ACTUARIAL APPLICATIONS OF A HIERARCHICAL

INSURANCE CLAIMS MODEL

BY

EDWARD W. FREES, PENG SHI AND EMILIANO A. VALDEZ

ABSTRACT

This paper demonstrates actuarial applications of modern statistical methods

that are applied to detailed, micro-level automobile insurance records. We con-
sider 1993-2001 data consisting of policy and claims files from a major Singa-
porean insurance company. A hierarchical statistical model, developed in prior
work (Frees and Valdez (2008)), is fit using the micro-level data. This model
allows us to study the accident frequency, loss type and severity jointly and to
incorporate individual characteristics such as age, gender and driving history
that explain heterogeneity among policyholders.
Based on this hierarchical model, one can analyze the risk profile of either
a single policy (micro-level) or a portfolio of business (macro-level). This paper
investigates three types of actuarial applications. First, we demonstrate
the calculation of the predictive mean of losses for individual risk rating. This
allows the actuary to differentiate prices based on policyholder characteristics.
The nonlinear effects of coverage modifications such as deductibles, policy
limits and coinsurance are quantified. Moreover, our flexible structure allows
us to “unbundle” contracts and price more primitive elements of the contract,
such as coverage type. The second application concerns the predictive distrib-
ution of a portfolio of business. We demonstrate the calculation of various
risk measures, including value at risk and conditional tail expectation, that
are useful in determining economic capital for insurance companies. Third, we
examine the effects of several reinsurance treaties. Specifically, we show the
predictive loss distributions for both the insurer and reinsurer under quota
share and excess-of-loss reinsurance agreements. In addition, we present an
example of portfolio reinsurance, in which the combined effect of reinsurance
agreements on the risk characteristics of ceding and reinsuring company are
described.

KEYWORDS

Long-tail regression, copulas, risk measures, reinsurance.

Astin Bulletin 39(1), 165-197. doi: 10.2143/AST.39.1.2038061 © 2009 by Astin Bulletin. All rights reserved.
166 E.W. FREES, P. SHI AND E.A. VALDEZ

1. INTRODUCTION
Actuaries and other financial analysts that work with short term coverages
such as automobile insurance and healthcare expenditures typically have
massive amounts of in-company data. With modern computing equipment,
analysts can readily access data at the individual policyholder level that we
term “micro-level”. Actuaries use statistical models to summarize micro-level
data that subsequently need to be interpreted properly for financial decision-
making. For example, automobile insurers typically differentiate premium rates
based on policyholder characteristics such as age, gender and driving history.
Gourieroux and Jasiak (2007) have dubbed this emerging field the “micro-
econometrics of individual risk”.
Developing a statistical model to substantiate rate differentials may not be
straightforward because policyholder characteristics can affect the frequency
(number of accidents) and the severity (amount or intensity of the accident)
in different ways. Moreover, policies differ based on the type of coverage offered
as well as payment modifications (such as deductibles, upper limits and coin-
surance). Actuaries are also interested in other financial measures in addition
to those used for pricing. For example, actuaries use measures that summarize
portfolio risks for capital allocation and solvency purposes. As another exam-
ple, actuaries involved in managing risk through reinsurance use statistical
models to calibrate reinsurance treaties.
This paper demonstrates analyses that can be used for pricing, economic cap-
ital allocation, solvency and reinsurance based on a statistical model of a specific
database. The structure of the database encompasses policyholder and claim
files and is widely used. To illustrate, all property and casualty (general) insurers
domiciled in Singapore use this structure to report their data to a quasi-govern-
mental organization, the General Insurance Association (G.I.A.) of Singapore.
Our data are from a Singaporean insurance company. From the policyholder
file, we have available several policyholder characteristics that can be used to
anticipate automobile insurance claims. Additional data characteristics are in
Section 2.2. For each policy i, we are interested in predicting:
• Ni – the number of losses and
• yijk – the loss amount, available for each loss, j = 1, …, Ni , and the type of
loss k = 1, 2, 3.

When a claim is made, it is possible to have one or a combination of three types

of losses. We consider: (1) losses for injury to a party other than the insured
yij1, (2) losses for damages to the insured, including injury, property damage, fire
and theft yij2, and (3) losses for property damage to a party other than the
insured yij3. Occasionally, we shall simply refer to them as “injury”, “own dam-
age” and “third party property”. It is not uncommon to have more than one
type of loss incurred with each accident.
The hierarchical model was developed in our prior work (Frees and Valdez
(2008)). This model allows for:
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 167

• risk rating factors to be used as explanatory variables that predict both the
frequency and multivariate severity,
• the long-tail nature of the distribution of insurance claims through the GB2
(generalized beta of the second kind) distribution and
• the “two-part” distribution of losses. When a claim occurs, we do not
necessarily realize all three types of losses. Each type of loss may equal zero
or may be continuous and hence be comprised of “two parts”. Further, we
allow for
• losses to be related through a t-copula specification.

