See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/259432335

Prediction of RBNS and IBNR claims using claim amounts and claim counts

Article  in  ASTIN Bulletin · November 2010

DOI: 10.2143/AST.40.2.2061139

CITATIONS READS

24 767

3 authors, including:

Richard Verrall Jens Perch Nielsen

City, University of London City, University of London
61 PUBLICATIONS   1,345 CITATIONS    149 PUBLICATIONS   2,685 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

generalized additive models View project

Operational risk View project

All content following this page was uploaded by Richard Verrall on 03 September 2014.

The user has requested enhancement of the downloaded file.

 Prediction of RBNS and IBNR claims using claim
amounts and claim counts

Richard Verrall, Jens Perch Nielsen and Anders Hedegaard Jessen

April 9, 2010

Abstract
A model is proposed using the run-off triangle of paid claims and also the
numbers of reported claims (in a similar triangular array). These data are
usually available, and allow the model proposed to be implemented in a large
variety of situations. On the basis of these data, the stochastic model is built
from detailed assumptions for individual claims, but then approximated using
a compound Poisson framework. The model explicitly takes into account the
delay from when a claim is incurred and to when it is reported (the IBNR delay)
and the delay from when a claim is reported and to when it is fully paid (the
RBNS delay). These two separate sources of delay are estimated separately,
unlike most other reserving methods. The results are compared with those of
the chain ladder technique.

1
1 Introduction

There are a number of stochastic models that can be used to estimate reserves
in non-life insurance mathematics; see Schmidt (2007) for an extensive litera-
ture list. Wüthrich and Merz (2008), England and Verrall (2002) and Taylor
(1986) provide useful overviews of stochastic claims reserving. Most of these
models have been designed to deal with data which have been aggregated in
some way, as this often makes the presentation of the data convenient. How-
ever, this aggregation of the data may lead to a loss of information that in some
cases can give relatively poor estimation and prediction of the outstanding li-
abilities. This has been the subject of some recent papers on reserving: for
example, Taylor and McGuire (2004) uses a generalized linear model framework
to model the characteristics of individual claims. Norberg (1993, 1999) sets out
a framework for the claims occurrence, reporting and payment process, at an
individual claims level. Also relevant in this context is Norberg (1986). Models
based on individual claims data tend to be very detailed, often rather complex
and use extensive data to estimate parameters. For the practising actuary how-
ever, they have certain limitations: in particular, they are difficult to implement
because the use of data at an individual level is particularly computationally
challenging. Furthermore, very large and elaborate data sets are often hard to
get in insurance companies, and it is often the case that a model will only get
used in a practical situation if it can be applied to a wide variety of data sets
across a wide variety of business lines.
It can be seen from this that there is a difficult choice to be made of whether
to use individual data, which is theoretically appealing but computationally dif-
ficult, or whether to use aggregate data, which is much easier to deal with but
from which some (possibly important) information has been lost.
In this paper, we propose a stochastic model for loss reserving which is
based on incremental reported claim numbers Nij and paid amounts Xij and
which serves to predict RBNS and IBNR claims separately. We take a similar
approach as Ntzoufras and Dellaportas (2002) and Wright (1990), in that we
build a model for aggregate paid claims from basic principles at the level of
individual data. We believe that the use of the aggregated counts data, which
is readily available in most actuarial offices, can improve the reserving accuracy.
Other interesting possibilities of adding extra data compared to the simple chain
ladder method are Liu and Verrall (2009a, 2009b), Ntzoufras and Dellaportas
(2002), Schnieper (1991), Verrall (1990) and Verrall and England (2005).
Including the data on incurred counts enables us to model the payment pat-
terns for RBNS claims. In this way, we separate the reporting delay (in the in-
curred counts) from the payment delay (in the paid claims). In contrast, widely
used methods such as the chain ladder technique simply include all sources of
delay in a single development pattern.
The chain ladder technique was originally introduced without a stochastic
model specified using heuristic reasoning to estimate the sum of Incurred But
Not Reported (IBNR) claims and RBNS claims. In Hachemeister and Stanard

