Bayesian tains a reporting delay, a settlement delay, and a
context, the usual measure of variability is the partial payment process. The distribution of the
standard deviation of the predictive distribution claims process is described by 14 components. A
of the reserves. This is a natural way of doing priori, the intensities are assumed piecewise con-
analysis in the Bayesian approach and, as noted stant and their levels, as well as the intervals, are
in England and Verrall (2002), they “offer the the parameters. This allows them to apply the
best way forward” (p. 21). nonparametric approach of Ferguson (1973). The
Bayesian analysis of IBNR reserves has been analysis is done using Gibbs sampling, a particu-
considered before by Jewell (1989, 1990), Verrall lar form of MCMC.
(1990), and Haastrup and Arjas (1996). Jewell More recently, Bayesian results are provided in
(1989) formulated the problem considering that England and Verrall (2002), notably for the Born-
an IBNR “. . . claim in insurance is an event huetter-Ferguson (B-F) technique. In the tradi-
whose occurrence in some fixed exposure inter- tional B-F method, use is made explicitly of
val is not known until some later date because of perfect prior (expert) knowledge of “row” param-
random reporting delays” (p. 25). He considered eters. This is clearly well suited for the applica-
that the natural formulation of the IBNR problem tion of Bayesian methods when knowledge is not
is in continuous time because of the underlying perfect (England and Verrall 2002). Ntzoufras and
Poisson generation of claims and continuous re- Dellaportas (2002) consider various competing
porting of delays. His model considers different models using Bayesian theory and MCMC meth-
types of data that may become available, such as ods. Claim counts are used to add a further hier-
occurrence and reporting dates, and concludes archical stage in the model with log-normally dis-
that at least the observation of reporting dates tributed claim amounts and its corresponding
must be available. He assumes a homogeneous state space version. This way, they incorporate
process but recognizes that this may be unneces- information from both the outstanding claim
sarily restrictive. amounts and count data resulting in new model
In a sequel paper, Jewell (1990) considers the formulations. It is similar to what we do here.
more realistic situation where reporting occurs The chain-ladder is used as a benchmark in
only periodically, so that the reporting of dates is several of the references mentioned above, be-
discretized into intervals of the basic exposure cause of its generalized use and ease of applica-
interval. Doray (1996) presents an UMVUE (uni- tion. This facilitates comparison between meth-
formly minimum variance unbiased estimator) of ods. However, this paper does not aim to develop
IBNR reserves in a lognormal regression model, in Bayesian methods that provide results close to
which he can compute its variance. The last sec- those of the chain-ladder method. Rather, it aims
tion of this paper uses his data and compares the at using “objective” Bayesian methods to model
present results with his. both claim intensity and severity using some
Verrall (1990) approaches the subject of pre- common assumptions and to use the resulting
dicting outstanding claims using hierarchical predictive distributions to estimate loss reserves.
Bayesian linear models, considering the fact that
the chain-ladder technique is based on a linear
model: the two-way analysis of variance model 4. A MODEL FOR THE NUMBER OF CLAIMS
(ANOVA). He essentially carries out a Bayesian De Alba (1988) and de Alba and Mendoza (1996)
analysis of the two-way ANOVA model to obtain considered a model that can be stated in terms of
Bayes and empirical Bayes estimates. The latter IBNR as follows: Let Xit ⫽ number of claims in t-th
6 NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 6, NUMBER 4

development year corresponding to year of origin each year of origin, using the available informa-
(or accident year) i, that is, {Xit; i ⫽ 1, . . . , k tion in the runoff triangle.
t ⫽ 1, . . . , k}. Let m ⫽ number of complete years To estimate the unknown number of claims for
of origin for which there is complete data. Under succeeding accident years, we use a multinomial
the usual structure of the runoff triangle, m ⫽ 1, model as previously and proceed as follows. Un-
that is, the only year of origin for which all the der the same assumptions about the distribution
claims have been made is the first one. However, of the number of claims and notation as in Sec-
with the current procedure, we can easily accom- tion 2, let ai ⫽ k ⫺ i ⫹ 1 and define x*i ⫽ ¥t⫽1
ai
xit.
modate any other value of m; in what follows, we Then, using properties of the multinomial distri-
take m ⫽ 1, so the available information is bution (Bernardo and Smith 1994), write

兵Xit; i ⫽ 1, . . . , k t ⫽ 1, . . . , k i ⫹ t ⱕ k ⫹ 1其. f k⫺i⫹1 共x i兩ni, p兲 ⫽ f1共x*i 兩ni, p*k⫺i⫹1兲

Thus, we have the runoff triangle as given in

Table 2.
冉 冏 冉 冊
⫻ fk⫺i xi1, . . . , xi,k⫺i⫹1 x*i,
p1
p*k⫺i⫹1
,...,
pk⫺i⫹1
p*k⫺i⫹1
, 冉 冊冊
Let N1 ⫽ Total number of claims corresponding (3)
k
to year of origin 1, so that ¥t⫽1 X1t ⫽ N1. Assume
that the number of claims in the t-th development i ⫽ 1, . . . , k, where fk(䡠兩n, u) denotes a k-dimen-
year follows a Poisson distribution with parame- sional multinomial pf, and f1(䡠兩n, ␪) is a Bin(n, ␪).
ter ␭t. This is similar to Assumption I in Norberg The vectors x1 ⫽ ( x11, x12, . . . , x1k)⬘, x2 ⫽ ( x21,
(1986), with fixed general risk conditions. How- x22, . . . , x2k⫺1)⬘, . . . , xk ⫽ ( xk1) contain the
ever, we use a different approach here based on known data for the number of claims in each row
conditional probabilities. Suppose N1 is given, so of the triangle. Notice that in Equation (3), when
that by conditioning on N1 and assuming that the i ⫽ 1, f1( x*1兩n1, p*k) ⫽ 1 and p*k ⫽ 1, so that
number of claims in one development year is fk(x1兩n1, p) ⫽ Multk(n1; p) as in Equation (1).
independent of the number in another year, the The joint pf of the number of claims for all cells
vector X1 ⫽ (X11, X12, . . . , X1k)⬘ has a k-dimen- is then
sional multinomial distribution, that is,
写f
k

X 1兩N1 ⫽ n1 ⬃ Multk共n1; p1, . . . , pk兲, (1) f共x 1, . . . , xk兩n1, n2, . . . , nk, p兲 ⫽ k⫺i⫹1共xi兩ni, p兲.
i⫽1

with (4)
Since n1 is known, this is the likelihood function
冘 ␭,
k

p t ⫽ ␭ t/ t
for (N2, N3, . . . , Nk, p). Next, we specify a prior
t⫽1 distribution for the parameters, assuming they
are independent a priori, so that
(Bernardo and Smith 1994). Thus, by condition-

写 f共n 兲 f共p兲
ing on the total number of claims of origin year 1, k

we go from a parameterization in terms of ␭’s to f共n 2 , . . . , n k , p兲 ⫽ i

one in terms of the p’s. From the assumption of a i⫽2

stable pattern of development, which is also in- and with noninformative priors
cluded in Norberg (1986), it follows that pt does
not depend on the year of origin i. If m ⬎ 1, we
写 共1/n 兲.
k
could let X* ⫽ X1 ⫹ . . . ⫹ Xm so that f共n 2 , . . . , n k , p兲 ⬀ i (5)
i⫽2
X*兩N1 ⫽ n1, . . . , Nm
This follows from the noninformative prior of
⫽ n m ⬃ Multk共n1 ⫹ . . . ⫹ nm; p1, . . . , pk兲. for ni

(2) f共n i 兲 ⬀ 1/n i , (6)

In the IBNR setup, we are interested in estimat- assuming that Ni ⬎ 0, i ⫽ 2, . . . , k (de Alba and
ing N2, N3, . . . , Nk, the total number of claims for Mendoza 1996). The prior for p is a Dirichlet
BAYESIAN ESTIMATION OF OUTSTANDING CLAIM RESERVES 7

(1, 1, . . . , 1) (O’Hagan 1994). The joint posterior i⫹1

¥j⫽1 xjt ⫹ 1. With this, we obtain the posterior
distribution is then mean of the outstanding claims in the i-th acci-
dent year and show that it is given approximately

写 f共n 兩p, D兲 f共p兩D兲,

k
by
f共n 2 , . . . , n k , p兩D兲 ⬀ i
i⫽2 E共Ri兩D兲 ⬵
where D represents all the known information
included in the posterior distribution, that is, D ⫽
{x, . . . , xk, x*1, x*2, . . . , x*k, n1}. We then have
冋 冘 冒冉 冘
k⫺共i⫺1兲

xit
k⫺共i⫺1兲

t⫽1
共x1t ⫹ x2t ⫹ · · · ⫹ x共i⫺1兲t兲 冊册
f共n 2 , . . . , n k , p兩D兲 ⫽ k
n1! k

px1t
兿t⫽1 x1t! t⫽1 t
写 ⫻ 冋i⫺1

冘 冘
k*

j⫽1 t⫽k⫺共i⫺2兲
册
xjt ,

共n2 ⫺ 1兲! x*2! where xjt ⫽ E(Rj兩D) whenever j ⫹ t ⬎ k ⫹ 1. It is

x*2
⫻ 共1 ⫺ p*k⫺1兲x2kp*k⫺1 interesting to note that this result is very similar
xk!共 x*2 ⫺ 1兲!x*2 兿t⫽1 x2t!
k⫺1
to that of Verrall (2000). However, in the present

写冉 冊
k⫺1
pt x2t
共n3 ⫺ 1兲! applications the exact results are used to obtain
⫻ the posterior distribution of the required reserves
p*k⫺1 共n3 ⫺ x*3兲!共x*3 ⫺ 1兲!x*3
t⫽1 via Monte Carlo simulation by iteratively follow-
ing steps 1–7 in Appendix B for j ⫽ 1, . . . , N,
x*3! k⫺2 pt
写冉 冊x3t
⫻ 共1 ⫺ p*k⫺2兲 n3⫺x*3
p* x*3 (Tanner 1996). This process provides a set of N
k⫺2
兿t⫽1
k⫺2
x3t! t⫽1 p*k⫺2 randomly generated values for the number of
claims in each cell of the (unobserved) lower right
· ( j)
triangle xit , i ⫽ 2, . . . , k, t ⬎ k ⫺ i ⫹ 1. The j-th
·
· randomly generated observation for the total
共nk ⫺ 1兲! number of outstanding claims will be
⫻
共nk ⫺ xk1兲!共x*k ⫺ 1兲!x*k

冘 冘
k

写冉 冊
x*k! p1 1 xk1 R 共 j兲
⫽ x it共 j兲 , j ⫽ 1, . . . , N.
⫻ 共1 ⫺ p*1兲nk⫺xk1p*1 xk
. (7) i⫽2 t⬎k⫺i⫹1
xk1! t⫽1 p*1
The mean and variance can be computed as
This is the product of (k ⫺ 1) independent nega-
⫺ R៮ 兲 2
冘 共R 冘
共 j兲
tive binomials for the ni, i ⫽ 2, . . . , k, and a 1 N
1 N 共 j兲
Dirichlet for p. We can then easily integrate ␴ R2 ⫽ and R៮ ⫽ R . (8)
N N N j⫽1
(sum) with respect to the ni, i ⫽ 2, . . . , k, and j⫽1

have f(p兩D). Since we do not have the same infor- The standard deviation ␴R thus obtained is an
mation on p for each accident year, we need to “estimate” for the prediction error of the number
express this posterior pdf differently. Appendix A of claims to be paid. The simulation process has
uses the well-known result that the added advantage that it is not necessary to
obtain explicitly the covariances that may exist
f共p兩D兲 ⬀ f共 pk兩D兲 f共 pk⫺1兩pk, D兲
between parameters. They are dealt with implic-
⫻ f共 pk⫺2兩pk⫺1, pk, D兲. . . f共 p2兩p3, p4. . .pk, D兲 itly.
⫻ f共 p1兩p2. . . pk, D兲,
where 5. A MODEL FOR AGGREGATE CLAIMS
This section presents a model for the unobserved
f共 p k⫺i 兩p k⫺i⫹1 , p k⫺i⫹2 ,. . ., p k , D兲 aggregate claim amounts and, hence, the neces-

⬀ 冉 冊 冉
p k⫺i
p *k⫺i
c k⫺i ⫺1
1⫺
p k⫺i
p *k⫺i 冊 d k⫺i ⫺1
,
sary reserves for outstanding claims. Let the ran-
dom variable Zit ⱖ 0 represent the value of aggre-
gate claims in the t-th development year of
with ck⫺i ⫽ ¥j⫽1
i⫹1
xjk⫺i ⫹ 1 and dk⫺i ⫽ ¥t⫽1
k⫺i⫺1
accident year i, i, t ⫽ 1, . . . , k. As in the model
8 NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 6, NUMBER 4

for the number of claims, let m ⫽ number of years Y it ⫽ ␮ ⫹ ␣ i ⫹ ␤ t ⫹ ⑀ ij ⑀ ij ⬃ N共0, ␴ 2 兲, (10)

of origin, or business years, for which there is
complete data. As above, in this paper we refer to i ⫽ 1, . . . , k, t ⫽ 1, . . . , k and i ⫹ t ⱕ k ⫹ 1, that
years of origin and development years, but the is, an unbalanced two-way ANOVA model (Verrall
procedures are equally applicable with other time 1990 and Doray 1996).
units. Under the usual structure of the runoff Thus, Mit follows a lognormal distribution, and
triangle m ⫽ 1, that is, the only year of origin for f共 y it 兩␮, ␣ i , ␤ t , ␴ 2 兲
which the total claim amount has been paid is the
first one.
As in the previous model, the Zit are known for
⬀
1
␴
1
冋
exp ⫺ 2 共 yit ⫺ ␮ ⫺ ␣i ⫺ ␤t兲2 .
2␴ 册 (11)
i ⫹ t ⱕ k ⫹ 1. Assume that
It is well known in ANOVA that certain restric-

Y it ⫽ log冉冊 Zit
Xit
⫽ log共Mit兲, (9)
tions must be imposed on the parameters to at-
tain estimability. One option is to use ¥i⫽1 n

¥j⫽1 ␤j ⫽ 0 (Kremer 1982). We use the alternative

n
␣i ⫽

where we can use alternative specifications for assumption that ␣1 ⫽ ␤1 ⫽ 0. Also, we define
Xit. The first one of these specifications to con- TU ⫽ (k ⫹ 1)k/2 ⫽ number of cells with known
sider is to let Xit be, as before, the number of claim information in the upper triangle, and TL ⫽
closed claims in the t-th development year corre- (k ⫺ 1)k/2 ⫽ number of cells in the lower triangle
sponding to year of origin i. Hence, we choose to whose claims are unknown.
keep the previous notation. In this case, Mit ⫽ Ntzoufras and Dellaportas (2002) use the fol-
Zit/Xit is the corresponding average claim. This is lowing similar model Yit ⫽ log Mit, Yit ⬃ N(␮it, ␴2)
the structure used in Taylor and Ashe (1983). with ␮it ⫽ ␮ ⫹ ␣i ⫹ ␤t ⫹ log( xit) and where Mit
By modeling the average claim per period, we includes an inflation correction. For the number
implicitly allow for the possibility that it increases of claims, they use ( xi1, xi2, . . . , xik)T ⬃ Mult( p1,
payment delay (Wright 1990). The second speci- p2, . . . , pk; Ni) with log( pt/p1) ⫽ ␤*t. The main
fication considered in this paper is to let Xit ⫽ Xi difference with ours is that they assume Ni is
for all t ⫽ 1, . . . , k, that is, Xi is some measure of known and, therefore, fixed. This is necessary for
exposures in each year of origin, for example, the them to able to use Gibbs sampling for the simu-
size of portfolio in year i. It is used as a standard- lations. They compare four models. The first two
izing measure of business volume. This is the are equivalent to the second and third specifica-
formulation used in most of the references, for tions described in this paper above. We are able
example, Verrall (1990) and Doray (1996). to remove the restriction that Ni is known by
The models in Renshaw and Verrall (1994) and conditioning on the available observations for
England and Verrall (1999) use the same linear each accident year, which leads to Equation (7).
predictor structure, but the error distributions In so doing, it is not possible to obtain the full
are different. The first assumes a Poisson distri- conditional posterior distributions that are nec-
bution while the latter considers both an “over- essary to use Gibbs sampling (Tanner 1996), but
dispersed” Poisson and a gamma distribution. we can still use direct Monte Carlo as in steps 1–7
Also, both are formulated in a generalized linear of Appendix B.
model framework. A possible third specification Using matrix notation, the model in Equation
would be to use Equation (9) to model aggregate (10) can be written as follows:
claim amounts without including any information y ⫽ W␪ ⫹ ⑀ ⑀ ⬃ N共0, ␴2I兲, (12)
on number of claims. That is, use Mit ⫽ Zit or,
equivalently, let Xit ⫽ 1, i, t ⫽ 1, . . . , k. This where y ⫽ { yit; i, t ⫽ 1, . . . , k, i ⫹ t ⱕ k ⫹ 1} is
specification is equivalent to the model of Doray a TU-dimension vector that contains all the ob-
(1996). served values of Yit, ␪ ⫽ (␮, ␣2, . . . , ␣k, ␤2, . . . ,
Since the first formulation of the three men- ␤k)⬘ is the ((2k ⫺ 1) ⫻ 1) vector of parameters, ␧
tioned above is more general (Xit depends on t), is the (TU ⫻ 1) vector of errors, and W is the TU ⫻
let’s consider it in more detail. In either case, we (2k ⫺ 1)) design matrix of the model. This spec-
assume in addition that ification is used in the following sections.
BAYESIAN ESTIMATION OF OUTSTANDING CLAIM RESERVES 9