2
(1975), Mack (1991), Neuhaus (2004) and Renshaw and Verrall (1994) stochas-
tic models have since been formulated that lead to the same estimates as the
chain ladder method. In all these papers, the models take the data as given and
do not attempt to build a model based on the commonly accepted compound
Poisson framework, used elsewhere in risk theory. It could be argued that the
over-dispersed Poisson model could be interpreted in this way (see, for example,
England and Verrall, 1999), but this was not the original approach taken. In this
paper we derive model which is an approximation to an exact model based on
more detailed data, and which is a compound Poisson model. Separate models
are defined for IBNR and RBNS claims, allowing for the prediction of IBNR and
RBNS claims separately. In this way, we take a similar approach to Bühlmann
et al (for example), who also split the reserve into two elements.
The paper is set out as follows. In Section 2 we define the notation and de-
scribe the data which we will assume is available. In Section 3, the theoretical
development is given, working from assumptions at the level of individual data,
which shows how the claims development is split into the IBNR and RBNS
delays. In Section 4, the model which we will actually apply is given, as an ap-
proximation to the more detailed models for individual data. Section 5 considers
prediction and Section 6 examines results based on the model. The conclusion
is contained in Section 7.

2 Data and notation

In deciding which data to use, the two main considerations are that this data
should be readily available for most practical actuaries and that this extra data
should have the potential to provide improved estimates of the RBNS and IBNR
reserves. In general, we assume that a run-off triangle consists of the random
variables △m = {Xij : (i, j) ∈ Am } where Am = {(i, j) ∈ N × N0 : 1 ≤ i + j ≤
m}.

X10 X11 ... X1,m−2 X1,m−1

X20 X21 ... X2,m−2
.. ..
. .
Xm0

The random variables in △m could represent either the paid or the incurred
data. For the paid data, Xij , (i, j) ∈ Am , are the total claims incurred in period
i which are paid with j periods delay from when they were incurred. For the in-
curred data, Xij are total claims incurred in period i with delay j . In this case,
the claims which are reported but not paid a claim estimate is included, rather
than the actual payment. In a practical context, a decision has to be made as
to whether to use paid or incurred data (or to try to use both). The advantage
of using only paid data is that △m then contains no human judgement: it is
"real data". However, it is possible that the case estimates, which are included

3
in the incurred data, cantain useful information about future payments. On
the other hand, the inclusion of claim estimates is debatable since these are not
"real data", and there may be political or business related considerations which
make the individual claim estimates unreliable. There is further some variabil-
ity that is disregarded, since the claim estimates and the actual paid amount
often differ. Finally, claims estimates appear as paid at the wrong point in time
which can disrupt the cashflow modeling.
For these reasons the approach taken in this paper is to use the triangle
of paid claims, which is easily accessible in most companies. Thus, Xij is the
total claims incurred in period i and paid with j periods delay. We combine
this paid triangle with a second triangle, in the same format as the paid trian-
gle above, containing the number of incurred claims. Note that these data are
obtained from the incurred claims, and therefore use some of the information
not used when just the aggregate paid claims are used. The random variables
are denoted by ℵm = {Nij : (i, j) ∈ Am } where Nij represents the number of
claims incurred in period i and reported with j periods delay (in period i + j)
for (i, j) ∈ Am . It would also have been possible to consider the number of
payments. This would remove one difficulty with the incurred claims (as dis-
cussed below), in that some reported claims end up without a payment being
made - the zero-claims. However, the use of the number of payments can lead
to a number of other difficulties. For example, the number of payments is rarely
easily accessible in insurance companies. The number of reported claims on the
other hand is usually relatively easy to obtain.
We therefore assume that two triangles of data are available: the triangle
of aggregate paid claims and the triangle of the number of incurred claims. As
mentioned in the introduction, restricting the analysis to just these triangles to
some extent complicates the statistical analysis, and it would be better, from
this point of view at least, to assume that data was available at whatever level
of detail was required. A disadvantage of this would be that the estimation of
the models would become much more computationally intense, and the models
could not be used when the data requirements were not satisfied.
Thus, a compromise about the data has been made, but it will be seen in
Section 5 that by just including the count data for the incurred claims, it is pos-
sible to improve significantly on the chain ladder technique without completely
giving up the well known chain ladder idea. The chain ladder technique is, in
fact, a special case of the model specified in this paper.
In the next section we define the model for △m and ℵm using some un-
observable random variables. The structure which are intended to mimic the
models from Norberg (1993, 1999), but using a discrete time framework.