5.1 LOSS RESERVING USING CLAIMS

写 f共n 兲 f共p兲.
k

PER PERIOD f共␪, ␴, n2, . . . , nk, p兲 ⫽ f共␪兲 f共␴兲 i

i⫽2
Now let’s estimate (or obtain the distribution of)
aggregate claims for accident year i given infor- Using the same noninformative priors for ni and p
mation on at least one year that has fully devel- given in Equation (5) and noninformative priors
oped and perhaps on m previous completely for ␪ and ␴ (Zellner 1971), we get
known accident years. Using a definition analo-
写 共1/n 兲.
k
gous to the one used previously, let Z*it ⫽ ¥j⫽1 t
Zij
for 1 ⱕ t ⱕ k. Hence, in the runoff triangle setup, f共␪, ␴, n2, . . . , nk, p兲 ⬀ 共1/␴兲 i
i⫽2
we are really interested in estimating Z*ik i ⫽
2, . . . , k, given Z*1k, Xit and Zit, i ⫽ 1, . . . , k The joint posterior distribution is then
t ⫽ 1, . . . , k, with i ⫹ t ⱕ k ⫹ 1. Conditioning on
Xit is particularly important; the results obtained f共␪, ␴, n2, . . . , nk, p兩D兲 ⬀ ␴⫺共TU⫹1兲
here must later be combined with the marginal
posterior distribution of the Xit, i, t ⫽ 1, . . . , k, to
estimate outstanding aggregate claims. Now let
⫻ exp ⫺ 冋 1
2␴2
共y ⫺ W␪兲⬘共y ⫺ W␪兲 册
Ri ⫽ Z*ik ⫺ Z*iai, for i ⫽ 2, . . . , k, with ai ⫽ k ⫺ i ⫹
写p
k
n 1!
1, so that Z*iai is the accumulation of Zit up to the ⫻ k x 1t

latest development period and Ri ⫽ the total of 兿 t⫽1 x 1t ! t⫽1

t

the aggregate claims process for the development

years for which it is unknown, both correspond- 共n 2 ⫺ 1兲! x*2
⫻ 共1 ⫺ p *k⫺1 兲 x 2kp *k⫺1
ing to business year i. x 2 !共 x 2 ⫺ 1兲!
*
Hence, using matrix notation, the model in
x*2! k⫺1 pt
写冉 冊 共n3 ⫺ 1兲!
x2t
Equation (12) can be written as
⫻
兿t⫽1
k⫺1
x2t! t⫽1 p*k⫺1 共n3 ⫺ x*3兲!共x*3 ⫺ 1兲!
f共y兩␪, ␴2, W, x1, x2, . . . , xk兲

冋 册 x*3! k⫺2 pt
写冉 冊x3t
1 *
⬀ ␴⫺TU exp ⫺ 共y ⫺ W␪兲⬘共y ⫺ W␪兲 . ⫻ 共1 ⫺ p*k⫺2兲n3⫺x*3 p*k⫺2
x3
2␴2 兿t⫽1
k⫺2
x3t! t⫽1 p*k⫺2

As before, the vectors x1 ⫽ ( x11, x12, . . . , x1k)⬘, ·

·
x2 ⫽ ( x21, x22, . . . , x2k⫺1)⬘, . . . , xk ⫽ (xk1) con- ·
tain the known data for the number of claims in 共n k ⫺ 1兲!
each row of the triangle. Using the same assump- ⫻ 共1 ⫺ p *1 兲 n k⫺x k1p *1 x k
共n k ⫺ x k1 兲!共 x *k ⫺ 1兲!
tions about the distribution of the number of

写 冉 pp*冊
claims and notation as in Section 4, as well as x *k ! 1 x k1
1
from independence of the number of claims and ⫻ , (14)
x k1 ! 1
the average claim per cell, we can rewrite the t⫽1

joint pdf as where D represents all the known information

included in the posterior distribution; that is, D ⫽
f共y, x1, . . . , xk兩␪, n1, n2, . . . , nk, p, ␴, W兲 {x1, . . . xk, x*1, x*2, . . . , x*k, n1, W, y}. We can
rewrite Equation (14) as
写f
k

⫽ f共y兩␪, ␴, W兲 ⫻ 共xi兩ni, p兲,

k⫺i⫹1 (13) f共␪, ␴, n2, . . . , nk, p兩D兲 ⬀ f共␪, ␴兩D兲
i⫽1

写 f共n 兩p, D兲 f共p兩D兲,

k
where fk⫺i⫹1(xi兩ni, p) is defined as in Equation ⫻ i
(3). i⫽2

Next we specify a prior distribution for the

where
parameters. Assume they are independent a pri-
ori so that f共␪, ␴兩D兲 ⫽ f共␪兩␴, D兲 f共␴兩D兲.
10 NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 6, NUMBER 4

Since ␪ ⫽ (␮, ␣2, . . . , ␣k, ␤2, . . . , ␤k)⬘, and from It is not difficult to show that

冋 冏册
the first factor in Equation (14), we can write
pt
E共Xit兩D兲 ⫽ x*i E D .
f共␪兩␴, D兲 ⬀ ␴⫺共2k⫺1兲 p p*ai

⫻ exp ⫺ 冋 1
2␴2
共␪ ⫺ ␪ˆ 兲⬘W ⬘W共␪ ⫺ ␪ˆ 兲 册 (15) However, it is not possible to obtain E(Mit兩D). It is
even more complicated to compute the variance
of the predicted loss reserves.
and
To compute Equation (18), we would need to
f共␴兩D兲 ⬀ ␴ ⫺共T U⫺2k⫹2兲 find Cov(Zis, Zit兩D). But this involves both M and

冋 册
x. Doray (1996) considers this problem when the
1 latter is not involved explicitly, based on the fact
⫻ exp ⫺ 共y ⫺ W␪ˆ 兲⬘共y ⫺ W␪ˆ 兲 , (16)
2␴2 that the vector of “future” values z ⫽ (Zu) has a
with ␪ˆ ⫽ (W⬘W )⫺1W⬘y. This is the “square-root multivariate lognormal distribution. Then, follow-
inverted-gamma” distribution (Bernardo and ing Crow and Shimizu (1988), he gets
Smith 1994). Furthermore, recalling ai ⫽ k ⫺ Cov共Zit, Zjs兲 ⫽ exp关b⬘it␪ˆ ⫹ b⬘js␪ˆ ⫹ ␴˜ 2共␯it ⫹ ␯js兲/ 2兴
i ⫹ 1,
⫻ 兵exp关␳␴˜ 2共␯it␯js兲1/ 2兴 ⫺ 1其.
共ni ⫺ 1兲!
f共n i 兩p, D兲 ⫽ 共 p* 兲x*i 共1 ⫺ p*ai兲ni⫺x*i, where ␳(Zit, Zjs兩D) ⫽ ␳ is the correlation between
共x*i ⫺ 1兲!共ni ⫺ x*i 兲! ai Zit and Zjs, and ␴˜ is some estimate of the standard
(17) deviation. Thus, the covariance for each pair of
elements in the lower triangle would need to be
i ⫽ 1, 2, . . . , k, and from Equation (A1) in Ap-
evaluated to find the variance of the IBNR claims.
pendix A
It is clear that these formulas can be very cum-