3 Modelling IBNR and RBNS claims

In this section a micro model is introduced, using a number of (in practise

often) unobservable random variables. Based on this micro model, a compound

4
poisson interpretation of IBNR and RBNS claims (at the individual level) is
derived. The aim of specifying a model for the individual claims is to derive a
suitable model for the aggregated data which is assumed to be available.
Consider the kth claim of the Nij claims incurred in period i and reported
with j periods delay. Usually a claim is not paid immediately upon notification
to the insurance company. The final claim amount is generally paid with some
waiting time from notification, often due to general consideration of the case,
legal issues, collection of further information concerning the case, etc. In other
words, there is a delay from a claim being reported until it is fully paid. The
claims that have been reported but are not yet paid are the so called RBNS
claims (or Incurred But Not Enough Reported (IBNER) claims). The related
delay in payment is referred to as the RBNS delay.
paid
Denote by Nijk the part of the Nij claims which are (fully) paid with k
periods delay (after being reported), k = 0, ..., d. Here k = 0 corresponds to a
claim being paid in the same period as it was reported whereas k = d is the
maximal possible RBNS delay in the model. d could be chosen using information
from the underlying data or the judgement from a claims handler.
The aggregate paid claims will depend on the number of claims paid in each
development period, rather than the number of reported claims. The number
of claims incurred in period i and (fully) paid with j periods delay after being
paid
reported is denoted by Nij , where
min{j,d}
paid paid paid paid
X paid
Nij = Nij0 + Ni,j−1,1 + · · · + Ni,j−min{j,d},min{j,d} = Ni,j−k,k
k=0

for (i, j) ∈ Am .
Thus, the lifetime of a claim is divided into two: the IBNR delay and the
RBNS (and IBNER) delay. These two sources of delay are modelled separately.
The IBNR delay is considered when a model is specified for the reported numbers
of claimssince the outstanding numbers for this triangle are the claims still to
be reported. For the reported claims, the RBNS delay can be considered by
specifying a model for the number of claims paid, given the numbers of reported
paid
claims. In other words, we specify a model for Nijk | Nij .
To begin with the the numbers of reported claims, it is assumed that the
number of claims incurred in period i and reported with j periods delay, Nij ,
are independently distributed and have an over-dispersed Poisson distribution.
It is well-known that for this model simplest way to obtain the predictions of
future incurred claims is to use the chain ladder technique (see, for example,
Hachemeister and Stanard, 1975 and Renshaw and Verrall, 1994, for proofs of
this well known result). Thus, a straightforward way to obtain forecasts of the
numbers of claims that will be reported for future delay periods is simply to
apply the chain ladder technique to the triangle of the numbers of reported
claims. In this way, the IBNR delay can be estimated.
For the RBNS delay, it is necessary to consider the delay in a reported
claim being paid, and also consider the issue of a claim being paid in partial
payments. For the estimation of the outstanding claims, the distribution of

5
the claim severities is also required. In this paper, it is assumed that claims
are settled with a single payment, which simplifies the theory, estimation and
data questions considerably. Although it may often be the case that there
is often more than one payment per claim, including this in the model leads
to a much more complicated approach. Also, we believe that the simplified
approach used in this paper should provide some useful and realistic insights
into the different sources of the claims delay patterns. Finally, data are often
not available on the development of the payment patterns and definitions on
payments may differ from one insurance company to the other, or even within
the same insurance company. With this assumption,we can now specify the
paid
distribution of Nijk | Nij . Given Nij we assume that the distribution of the
numbers of paid claims follows a multinomial distribution. i.e.
paid paid
(Nij0 , ..., Nijd ) ∼ Multi(Nij ; p0 , ..., pd )

for (i, j) ∈ Am and p0 + · · · + pd = 1 where pi ∈ (0, 1), 0 ≤ i ≤ d.

The aggregated incremental claims can be obtained by summing the indi-
vidual payments:
paid
Nij
(k)
X
Xij = Yij (1)
k=1

(k)
for (i, j) ∈ Am . Here Yij , (i, j) ∈ Am , k ≥ 1, denotes an individual claim pay-
ment. The final part of the model is to specify the distribution of the individual
(k)
claims, Yij . It is assumed that these are independent of the numbers of claims,
and do not depend on the IBNR delay or the RBNS delay. It is also assumed
that they are independently and identically distributed. We recognise that the
assumption that the claim payments are identically distributed is unlikely to be
valid in practice: in particular, the sizes of the payments are likely to depend
on the delays. However, the model with this basic assumption provides a rea-
sonable starting point and ways in which this can be relaxed can be explored in
further developments of the approach.
With these distributional assumptions, the likelihood function can be written
as

Lℵm ,△m = Lℵm × L△m |ℵm

 
m m−i
Y Y
=  P (Nij = nij )
i=1 j=0
m
!
Y
× fXi0 ,...,Xi,m−i |Ni0 ,...,Ni,m−i (xi0 , ..., xi,m−i |ni0 , ..., ni,m−i ) .
i=1

Since Lℵm and L△m |ℵm are not functions of the same parameters, it is suffi-
cient to maximize Lℵm and L△m |ℵm separately to maximize L△m ,ℵm . As stated
above, the likelihood function of ℵm can be maximised using the chain ladder

6
method.
Having defined the general framework, the next section formulates an ap-
proximation to this the model and discusses its possibilities and limitations.