写冉 冊
bersome to compute (Doray 1996, England and
写
k
n1! x*2! k⫺1 pt x2t
f共p兩D兲 ⫽ k p ⫻ k⫺1
x1t Verrall 1999, and Verrall 1991), and we still
兿t⫽1 x1t! t⫽1 t 兿t⫽1 x2t! t⫽1 p*k⫺1 would not have the distribution. However, it is
relatively easy to obtain an approximation to the
x*3! k⫺2 pt
写冉 冊
x3t
distribution of the reserves by simulation, as fol-
⫻ . . . ⫻ k⫺2 lows: For j ⫽ 1, . . . , N, use steps 1–7 given in
兿t⫽1 x3t! t⫽1 p*k⫺2
Appendix B to obtain a sample of randomly gen-

写冉 冊
x*k! 1 p1 xk1 erated values for the number of claims in each
⫻...⫻ . cell of the (unobserved) lower right triangle.
xk1! t⫽1 p*1
Then, also for j ⫽ 1, . . . , N, follow steps 8 –12 in
the second portion of Appendix B to generate
To compute the reserves for the outstanding ( j)
random values of the of the average payment Mit
aggregate claims, we would need to estimate the
for each one of those same cells (steps 8 –10), and
lower portion of the triangle by obtaining the
for the corresponding pending aggregate loss pay-
mean and variance of the predictive distribution.
ment Zit ( j)
⫽ Xit
( j)
ⴱ Mit
( j)
, i ⫽ 2, . . . , k and t ⬎ k ⫺
Hence, for each cell we have
i ⫹ 1. These Mit values include both parameter
( j)

E共Zit兩D兲 ⫽ E共Mit Xit兩D兲 ⫽ E共Mit兩D兲E共Xit兩D兲, variability and process variability (see explanation
in Appendix B). Thus, we can compute a random
because of the independence of M and (x1, . . . , xk), value of the total required reserves R( j) ⫽ ¥i,t Zit ( j)
.
where M ⫽ {Mit; i, t ⫽ 1, . . . , k, i ⫹ t ⱕ k ⫹ 1}. Then The mean and variance of the approximate pre-
the Bayes estimate of outstanding claims for year of dictive distribution can be computed using Equa-
business i is ¥t⬎k⫺i⫹1 E(Zit兩D) and the Bayes “esti- tion (8) in Section 4. The R( j), j ⫽ 1, . . . , N, can
mator” of the variance (the predictive variance) for be used to analyze the behavior of reserves. As in
that same year is the model for “number of claims,” it is not nec-

冋 冘 Cov共Z , Z 兩D兲册.
essary to obtain explicitly the covariances be-
冘 Var共Zit兩D兲 ⫹ 2 is it (18) tween parameters, since this process incorpo-
t⬎k⫺i⫹1 s⬎t rates them implicitly.
BAYESIAN ESTIMATION OF OUTSTANDING CLAIM RESERVES 11

Table 3
Cumulative Loss Payments through Development Years

Cumulative Loss Payments (in thousands)

Accident Development Year Ultimate Number
Year 0 1 2 3 4 of Claims
1988 2,000 6,000 9,000 11,200 14,000 1,000
1989 2,600 6,840 10,920 15,600 1,200
1990 2,380 8,960 14,400 1,400
1991 3,120 10,800 1,500
1992 3,800 1,500
Source: Brown (1993).

5.2 Loss Reserving Using a Measure of arrived at using all the information on p and this
Exposure Per Accident Year vector does not appear when using Xit ⫽ Xi ⫽
This section presents the second specification for exposures per accident year, as in the “exposure”
Equation (9). Let Xit ⫽ Xi ⫽ some exposure factor model of Section 5.2, then there is no need to
or the (known) exposures in each year of origin model Xit when it is defined this way.
(Verrall 1990, Renshaw and Verrall 1994, and
England and Verrall 1999). Since Xi, i ⫽ 1, . . . , k 6. APPLICATIONS
are no longer random variables, we only need to
model Mit as before. The expressions for To exemplify the application of the methods pre-
Var(Zit兩D) and Cov(Zis, Zit兩D) will simplify some- sented in the previous sections, three sets of data
what, but they will still be cumbersome to com- will be used: (1) the data given in Tables 4.8 and
pute. Hence, we obtain the predictive distribution 4.10 of Brown (1993), reproduced here as Tables
by simulation but do not need steps 1–7 from 3 and 4; (2) the data of Verrall (1990), originally
Appendix B. We start with step 8, and step 11 will attributable to Taylor and Ashe (1983); and (3)
be modified as follows: the data in Ntzoufras and Dellaportas (2002), for
a comparison of results. All references to simula-
Compute Zit共 j兲 ⫽ Xi ⴱ Mit共 j兲, tion steps are to those in Appendix B.
With the data from Brown (1993), we first an-
all i ⫽ 2, . . . , k, t ⬎ k ⫺ i ⫹ 1. (19)
alyze number of claims and then aggregate
Steps 1 through 10 and 12 remain unchanged. claims. The numbers of claims are given in Table
Using this procedure, we can obtain the predic- 4. The estimated number of outstanding claims
tive distribution of the reserves that will be com- obtained by the Bayesian method using the pos-
parable to those given in the references, but with terior mean, as described in Section 4 (also in
the advantage of having the complete predictive steps 1–7 in Appendix B), is 1,870. Figure 1 shows
distribution. Notice that, since the global model is the corresponding distribution obtained with

Table 4
Cumulative Closed Claims through Development Years

Cumulative Closed Number of Claims

Accident Development Year Estimated Ultimate Number
Year 0 1 2 3 4 of Claims
1988 400 700 850 930 1,000 1,000
1989 480 790 1,000 1,140 1,200
1990 500 950 1,190 1,400
1991 570 1,050 1,500
1992 600 1,500
Source: Brown (1993).
12 NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 6, NUMBER 4

Figure 1 Figure 2 shows the histogram for a sample from

Distribution for the Number of Closed Claims the predictive distribution of the reserves gener-
from Brown’s Data ated with 5,000 simulations of the Bayesian
model following steps 1 through 12 of Appendix B,
using the number of closed claims in each devel-
opment-year/business-year cell (Equation B1). In
this case, the mean of the predictive distribution
of the reserves is $57,414, while the correspond-
ing median is $57,027 and the standard deviation
of the predictive distribution of aggregate reserves
is $5,671. When the model is used with the busi-
ness year exposures, then the predictive mean is
$57,490, the median is $55,731, and the standard
deviation is $12,961. To obtain this, we simulate
using steps 8 –12, with step 11 modified as indi-
cated in Equation (19).
5,000 simulations. The estimate for the number In this case, the exposure is taken as the esti-
of outstanding claims using the chain-ladder mated ultimate number of claims. Figure 3 shows
methods yields 1,861. the corresponding histogram of a sample from the
The estimate of the reserves with the straight- distribution of the reserves. There is a clear dif-
forward classical chain-ladder model, without in- ference between these two Bayesian “estimates.”
cluding exposures or closed claims, is $59,240. However, both of these methods produce the dis-
Brown (1993) also reports other results where the tribution and, hence, are preferable to others
data have been adjusted for speed of finalization. since they yield much more information—we
In this case, the required reserves are $54,863. have an approximation to the complete and exact
We applied our Bayesian model without including distribution. In the case where we use only the
any information on claims or exposure; that is, we aggregate data, the predicted reserves are very
only used the aggregate claim amounts per cell. close to that of the chain-ladder. When we use
This is the third specification of Xit mentioned in more information, they tend to be smaller than
Section 5. To do this, skip steps 1–7 in the sim- the latter but not by very much. There is no
ulation and use Equation (B1), with Mit ⫽ Zit (or evident pattern on the behavior of the standard
Xit ⫽ 1) in step 11. The result is $60,073, very deviation.
close to the chain-ladder estimator, with the ad- Now let’s apply the Bayesian methods to an
ditional advantage of having (an approximation example from Verrall (1990). Again, we first an-
to) the predictive distribution of the reserves. alyze the number of claims and then the claim

Figure 2 Figure 3
Distribution of the Reserves Using Information Distribution of the Reserves Using Information
on Closed Claims from Brown’s Data on Exposure from Brown’s Data
BAYESIAN ESTIMATION OF OUTSTANDING CLAIM RESERVES 13

Table 5
Claim Payments through Development Years from the Portfolio of General Insurance Policies

General Insurance
Amount of Claims
Accident Development Years
Year 1 2 3 4 5 6 7 8 9 10
1 $357,848 $ 766,940 $ 610,542 $ 482,940 $527,326 $574,398 $146,342 $139,950 $227,229 $67,948
2 352,118 884,021 933,894 1,183,289 445,745 320,996 527,804 266,172 425,046
3 290,507 1,001,799 926,219 1,016,654 750,816 146,923 495,992 280,405
4 310,608 1,108,250 776,189 1,562,400 272,482 352,053 206,286
5 443,160 693,190 991,983 769,488 504,851 470,639
6 396,132 937,085 847,498 805,037 705,960
7 440,832 847,631 1,131,398 1,063,296
8 359,480 1,061,648 1,443,370
9 376,686 986,608
10 344,014

Source: Verrall (1990).

amounts. Table 5 shows the claim amounts paid tion of Table 7 shows this, along with the results
by development year for each year of origin. The of estimating the required reserves under the
number of claims is given in Table 6. Numbers in Bayesian model using a measure of exposure per
italics in the lower right triangle of this table are accident year: Simulate using steps 8 –12 of Ap-
the estimates obtained by the chain-ladder pendix B, with step 11 modified as indicated in
model. Equation (19). This is the specification that can
In Figure 4 we show the distribution of the be compared to the results of Verrall (1990),
number of outstanding claims obtained using the Mack (1991), Renshaw and Verrall (1994), and
this model, simulated using steps 1–7 of Appendix England and Verrall (1999), since all are based on
B. From the simulation, we obtained a mean of the use of some measure of business year expo-
919 for the number of outstanding claims, with a sures.
standard deviation of 79.51. These values com- Figure 5 shows the predictive density of the
pare with 1,270 from the chain-ladder method. reserves using this model. The mean is $19,630
This value is in the extreme tail of the distribution and the standard deviation $3,284. These values
and is, thus, seen to be overly “conservative.” are slightly larger than those obtained by Verrall
Results of applying the Bayesian methods to (1990) using prior information, but they are es-
obtain estimates of the required reserves follow. sentially the same as when he uses no prior in-
The regular chain-ladder method yields an esti- formation or the chain-ladder. This may be attrib-
mate for the reserves of $18,681. The upper por- utable to the fact that the latter method utilizes

Table 6
Claims Finalized through Development Years from the Portfolio of General Insurance Policies

General Insurance
Number of Claims
Accident Development Years Outstanding Claims
Year 1 2 3 4 5 6 7 8 9 10 Chain-Ladder
1 40 124 157 93 141 22 14 10 3 2
2 37 186 130 239 61 26 23 6 6 0 0
3 35 158 243 153 48 26 14 5 1 0 1
4 41 155 218 100 67 17 6 4 1 0 5
5 30 187 166 120 55 13 12 4 1 0 17
6 33 121 204 87 37 16 13 7 1 0 37
7 32 115 146 103 42 17 20 8 2 0 89
8 43 111 83 80 46 27 23 8 1 1 186
9 17 92 142 88 72 31 25 7 3 4 372
10 22 101 156 138 83 34 20 13 11 7 563
Source: Verrall (1990).
14 NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 6, NUMBER 4

Figure 4 Figure 5
Distribution of the Number of Claims Distribution of the Reserves Using Information
from Verrall’s Data on Exposure from Verrall’s Data

of Table 7. Their estimate for the reserves was

only the aggregate claims data to estimate. It does
$22,301. The lower portion of Table 7 shows their
not even use the exposures for each accident
results, together with the predictive mean of the
year. Verrall (1990, 1991), in contrast, divides
reserves, also obtained using the information on
the exposures for each year of business into the
“closed” claims.
claims data before carrying out the analysis. This
The resulting distribution for the total reserves
is equivalent to the second specification given in
is shown in Figure 6. The mean is $25,076 and the
Section 5 (Equation 19, with Mit ⫽ Zit, the aggre-
standard deviation is $6,803. The resulting mean
gate claim amounts per cell). In addition, we have
is higher than Taylor and Ashe’s (1983) estimate.
a simulated sample from the complete and exact
This may be because we are considering the num-
predictive distribution. The results labeled as
ber of closed claims as random while they as-
Renshaw and Christofides are taken directly from
sumed them to be fixed and known. Again, they
the last reference. The results from Mack (1991)
only provide a point estimate while we have much
are somewhat smaller, but they were obtained
more information. Furthermore, their estimate is
assuming a gamma distribution for the incremen-
around the 38.9th quantile, so there is a very high
tal claims.
probability of larger losses and, hence, insuffi-
Now let’s apply the simulation process to the
cient reserves. Notice also that the distribution of
Verrall (1990) data using the information on fi-
the reserves is quite skewed.
nalized (closed) claims in each development year
Finally, the methods here are applied to the
for every accident year, as in Taylor and Ashe
(1983). The results are given in the lower portion
Figure 6
Table 7 Distribution of the Reserves Using Information
Comparison of Reserve Estimates and their on Closed Claims from Verrall’s Data
Standard Deviation Using Different Methods

Reserves Standard
Model (1,000) Deviation
Chain ladder $18,681 —
Verrall (no prior info) 19,512 $3,194
Verrall (prior info) 16,743 1,996
Mack 18,085 —
Renshaw/Christofides 19,512 3,121
Bayesian (global) 19,627 3,290
Bayesian (exposure) 19,630 3,284
Taylor and Ashe 22,301 —
Bayesian (claims) 25,075 6,803
BAYESIAN ESTIMATION OF OUTSTANDING CLAIM RESERVES 15

data in Ntzoufras and Dellaportas (2002) and Figure 7

compared with the results from their first two Distribution of the Reserves Using Information
models, plus the chain-ladder (see Table 8). The on Claims per Period from Ntzoufras/
upper portion of the table shows results from the Dellaportas’ Data
chain-ladder and our model under the third spec-
ification, that is, without using data on number of
claims (Xit ⫽ 1 in Equation B1); our estimate of
the reserves is larger. The middle portion shows a
comparison between their Model-1 and our sec-
ond specification, where only exposure data are
used (Xit ⫽ Xi and known, in Equation 19).
The results are very close, $2,911 and $2,884
respectively, but our standard deviation is
smaller. Something analogous occurs when all the
information—that is, number and amount of
claims paid per period (Ntzoufras and Dellapor-
tas’s Model-2 and our use of paid claims per pe-
riod)—is used. The lower portion of Table 8 obtain results that are comparable to frequentist
shows the corresponding reserves and their stan- ones. In fact, Bayesian analysis with objective
dard deviation. In Figure 7, we see that the dis- prior is being increasingly used to obtain “opti-
tribution of reserves is skewed, so that using a mal” frequentist procedures. One must be care-
normal approximation when computing the prob- ful, however, since the interpretation of results is
ability of extreme values may be misleading. An- very different in the two approaches to inference.
other difference in this case is that we also have The Bayesian method has the advantage of pro-
the posterior distribution for the total number of ducing the predictive distribution of the required
outstanding claims. reserves, whereas with other models, even though
you may have the variance, the complete distri-
bution is not available (Renshaw and Verrall 1994
7. FINAL REMARKS and England and Verrall 1999). This is also the
From the examples presented, it is clear that the case with Doray (1996), who computes UMVUEs
Bayesian procedure is more general than others for the mean and variance, but cannot display the
in that it provides more information: mean, stan- distribution. This is necessary to compute the
dard deviation, an approximation to the predic- probability of catastrophes, especially if use of the
tive distribution of outstanding claims and the normal approximation is not warranted. In some
reserves, and other characteristics such as prob- of the cases presented here, there is a clear skew-
abilities of extreme values. ness of the distribution, so that assumption of
In using noninformative (objective) priors, we normality is not adequate. In most, there is ample
evidence of non-normality and so confidence in-
tervals obtained with normal approximations can
Table 8 be very different from exact ones.
Comparison Between Three Bayesian The method also has the advantage of making
Specifications, the Chain-Ladder, and Two
use of all the information available, both on the
number and amounts of claims. Whatever infor-
Ntzoufras/Dellaportas Models
mation is available on exposure, closed claims,
Reserves Standard
and so forth can be incorporated into the model
Model (106) Deviation without major changes. In addition, with our
Chain Ladder $2,722 — Bayesian models, it is not necessary to obtain
Bayesian (global) 2,875 $569 explicitly any covariances between parameters,
Ntzoufras/Dellaportas 1 2,911 730
Bayesian (exposure) 2,884 571 because the simulation process incorporates
Ntzoufras/Dellaportas 2 2,303 718 them implicitly.
Bayesian (claims) 2,331 556
The method presented here also has the advan-
16 NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 6, NUMBER 4