4 Approximating the Likelihood with an over-

dispersed Poisson Distribution

As was stated in the introduction, the process we have followed is to derive a

model as far as possible, based on very basic unobservable random variables, and
then approximate the model as closely as possible to motivate a model for the
data available. The previous section has looked at the process that generates the
claims, and we now look at the resulting model for the aggregate data and derive
an approximation which is easier to use in practice. The approximation to the
log-likelihood function for △m given ℵm is based on the theory of generalised
linear models (see, for example, McCullagh and Nelder, 1989). The approach
taken is to construct a quasi-log likelihood which (for generalised linear models)
requires just the first two moments, E[Xij |ℵm ] and V [Xij |ℵm ].
paid
E[Xij |ℵm ] = E[E[Xij |Nij ]|ℵm ]
  paid  
Nij
  X (k) paid 
= E E  Yij |Nij  |ℵm 

k=1

paid (k)
= E[Nij E[Yij ]|ℵm ]
paid (k)
= E[Nij |ℵm ]E[Yij ]

paid paid
V [Xij |ℵm ] = E[V [Xij |Nij ]|ℵm ] + V [E[Xij |Nij ]|ℵm ]
paid
Nij
(k) (k)
X paid paid
= E[V [ Yij |Nij ]|ℵm ] + V [Nij E[Yij ]|ℵm ]
k=1
paid (k) paid (k)
= E[Nij V [Yij ]|ℵm ] + V [Nij E[Yij ]|ℵm ]

Since the claim severities are assumed to be independently, identically dis-

(k) (k)
tributed, we can write E[Yij ] = µ and V [Yij ] = σ 2 . Hence
paid paid
V [Xij |ℵm ] = E[Nij |ℵm ]σ 2 + V [Nij |ℵm ]µ2 (2)
paid
It can be seen that the distribution of Nij |ℵm plays a crucial role in the first
two moments of Xij |ℵm . This distribution can be derived by considering sums

7
of the numbers of claims paid with the appropriate length of delay, since (as
paid Pmin{j,d} paid
shown in section 3) Nij = k=0 Ni,j−k,k . Considering first the mean,
 
min{j,d}
paid
X paid
E[Nij |ℵm ] = E  Ni,j−k,k |ℵm 
k=0
min{j,d}
X paid
= E[Ni,j−k,k |ℵm ] (3)
k=0
min{j,d}
X
= Ni,j−k pk (4)
k=0

Hence
min{j,d}
X
E[Xij |ℵm ] = Ni,j−k pk µ. (5)
k=0

Assuming that the numbers of claims paid from different origin years are
uncorrelated,
 
min{j,d}
paid
X paid
V [Nij |ℵm ] = V  Ni,j−k,k |ℵm 
k=0
min{j,d}
X paid
= V [Ni,j−k,k |ℵm ] (6)
k=0
min{j,d}
X
= Ni,j−k pk (1 − pk ) (7)
k=0

Hence,
min{j,d} min{j,d}
X X
V [Xij |ℵm ] = Ni,j−k pk σ 2 + Ni,j−k pk (1 − pk )µ2
k=0 k=0
min{j,d}
X
Ni,j−k pk σ 2 + µ2 (1 + pk )


=
k=0
min{j,d}
X
Ni,j−k pk σ 2 + µ2


≈
k=0

With this approximation,

min{j,d}
X
2 2


V [Xij |ℵm ] ≈ σ + µ Ni,j−k pk = ϕE[Xij |ℵm ] (8)
k=0

8
 σ 2 +µ2
where ϕ = µ .