tage that the number of claims is modeled as where p*s ⫽ ( p1 ⫹ p2 ⫹ p3 ⫹ . . . ⫹ ps). Now write
Poisson, whereas claim amounts are modeled us- f(p兩D) as
ing a lognormal distribution. In modeling severity
of claims in terms of average claim per develop- ⬀ f共 p k 兩D兲 f共 p k⫺1 兩p k , D兲
ment year, by year of origin, we are also implicitly ⫻ f共 pk⫺2兩pk⫺1, pk, D兲. . . f共 p2兩p3, p4. . .pk, D兲
allowing for the possibility that it is a function of
payment delay or for possible patterns in the ⫻ f共 p1兩p2. . .pk, D兲
average amount paid. For instance, it may happen
that there is an increase in average payment with with f( pk⫺i兩pk⫺i⫹1, pk⫺i⫹2 . . . pk, D) ⬀ ( pk⫺i/
delay (development) (Wright 1990). This would p*k⫺i)ck⫺i⫺1(1 ⫺ ( pk⫺i/p*k⫺i))dk⫺i⫺1, ck⫺i ⫽ ¥j⫽1
i⫹1

be reflected in the values of the ␤ parameters of xjk⫺i ⫹ 1 and dk⫺i ⫽ ¥t⫽1 ¥j⫽1 xjt ⫹ 1.
k⫺i⫺1 i⫹1

the model. Other specific behaviors may be mod- To do this, notice that

冉写 冊 冋 写 冉 冊 册
eled by including exogenous variables in the x2t
k k⫺1
model. xit pt
p t
f共p兩D兲 ⬀ t⫽1 t⫽1
p*k⫺1
APPENDIX A A1 A2

冋写 冉 冊 册 冋 写 冉 冊册
In the text, we integrate Equation (14) with re- x 3t x k⫺1,t
spect to ␪ to get k⫺2
pt 2
pt
⫻ p *k⫺2 ⫻...⫻ p *2
写
k t⫽1 t⫽1
n 1!
f共n 2 , . . . , n k , p兩D兲 ⬀ px1t A3 A k⫺1
兿t⫽1 x1t! t⫽1 t
k

(A1)
共n 2 ⫺ 1兲! x *2
x *2 ! using: p*k⫺1 ⫽ 1 ⫺ pk, p*k⫺2 ⫽ 1 ⫺ pk⫺1 ⫺ pk and,
⫻ 共1 ⫺ p *k⫺1 兲 x 2kp *k⫺1 in general, p*s ⫽ p1 ⫹ p2 ⫹ . . . ⫹ ps ⫽ 1 ⫺ ps⫹1
x 2 !共 x 2 ⫺ 1兲!
* 兿 t⫽1 x 2t !
k⫺1

⫺ . . . ⫺ pk. Now rewrite the term A1 in the ex-

写 冉 pp* 冊
k⫺1 x 2t
共n 3 ⫺ 1兲! pression above as follows:
t
⫻
k⫺1 共n 3 ⫺ x *3 兲!共 x *3 ⫺ 1兲!
写p
t⫽1 k⫺1

A 1 ⫽ p kx 1k x 1t

写 冉 pp* 冊
t
k⫺2 x 3t
* x *3 ! t
t⫽1
⫻ 共1 ⫺ p *k⫺2 兲 n 3⫺x *3 p *k⫺2
x3
兿 t⫽1 x 3t !
k⫺2

写 冉 pp* 冊
t⫽1 k⫺2

写 共 p*
k⫺1 x 1t k⫺1
t
⫽ p kx 1k k⫺1 兲 x 1t
· k⫺1
· t⫽1 t⫽1
·
共n k ⫺ 1兲!
⫻
写 冉 pp* 冊
k⫺1 x 1t
共n k ⫺ x k1 兲!共 x *k ⫺ 1兲! ⫽ p 共1 ⫺ p k 兲
x 1k
k
k⫺1
¥ t⫽1 x 1t t

t⫽1 k⫺1

写 冉 pp*冊
1 x k1
x *k ! 1
⫻ 共1 ⫺ p *1 兲 n k⫺x k1p *1 x k
写冉 冊
. k⫺1 x 1t
x k1 ! 1 k⫺1 pt
t⫽1
⫽ p 共1 ⫺ p k 兲
x 1k
k
¥ t⫽1 x 1t
.
t⫽1
p *k⫺1
It is now possible to integrate (sum) out the n’s,
and we are left with
Therefore, now

冋 册冋 冉 冊 册冋 冉 冊 册
x2t x3t

写p 写 写
k k⫺1 k⫺2
pt pt k⫺1

f共p兩D兲 ⬀ x1t A 1 ⫻ A 2 ⫽ 关 p kx 1k共1 ⫺ p k 兲 ¥ t⫽1 x 1t

兴
t
t⫽1 t⫽1
p*k⫺1 t⫽1
p*k⫺2

冋 写 冉 冊 册 冋写 冉 冊 册 冋写 冉 冊 册
B1
x4t xkt x 1t ⫹x 2t
k⫺3 2 k⫺1
pt pt pt
⫻ ... , ⫻ (A2)
t⫽1
p*k⫺3 t⫽1
p*2 t⫽1
p *k⫺1
BAYESIAN ESTIMATION OF OUTSTANDING CLAIM RESERVES 17

冉 冊 冉 冊 and so on, until we have A1 ⫻ A2 ⫻ . . . ⫻ Ak⫺1 ⫽

k⫺2
x 1k⫺1 ⫹x 2k⫺1 ¥ t⫽1 共 x 1t ⫹x 2t 兲
p k⫺1 p k⫺1
⫽ B1 1⫺ B1 ⫻ B2 ⫻ . . . ⫻ Bk⫺1 ⫻ Bk, with
p *k⫺1 p *k⫺1

⫻ 写 冉 冊 冉
k⫺2

⫻
pt
p*k⫺1
x1t⫹x2t
1⫺
pk⫺1
p*k⫺1 冊
k⫺2
⫺¥t⫽1 共 x1t⫹x2t兲
Bi ⫽ 冉 p k⫺i⫹1
p *k⫺i⫹1 冊 x 1k⫺i⫹1 ⫹x 2k⫺i⫹1 ⫹. . .⫹x ik⫺i⫹1

冉 冊
t⫽1 k⫺i
共 x1t⫹x2t⫹. . .⫹xit兲

冉 冊 冉 冊
¥t⫽1
x 1k⫺1 ⫹x 2k⫺1 k⫺2
共 x 1t ⫹x 2t 兲
pk⫺i⫹1
p k⫺1 p k⫺1 ¥ t⫽1
⫻ 1⫺ , (A5)
⫽ B1 1⫺ p*k⫺i⫹1
p *k⫺1 p *k⫺1
B2
where Bk ⫽ 1 and Bi ⬀ f( pk⫺i⫹1兩pk⫺i⫹2,
pk⫺i⫹3, . . . , pk, D), as required. Recall that x*i ⫽
xit, p*ai ⫽ p1 ⫹ . . . ⫹ pai, p*k⫺1 ⫽ 1 ⫺ pk. Let
写冉 冊 冉 冊
ai
x 1t ⫹x 2t k⫺2
⫺¥ t⫽1 共 x t ⫹x 2t 兲 ¥t⫽1
k⫺2
pt p *k⫺1 ⫺ p k⫺1 p*k⫺2 ⫽ 1 ⫺ pk⫺1 ⫺ pk ⫽ 1 ⫺ p̃k⫺1, with p̃k⫺1 ⫽
⫻
t⫽1
p *k⫺1 p *k⫺1 pk⫺1 ⫹ pk and p̃i ⫽ pi ⫹ pi⫹1 ⫹ . . . ⫹ pk ⫽ 1 ⫺
p*i⫺1. Using these definitions, it follows immedi-