Since the variance is proportional to the mean, an over-dispersed Poisson

model can be used, with the model for the mean being
min{j,d} min{j,d}
X X
E[Xij |ℵm ] = Ni,j−k pk µ = Ni,j−k ψk (9)
k=0 k=0

where ψk = µpk . Now

d
X d
X d
X
ψk = µpk = µ pk = µ (10)
k=0 k=0 k=0
ψ
and hence pk = . Also, the first two moments of the claim severity
Pd k
ψk
k=0
Pd Pd
distribution can be derived using µ = k=0 ψk and σ 2 + µ2 = ϕµ = ϕ k=0 ψk .
Thus,
d
X
µ= ψk (11)
k=0

and
d d
!2
X X
2
σ =ϕ ψk − ψk (12)
k=0 k=0

To summarise, the chain-ladder technique will be applied to the triangle of

the numbers of reported claims, an overdispersed Poisson model is fitted to the
paid claims traingle, with mean
min{j,d}
X
E[Xij |ℵm ] = Ni,j−k ψk
k=0

to obtain the maximum likelihood estimates of the parameters, ψˆ0 , ψˆ1 , ..., ψˆd .
The maximum likelihood estimates of the parameters required to obtain predic-
tions of RBNS and IBNR claims can then be obtained as follows:
ψ̂k
p̂k = Pd (13)
k=0 ψ̂k
Xd
µ̂ = ψ̂k . (14)
k=0

This section has derived the model which will be applied to estimate the
IBNR and RBNS delays and to predict the outstanding claims. The following
section considers how this model can be used to estimate the RBNS and IBNR
reserves.

9
5 Prediction

This section considers how the model outlined in section 4 can be used to pre-
dict outstanding claims. A key feature of the model described in this paper
is that it is possible to separate the reserves for RBNS and IBNR claims. In
the actuarial literature, a lot of attention has been given to the derivation of
stochastic models for existing deterministic models. In fact, we would argue
that too much attention has been paid to the minutiae of the philosophical un-
derpinning of methods such as the chain ladder technique. We believe that is is
more important to investigate ways to improve the insights and inferences that
can be obtained from stochastic methods that look at the problem in different
ways from the standard deterministic methods.
The model for the triangular arrays (ℵm , △m ) = {(Nij , Xij ) : (i, j) ∈ Am }
can be extended in a natural way to the random variables Nij , (i, j) ∈ Bm and
Xij , (i, j) ∈ Cm , where

Bm = {(i, j) ∈ N20 : 1 ≤ i ≤ m, 0 ≤ j ≤ m − 1}

and
Cm = {(i, j) ∈ N20 : 1 ≤ i ≤ m, 0 ≤ j ≤ m + d − 1}.
The random variables thus appear in this format

N10 ... N1,m−1 X10 ... X1,m+d−1

N20 ... N2,m−1 X20 ... X2,m+d−1
.. .. .. ..
. . . .
Nm0 ... Nm,m−1 Xm0 ... Xm,m+d−1 .

The estimates of outstanding claims are obtained by summing the predicted

values of incremental claims. We therefore require a prediction for the expected
incremental paid claims, Xil , where l > n − i, given (ℵm , △m ). The future
values of the paid claims will be made up of two separate elements, and the
prediction methods are different for each of these. The RBNS claims arise from
the claims which have already been reported: in other words, they come from
the values of Nij in ℵm which are already known. The IBNR claims arise from
the claims which are yet to be reported: they come from the future values of
Nij and, in this paper, these are predicted using the chain ladder technique.
Thus, the expected RBNS claims are
min{j,d}
X
µ pk Ni,j−k (15)
k=i−m+j

10
and the expected IBNR claims are
min{i−m+j−1,d}
X
µ pk N̂i,j−k (16)
k=0

where we have emphasised the difference from the RBNS claims by using the
notation N̂i,j−k which is a forecast of number of reported claims.
The estimates of the RBNS and IBNR claims can be obtained by substi-
tuting in the estimates of µ and pk , (13) and (14). This is illustrated in the
following section. Also of interest are prediction errors and predictive distribu-
tions. There are a number of ways to approach this: analytically, bootstrapping
or Bayesian methods. Given the relative complexity of the formulae for the
RBNS and IBNR claims, we do not believe that analytical expressions for the
prediction errors would be straightforward to derive. Also, prediction errors in
themselves are of limited practical usefulness: predictive distributions are really
required for capital setting and solvency requirements. For these reasons, we
would recommend that bootstrapping or Bayesian methods are preferable: see,
for example, England and Verrall (2006). In this paper, we concentrate on the
properties of the parameter estimates, the reserve estimates, and the implica-
tions of the proposed method in terms of understanding the characteristics of
the delays in greater detail. Prediction errors and predictive distributions are
not considered any further here: this will be considered in future research when
the new methodology has been developed further.

6 Data study

The application of the model is illustrated in this section, using data from Royal
& Sun Alliance. The data relate to a portfolio of motor policies, and in this
example the auto third part liability (TPL) data is considered. The reason for
choosing this data set is that we expect there to be reasonably long settlement
delays (RBNS delays). This could be of particular interest as the methodology
developed in this paper explicitly models the RBNS delay. The data displayed
in Table 1 is inflation corrected, so that
Yij
Xij :=
δi+j

where Yij , (i, j) ∈ A10 , are the observed payments and δi is an inflation index,
1 ≤ i ≤ 10. In a full analysis of a dataset such as this, the inflation index could
be modeled independently, for example by a time series which should then be
used in the prediction, Xij δj for j ≥ 10−i+1. For the purpose of this paper, we
assume that the claims inflation has already been estimated, and we concentrate
on modeling the inflation corrected payments, △10 , which are shown in Table
1.

11
i\j| 0 1 2 3 4 5 6 7 8 9
--------------------------------------------------------------------
1 |451288 339519 333371 144988 93243 45511 25217 20406 31482 1729
2 |448627 512882 168467 130674 56044 33397 56071 26522 14346
3 |693574 497737 202272 120753 125046 37154 27608 17864
4 |652043 546406 244474 200896 106802 106753 63688
5 |566082 503970 217838 145181 165519 91313
6 |606606 562543 227374 153551 132743
7 |536976 472525 154205 150564
8 |554833 590880 300964
9 |537238 701111
10 |684944

Table 1: The paid run-off triangle, Xij , (i, j) ∈ A10 , for the auto TPL data.

The incurred counts are shown in Table 2.

i\j| 0 1 2 3 4 5 6 7 8 9
----------------------------------------------------------
1 | 6238 831 49 7 1 1 2 1 2 3
2 | 7773 1381 23 4 1 3 1 1 3
3 |10306 1093 17 5 2 0 2 2
4 | 9639 995 17 6 1 5 4
5 | 9511 1386 39 4 6 5
6 |10023 1342 31 16 9
7 | 9834 1424 59 24
8 |10899 1503 84
9 |11954 1704
10 |10989

Table 2: The number of reported claims, Nij , (i, j) ∈ A10 , for the auto TPL.

For these data, as sometimes occurs in practice, some of the reported claims are
settled with no payment. This can occur if, for example, there is consideration
about who carries responsibility for a claim, fraud or similar. These are referred
to as zero claims, and are, by nature, different from the claims which result in
a payment. They can be dealt with in a number of different ways: for example,
it would be possible to include in this set of claim numbers a separate class
for those which are settled at zero. An alternative approach which leads to the
same model is to use a claim severity distribution which is of mixed type. In this
paper, we use the latter approach and use a claims severity distribution which a
discrete probability that a claim is zero combined with a continuous claim size
distribution. The probability that a claim is settled at zero is denoted by Q,
Q ∈ [0, 1), and Q is assumed to be a known constant for (i, j) ∈ Am . Thus, the

12
 (k)
distribution of claim payments is such that P (Yij = 0) = Q and the density
(k) (k)
ofYij |Yij > 0 is denoted by f . The most natural assumption for f is that
it is a Gamma distribution. However, for the purposes of this paper, we only
consider the first two moments of this distribution. The expert advice from the
company has been used to determine the fraction of reported zero-claims, , and
the maximal possible RBNS delay, d ≤ 10. In this case (Q, d) = (0.2, 7).
Note that the mean and variance of the (non-zero) claim distribution, f , can
be obtained from the estimates of µ and σ 2 , using:
(k) (k) (k)
µ = E[Yij ] = (1 − Q)E[Yij |Yij > 0]

and
 2
(k) (k) (k) (k) (k)
σ 2 = V [Yij ] = (1 − Q)V [Yij |Yij > 0] + Q(1 − Q) E[Yij |Yij > 0] .

(k) (k)
Thus, the estimate of E[Yij |Yij > 0] is

ψˆk
Pd
k=0
1−Q
(k) (k)
and the estimate of V [Yij |Yij > 0] is
h i
ψˆk (1 − Q)φ̂ − k=0 ψˆk
Pd Pd
k=0

(1 − Q)2

The chain ladder technique has been applied to the data in Table 2, in order
to estimate the numbers of IBNR claims. The overdispersed Poisson distribution
derived in Section 4 has then been applied and the prediction of IBNR and
RBNS claims has been conducted as proposed in Section 5. The chain ladder
technique applied to the triangle of the reported numbers of claims provides
estimates of the development factors, and it is straightforward to convert these
into the distribution of the IBNR delay. The distribution of the IBNR delay gives
the proportion of ultimate number of claims which is expected to be reported
in each development period. The development pattern and distribution of the
IBNR delay are shown in Table 4.