写 冉 pp* 冊
k⫺2
t
x 1t ⫹x 2t
k⫺2
ately that
⫽ B1 ⫻ B2 ⫻ 共 p *k⫺2 兲 ⫺¥ t⫽1 共 x it ⫹x 2t 兲

t⫽1 k⫺1 x ⫹1
1k
E共 pk兩D兲 ⫽ ⫽ p̂k,
k⫺2
¥ x ⫹2
k
t⫽1 1t
共 x 1t ⫹x 2t 兲
⫻ 共 p *k⫺1 兲 ¥ t⫽1
冉 冊
, (A3)
x1k⫺1 ⫹ x2k⫺1 ⫹ 1
since p*k⫺1 ⫺ pk⫺1 ⫽ 1 ⫺ pk ⫺ pk⫺1 ⫽ p*k⫺2. E共 pk⫺1兩pk, D兲 ⫽ 共1 ⫺ pk兲 k⫺1
¥t⫽1 共x1t ⫹ x2t兲 ⫹ 2
Furthermore,
⫽ 共1 ⫺ p k 兲 p̂ k⫺1 ,

写冉 冊
k⫺2 x 1t ⫹x 2t
pt
A1 ⫻ A2 ⫽ B1 ⫻ B2 ⫻ . E共 pk⫺2兩pk⫺1, pk, D兲 ⫽ 共1 ⫺ pk兲共1 ⫺ pk⫺1兲
p *k⫺2

冉 冊
t⫽1 3
¥i⫽1 xik⫺2 ⫹ 1
Proceed in a similar fashion to find ⫻ k⫺2 3
¥t⫽1 ¥i⫽1 xit ⫹ 2

写 冉 pp* 冊
k⫺2 x 1t ⫹x 2t
t ⫽ 共1 ⫺ pk兲共1 ⫺ pk⫺1兲p̂k⫺2,
A1 ⫻ A2 ⫻ A3 ⫽ B1 ⫻ B2 ⫻
k⫺2
t⫽1 and so on. The general unconditional expected

冋写 冉 冊 册
x 3t
value is then given by
k⫺2
pt
⫻ E共 pk⫺i兩D兲
t⫽1
p *k⫺2
⫽ 共1 ⫺ p̂k兲共1 ⫺ p̂k⫺1兲 䡠 . . . 䡠 共1 ⫺ p̂k⫺i⫹1兲p̂k⫺i.
⫽ B1 ⫻ B2 The conditional posterior pdf ’s given in Equation

冉 冊 冉 冊
x 1k⫺2 ⫹x 2k⫺2 ⫹x 3k⫺2 k⫺3
¥t⫽1 共 x1t⫹x2t⫹x3t兲 (A5) are essentially the same as the “predictive”
p k⫺2 pk⫺2
⫻ 1⫺ distributions given in Verrall (2000, Eq. 3.11).
p *k⫺2 p*k⫺2 The unconditional mean given in Equation (A6)
B3 below is also very similar to his results in Equa-
tion (4.1). In our case, the variances are much
冉 冊 写冉 冊
k⫺3
⫺¥t⫽1 共 x1t⫹x2t⫹x3t兲 k⫺3 x1t⫹x2t⫹x3t
pk⫺2 pt more complicated to compute because of the suc-
⫻ 1⫺
p*k⫺2 t⫽1
p*k⫺2 cessive conditioning that has to be carried out,
and to the fact that more detail is included. To
(A4) find the mean of outstanding claims (R) in acci-
so that dent year 2, we have:

A1 ⫻ A2 ⫻ A3 ⫽ B1 ⫻ B2 ⫻ B3 E共R2兩D兲 ⫽ E共 pkN2兩D兲 ⫽ E关E共N2兩pk兲 pk兴

pk

写 冉 pp* 冊 冋 册
k⫺3 x 1t ⫹x 2t ⫹x 3t
t x*2k⫺1
⫻ , ⫽ E pk ,
t⫽1 k⫺3
pk p*k⫺1
18 NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 6, NUMBER 4

and, from Raiffa and Schlaifer (1961), APPENDIX B

B1. SIMULATION STEPS FOR THE

E共R2兩D兲 ⫽ E pk
pk
冋
¥ x2t
1 ⫺ pk
Dt ⫽
k⫺1
t⫽1

t⫽1
x2tE
pk
冏册
pk 1 ⫺ pk
Dt 冘
k⫺1

冋 冏册 NUMBER OF CLAIMS
1. Use expression B1 (Equation A2 in Appendix
A) to generate pk(j) from a

⫽ 冘x
k⫺1

t⫽1
2t 冉 x 1k ⫹ 1
k⫺1
¥ t⫽1 x 1t
⬵ 冊 冉
x 1k
k⫺1
¥ t⫽1
k⫺1
x 2t
¥ t⫽1 x 1t
. 冊 冉
Beta x1k ⫹ 1, 冘x
k⫺1

t⫽1
1t ⫹1 冊
2. Use expression B2 (Equation A3 in Appendix
Similarly, A) to generate ␪˜ k⫺1
(j)
from

E共R3兩D兲 ⫽ E共 p̃k⫺1N3兩D兲

⫽ E E共N3兩p兲共 pk ⫹ pk⫺1兲
冉
Beta x1k⫺1 ⫹ x2k⫺1 ⫹ 1, 冘 共x
k⫺2

t⫽1
1t ⫹ x2t兲 ⫹ 1 冊
pk⫺1
3. Use the results of steps 1 and 2 to generate

冋 冉 冘 冊冏 册 ⫽ ␪˜ k⫺1 (1 ⫺ pk(j))
(j) (j)
k⫺2 pk⫺1
x3t 4. Use expression B3 (Equation A4 in Appendix
⫻ Ep 共 pk ⫹ pk⫺1兲 D
p*k⫺2 A) to generate ␪˜ k⫺2
(j)
from

冉
t⫽1

⫽ Ep 冋
共1 ⫺ p*k⫺2兲 k⫺2
p*k⫺2
x3t D 冘 冏册 (A6)
Beta x1k⫺2 ⫹ x2k⫺2 ⫹ x3k⫺2 ⫹ 1,

冊
t⫽1

冘 共x
k⫺3

1t ⫹ x2t ⫹ x3t兲 ⫹ 1
and approximating t⫽1

5. Use the results of step 1– 4 above to generate

⫽ ␪˜ k⫺2 ⫺ pk(j)), and so on. With

冤 冥
(j) (j) (j)
pk⫺2 (1⫺pk⫺1

冉 冊
k⫺1 this, we will have generated a vector p(j) ⫽
¥t⫽1 x2t
E共R3兩D兲 ⬵ 共 x1k ⫹ x1k⫺1兲 ⫹ x1k k⫺1 ⫹ x2k⫺1 (pk(j), pk⫺1
(j)
, . . . , p1(j)).
¥t⫽1 x1t (j)
6. Use p and
E共R2兩D兲 ⫽ x2k
共ni ⫺ 1兲!
⫻ 冉 k⫺2
¥ t⫽1 x 3t
¥ t⫽1 共 x 1t ⫹ x 2t 兲
k⫺2 . 冊 f共n i 兩p, D兲 ⫽
共ni ⫺ x*i 兲!共x*i ⫺ 1兲!
⫻ 共1 ⫺ p*k⫺i⫹1兲ni⫺x*ip*k⫺i⫹1
xi
,
*

In general,
i ⫽ 2, . . . , k, to generate an observation for
E共Ri兩D兲 each ni. Thus, we generate (n2, n3, . . . , nk) ⫽
n( j).

⬵ 冋 冒冉 冘
冘
k⫺共i⫺1兲

t⫽1
xit
k⫺共i⫺1兲

t⫽1
共x1t ⫹ x2t ⫹ · · · ⫹ x共i⫺1兲t兲 冊 7. Use n(j), p(j) to generate observations for the
unknown portion of xi(j) from each of (k ⫺ 1)
multinomials (one for each year); that is, use
the predictive distribution

⫻ 冋 冘 冘
i⫺1 k

j⫽1 t⫽k⫺共i⫺2兲
册
x jt ,
共 j兲
f共 x i1 共 j兲
, x i2 共 j兲
, . . . , x ik 兩n i共 j兲 , p 共 j兲兲
⫽ Multk共ni共 j兲; p共 j兲兲,
where xjt ⫽ E(Rj兩D) whenever j ⫹ t ⬎ k ⫹ 1. i ⫽ 2, . . . , k.
BAYESIAN ESTIMATION OF OUTSTANDING CLAIM RESERVES 19

B2. SIMULATION STEPS FOR CLAIM BERNARDO, J.M., AND A.F.M. SMITH. 1994. Bayesian Theory. New
York: John Wiley & Sons.
AMOUNTS BROWN, R.L. 1993. Introduction to Ratemaking and Loss Reserv-
8. Generate an observation ␴( j) from ing for Property and Casualty Insurance. Winsted, CT:
ACTEX Publications.
f共␴兩D兲 ⬀ ␴ ⫺共T U⫺2k⫹2兲 BÜHLMANN, H. 1967. “Experience Rating and Probability,” ASTIN