13
 | Development IBNR
j | Factor Delay
--------------------------------
0 | 0.8752
1 | 1.1353 0.1184
2 | 1.0038 0.0038
3 | 1.0009 0.0009
4 | 1.0003 0.0003
5 | 1.0003 0.0003
6 | 1.0002 0.0002
7 | 1.0001 0.0001
8 | 1.0003 0.0003
9 | 1.0004 0.0004

Table 4: Development factors and distribution of the IBNR delay for the
numbers of reported claims.

For the paid claim amounts the RBNS delay is given by the estimates of the
parameters p0 , p1 , · · · , p7 , which are shown in Table 5.

j | 0 1 2 3 4 5 6 7
------------------------------------------------------------
p | 0.3637 0.2881 0.1134 0.0852 0.0661 0.0358 0.0255 0.0222

Table 5: Maximum likelihood estimates of pl , 0 ≤ l ≤ 7.

Notice that IBNR delay (the last line in Table 4) and the IBNS delay (in Table
5) both sum to 1.
The estimates of the mean and variance of an individual (non-zero) claim
severity are 203.01 and 3,496,125.
As was mentioned at the beginning of this section, the TPL claims are
expected to have relatively long settlement delays (RBNS delays) as bodily
injury claims often take a long time to settle, and this can seen from the left
hand plot in Figure 1, since a long time after the majority of claims has been
reported there are still some significant payments. Also notice that p0 may be
relatively small. As claims happen, on average, in the middle of the year there
is on average only half a year to receive the final payment in order to finish in
the category of claims related to p0 . All other delay periods are full years.
Figure 1 compares the IBNR delay (from Table 4) with the RBNS delay
(from Table 5).

The average IBNR delay estimated is 0.14 years whereas the average RBNS
delay is 1.52 years. Hence the RBNS reserve is expected to be about ten times
as large as the IBNR reserve because the individual claims are assumed iid.

14
 0.8

0.6

0.4

0.2

0.0

0 1 2 3 4 5 6 7 8 9

Figure 1: The dotted line represents the IBNR delay and the solid lines repre-
sents the RBNS delay.

Using these parameter estimates, the IBNR and RBNS can be obtained us-
ing the expressions in section 5. Table 5 shows the IBNR reserves (from claims
which are not yet reported), the RBNS reserves (reported claims not yet paid),
and the total reserve.

15
 i | IBNR RBNS TOTAL CHAIN LADDER
---------------------------------------------------
2 | 628 605 1,233 1,685
3 | 1,350 4,514 5,863 29,379
4 | 1,510 43,623 45,133 60,638
5 | 1,967 94,526 96,493 101,158
6 | 2,579 171,633 174,212 173,802
7 | 3,168 299,136 302,304 249,349
8 | 5,349 509,334 514,684 475,992
9 | 14,280 852,144 866,423 763,919
10 | 254,499 1,135,678 1,390,177 1,459,860
Total | 285,329 3,111,192 3,396,521 3,315,779

Table 6: The row wise reserve estimates split into IBNR and RBNS claims,
together with the chain ladder estimates

As discussed above, the IBNR delay is (on average) shorter than the RBNS
delay, and hence the RBNS reserve is expected to be larger then the IBNR
reserve. The actual estimates divide the reserves such that the RBNS reserve
takes up 91.6% of the total reserve and the IBNR only 8.4%: roughly 10 : 1 as
suggested above. The chain ladder reserves include both the IBNR and a part
of the RBNS claims, but it is not possible to split them.

7 Conclusion

This paper has developed a new stochastic model for claims reserving, which has
a number of advantages over the standard approaches based on a single triangle
of data (such as the chain ladder technique). A significant extra element in the
results is that the sources of the delay in the claims process are split into the
IBNR and RBNS components. We believe that this approach has the potential
to make real improvements in the practical approaches to reserving, and the
data study in Section 6 illustrates this.
The approach taken in this paper steers a middle course between the crude
methods based on a single triangle and the very detailed methods based on
data at the individual claim level. We believe that, in a practical context,
this is a completely realistic approach: it does not throw away a lot of useful
information, as the chain ladder technique does, but nor does it make very heavy
extra computational demands.
The basic model described in this paper may be useful for some sets of
data, but we would suggest that more development is needed, particularly in
order to relax the assumption that the claims are identically distributed. Also,
it would be useful to develop full predictive distributions using, for example,
bootstrapping.

16
References

Bühlmann, H., Schnieper, R. and Straub, E. (1980): Claims reserves in casualty

insurance based on a probability model. Mitteilungen der Vereinigung Schweiz-
erischer Versicherungsmathematiker.

England, P. D. and Verrall, R. J. (1999): Analytic and bootstrap estimates of

prediction errors in claims reserving. Insurance: Mathematics and Economics,
vol. 25, 281-293.

England, P. D. and Verrall, R. J. (2002): Stochstic claims reserving in gen-

eral insurance (with discussion). British Actuarial Journal, vol. 8, 443-544

England, P. D. and Verrall, R. J. (2006): Predictive Distributions of Outstand-

ing Liabilities in General Insurance. Annals of Actuarial Science, vol. 1, 221-270.

Hachemeister, C. A. and Standard, J. N. (1975): IBNR claims count estima-

tion with static lag functions. 12th Astin Colloquium, IAA, Portimao, Portugal.

Liu, H. and Verrall, R. (2009a): Predictive Distributions for Reserves which

Separate True IBNR and IBNER Claims. Astin Bulletin, vol.39, 35-60.

Liu, H. and Verrall, R. (2009b): A Bootstrap Estimate of the Predictive Dis-

tribution of Outstanding Claims for the Schnieper Model. To appear in Astin
Bulletin.

Mack, T. (1991): A simple parametric model for rating automobile insurance

or estimating IBNR claims reserves. Astin Bulletin, vol. 21, no. 1, 93-108.

McCullagh, P. and Nelder, J. A. (1989): Generalized Linear Models, second

edition. Chapman & Hall, London.

Neuhaus, W. (2004): On the estimation of outstanding claims. Australian

Actuarial Journal, vol. 10, 485-518.

Norberg, R., (1986): A contribution to modelling of IBNR claims. Scandi-

navian Actuarial Journal, 155-203.

Norberg, R., (1993): Prediction of outstanding liabilities in non-life insurance.

Astin Bulletin, vol. 23, no. 1, 95-115.

Norberg, R., (1999): Prediction of outstanding claims: Model variations and

extensions. Astin Bulletin, vol. 29, no. 1, 5-25.

Ntzoufras, I. and Dellaportas, P (2002): Bayesian modeling of outstanding lia-

17
 bilities incorporating claim count uncertainty. North American Actuarial Jour-
nal, vol. 6, no. 1, 113-137.

Renshaw, A. E. and Verrall, R. (1994): The stochastic model underlying the

chain ladder technique. Actuarial research paper no. 63.

Schmidt, K. D. (2007): A bibliography on loss reserving.

Available on: http://www.math.tu-dresden.de/sto/schmidt/dsvm/reserve.pdf

Schnieper, R. (1991): Separating true IBNR and IBNER claims. Astin Bul-
letin, vol. 21, 111-127.

Taylor, G.C. (1986): Claims reserving in non-life insurance. North Holland.

Taylor, G. C. and McGuire, G. (2004): Loss reserving with GLM - a case study.
CAS discussion paper program, 327-391.

Verrall, R. J. (1990): Bayes and empirical Bayes estimates for the chains ladder.
Astin Bulletin, vol. 20, 217-243.

Verrall, R. J. (1991): Chain ladder and maximum likelihood. Journal of the

Institute of Actuaries, vol. 118, 489-499.

Verrall, R. J. and England, P. D. (2005): Incorporating expert opinion into

a stochastic model for the chain-ladder technique. Insurance, Mathematic and
Economics, vol. 37, 355-370 .

Wright, T. S. (1990): A stochastic method for claims reserving in general insur-

ance. Journal of the Institute of Actuaries, vol. 117, 677-731.

Wüthrich, M.V. and Merz, M. (2008): Stochastic Claims Reserving Methods

in Insurance. Wiley

Richard Verrall, Faculty of Actuarial Science and Insurance, Cass Business

School, City University, London, [email protected]
Jens Perch Nielsen, Faculty of Actuarial Science and Insurance, Cass Business
School, City University, London
Anders Hedegaard Jessen, Department of Mathematical Sciences, University of
Copenhagen

18

View publication stats