冋 册
Bulletin 4, 199 –207.
1 CROW, E.L., AND K. SHIMIZU. 1988. Lognormal Distributions. The-
⫻ exp ⫺ 共y ⫺ W␪ˆ 兲⬘共y ⫺ W␪ˆ 兲 3 ␴共 j兲;
2␴2 ory and Applications, New York: Marcel Dekker.
DE ALBA, E. 1988. “Pronostico Bayesiano de Agregados en Procesos
a “square-root inverted-gamma” distribution Estacionales Estables,” Revista de Estadistica 2(4): 1– 6.
with parameters ␣ ⫽ (TU ⫺ 2k ⫹ 1)/ 2 and DE ALBA, E., AND M. MENDOZA. 1996. “Discrete Bayesian Models for

␤ ⫽ (y ⫺ W␪ˆ )⬘(y ⫺ W␪ˆ ),/2. This can be done Forecasting with Stable Seasonal Patterns,” Advances in
by first generating an observation g( j) from a Econometrics, 2 (Part B): 267– 81.
gamma(␣, ␤) and then making ␴( j) ⫽ DORAY, L.G. 1996. “UMVUE of the IBNR Reserve in a Lognormal
1/公g( j). Linear Regression Model,” Insurance: Mathematics and
Economics 18(1): 43–57.
9. Generate an observation ␪(j) ⫽ (␮(j), ␣2(j), . . . ,
ENGLAND, P., AND R. VERRALL. 1999 “Analytic and Bootstrap Esti-
␣k(j), ␤2(j), . . . , ␤k(j))⬘ from mates of Prediction Errors in Claims Reserving,” Insur-
ance: Mathematics and Economics, 25: 281–93.
N共␪ˆ , ␴共 j兲2共W⬘W兲⫺1兲 3 ␪共 j兲. ———. 2002. “Stochastic Claims Reserving in General Insurance,
This is accomplished by using the Choleski Institute of Actuaries and Faculty of Actuaries, 1–76.” The
Faculty and Institute of Actuaries, Sessional Meeting Paper
decomposition.
sm0201. Online at http://www.actuaries.org.uk/sessional/
10. Generate an observation from the predictive sm0201-report.html.
distribution N(␮it (j)
, ␴(j)) 3 Yit
(j)
, with ␮it (j)
⫽ FERGUSON, T.S. 1973. A Bayesian Analysis of Some Nonparamet-
␮ ⫹ ␣i ⫹ ␤t for each (i, t), where t ⬎ k ⫺
(j) (j) (j)
ric Problems, The Annals of Statistics 1(2): 209 –30.
i ⫹ 1, and compute Mit (j)
⫽ exp{Yit (j)
}. This part GOOVAERTS, M.J., R. KAAS, A.E. VAN HEERWAARDEN, AND T. BAU-
of the process (steps 8–10) provides a set of WELINCKX. 1990. Effective Actuarial Methods. Amsterdam:

randomly generated values for the average North-Holland.

claim in each cell of the (unobserved) lower HAASTRUP, S., AND E. ARJAS. 1996. “Claims Reserving in Continu-
right triangle. These values include parameter ous Time; A Nonparametric Bayesian Approach,” ASTIN
Bulletin 26(2): 139 – 64.
variability (through the random values ␪(j))
HESSELAGER, O., AND T. WITTING. 1988. “A Credibility Model with
and process variability (through ␴(j)). In this Random Fluctuations in Delay Probabilities for the Predic-
(j)
sense, the Mit are values from the predictive tion of IBNR Claims,” ASTIN Bulletin 18(1): 79 –90.
lognormal distribution. HOSSACK, I.B., J.H. POLLARD, AND B. ZENWIRTH. 1999. Introductory
11. Compute Statistics with Applications in General Insurance. 2nd ed.
Cambridge: University Press.
Z it共 j兲 ⫽ X it共 j兲 ⴱ M it共 j兲 (B1) JEWELL, W.S. 1989. “Predicting IBNYR Events and Delays I. Con-
tinuous Time,” ASTIN Bulletin 19(1): 25–56.
for each i ⫽ 2, . . . , k, t ⬎ k ⫺ i ⫹ 1. JEWELL, W.S. 1990. “Predicting IBNYR Events and Delays II.
12. To obtain a value of the total required re- DiscreteTime,” ASTIN Bulletin 20(1): 93–111.
serves, add all the Zit( j)
, that is, R( j) ⫽ ¥i,t Zit
( j)
. KLUGMAN, S.A. 1992. Bayesian Statistics in Actuarial Science.
Boston: Kluwer.
KREMER, E. 1982. “IBNR Claims and the Two-way Model of
ACKNOWLEDGMENTS ANOVA,” Scandinavian Actuarial Journal 1: 47–55.
The author gratefully acknowledges support from MACK, T. 1994. “Which Stochastic Model is Underlying the Chain
Ladder Method?” Insurance: Mathematics and Economics
Asociación Mexicana de Cultura, A.C., as well as
15: 133– 8.
the contribution of Miguel A. Juárez in the initial MAKOV, U.E. 2001. “Principal applications of Bayesian Methods
research stages. in Actuarial Science: A Perspective,” North American Ac-
tuarial Journal 5(4): 53–73.
MAKOV, U.E., A.F.M. SMITH, AND Y.-H. LIU. 1996. “Bayesian Meth-
REFERENCES ods in Actuarial Science,” The Statistician 45(4): 503–15.
BERGER, J.O. 1985. Statistical Decision Theory and Bayesian NORBERG, R. 1986. “A Contribution to Modelling of IBNR Claims,”
Analysis. 2nd ed. New York: Springer-Verlag. Scandinavian Actuarial Journal, (3– 4): 155–203.
20 NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 6, NUMBER 4

———. 1993. “Prediction of Outstanding Liabilities in Non-Life TAYLOR, G.C., AND F.R. ASHE. 1983. “Second Moments of Esti-
Insurance,” ASTIN Bulletin 23(1): 96 –115. mates of Outstanding Claims,” Journal of Econometrics
NTZOUFRAS, I., AND P. DELLAPORTAS. 2002. “Bayesian Modelling of 23(1): 37– 61.
Outstanding Liabilities Incorporating Claim Count Uncer- VERRALL, R. 1990. “Bayes and Empirical Bayes Estimation for the
tainty,” North American Actuarial Journal. Forthcoming. Chain Ladder Model,” ASTIN Bulletin 20(2): 217– 43.
O’HAGAN, A. 1994. Kendall’s Advanced Theory of Statistics, Vol. ———. 1991. “On the Estimation of Reserves from Loglinear
2B, Bayesian Statistics. London: Edward Arnold. Models, Insurance: Mathematics and Economics 10(1):
OLIVER, R.M. 1987. “Bayesian Forecasting with Stable Seasonal 75– 80.
Patterns,” Journal of Business and Economic Statistics ———. 2000. “An Investigation into Stochastic Claims Reserv-
5(1): 77– 86.
ing Models and the Chain-Ladder Technique,” Insurance:
RAIFFA, H., AND R. SCHLAIFER. 1961. Applied Statistical Decision
Mathematics and Economics, 26(1): 91–9.
Theory, Boston: Harvard University Press.
WRIGHT, T.S. 1990. “A Stochastic Method for Claims Reserving in
RENSHAW, A.E., AND R. VERRALL. 1994. “A Stochastic Model Under-
General Insurance,” Journal of the Institute of Actuaries
lying the Chain-Ladder Technique,” Proceedings of the
117: 677–722.
XXV ASTIN Colloquium, Cannes.
ZELLNER, A. 1971. An Introduction to Bayesian Inference in
SCOLLNIK, D.P.M. 2001. “Actuarial Modeling With MCMC and
BUGS,” North American Actuarial Journal 5(2): 96 –125. Econometrics. New York: Wiley.
TANNER, M.A. 1996. Tools for Statistical Inference. 3rd. ed. New
York: Springer-Verlag.
Discussions on this paper can be submitted until
TAYLOR, G.C. 1986. Claim Reserving in Non-Life Insurance. New April 1, 2003. The author reserves the right to reply to
York: Elsevier Science Publishers. any discussion. Please see the Submission Guidelines
———. 2000. Claim Reserving. An Actuarial Perspective. New for Authors on the inside back cover for instructions
York: Elsevier Science Publishers. on the submission of discussions.