To keep this paper self-contained, this statistical model is described in the

Section 2.3. To give readers a feel for the historical precedents of this model,
Section 2.1 provides a review of the literature.
We fit this model to 1993-2000 company data. The statistical model allows
us to provide predictions for a 2001 portfolio of n = 13,739 policies. Several risk
rating factors known at the beginning of the year are used to develop predic-
tions. The focus of this paper is to show how predictions from this statistical
model can be used for rating and portfolio risk management. Due to the
hierarchical nature (multiple layers) of our model, we use a simulation-based
approach to calculate the predictions. The simulation procedures are described
in Appendix A.3.
Section 3 discusses individual risk rating by the predictive mean, a measure
that actuaries base prices upon. To begin, the model allows us to quantify
differences by risk rating factors. Moreover, although our data are based on a
comprehensive coverage that offers protection for “injury”, “own damage”
and “third party property”, we can compute predictive means for each type of
coverage. This allows the analyst to assess the effects of “unbundling” the com-
prehensive coverage.
We also quantify the effect of other policy modifications, such as if the
policy paid only for one claim during the year or the effect of policy limits such
as deductibles, upper coverage limit (by type of coverage) and coinsurance.
This type of flexibility allows the actuary to design a coherent pricing struc-
ture that is appealing to consumers.
Section 4 examines predictive distributions for a portfolio or group of poli-
cies. Here, we seek to gain insights into the effects of “micro-level” changes,
such as the introduction of a policy deductible or upper limit, on the “macro-
level” group basis. In principle, the group could be selected by geographic
region, sales agent type or some other method of clustering. For illustrative
purposes, we simply take random subsets of our portfolio of n = 13,739 policies.
Because of our concern of macro-level effects, Section 4 goes beyond simply
the predictive mean and examines other summary measures of the distribution.
We do this by Monte Carlo methods, where we simulate the loss frequency, type
and severity for each policy in the group.
The portfolio summary measures include well-known tail summary measures
such as the Value-at-risk (VaR) and the conditional tail expectation (CTE ).
168 E.W. FREES, P. SHI AND E.A. VALDEZ

Thus, the tools that we propose can be used to determine economic capital for
insurers as well as for solvency testing. For example, we are able to quantify
the effect of imposing an upper limit on injury protection on the VaR for the
portfolio. Although we allow for long-tail severity for individual policies, we
can give practical guidance as to when approximate normality (via a central
limit theorem) holds for a portfolio of policies.
Section 5 retains the focus on portfolio of policies but now with an eye
towards quantifying the impact of reinsurance treaties. Essentially, we are
interested in “macro-level” changes from our micro-level (individual policy)
model. We discuss proportional reinsurance as well as non-proportional rein-
surance, and examine the effect of reinsurance agreements on the losses dis-
tribution. An example of portfolio reinsurance is given, and the combined
effects of reinsurance agreements are investigated by showing the tail summary
measures of losses for both insurer and reinsurer.
Section 6 provides a summary and closing remarks.

2. HIERARCHICAL INSURANCE CLAIMS

2.1. Literature Review

There is a rich literature on modeling the joint frequency and severity distribu-
tion of automobile insurance claims. To distinguish this modeling from classical
risk theory applications (see, for example, Klugman, Panjer and Willmot, 2004),
we focus on cases where explanatory variables, such as policyholder characteris-
tics, are available. There has been substantial interest in statistical modeling of
claims frequency, see Boucher and Denuit (2006) for a recent example. How-
ever, the literature on modeling claims severity, especially in conjunction with
claims frequency, is less extensive. One possible explanation, noted by Coutts
(1984), is that most of the variation in overall claims experience may be attrib-
uted to claim frequency (at least when inflation was small). Coutts (1984) also
remarks that the first paper to analyze claim frequency and severity separately
seems to be Kahane and Levy (1975).
Brockman and Wright (1992) provide an earlier overview of how statistical
modeling of claims and severity can be helpful for pricing automobile coverage.
For computational convenience, they focused on categorical pricing variables
to form cells that could be used with traditional insurance underwriting forms.
Renshaw (1994) shows how generalized linear models can be used to analyze
both the frequency and severity portions based on individual policyholder level
data. Hsiao et al. (1990) note the “excess” number of zeros in policyholder
claims data (due to no claims) and compare and contrast Tobit, two-part and
simultaneous equation models, building on the work of Weisberg and Tomber-
lin (1982) and Weisberg et al. (1984). However, all of these papers use grouped
data, not individual policyholder level data as in this paper.
Pinquet (1997, 1998) provides a more modern statistical approach, fitting not
only cross-sectional data but also following policyholders over time. Pinquet
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 169

was interested in two lines of business, claims at fault and not at fault with
respect to a third party. For each line, Pinquet hypothesized a frequency and
severity component that were allowed to be correlated to one another. In par-
ticular, the claims frequency distribution was assumed to be bivariate Poisson.
Severities were modeled using lognormal and gamma distributions. Also at the
individual policyholder level, Frangos and Vrontos (2001) examined a claim fre-
quency and severity model, using negative binomial and Pareto distributions,
respectively. They used their statistical model to develop experience rated
(bonus-malus) premiums.

2.2. Data
Our statistical model of insurance claims is based on detailed, micro-level
automobile insurance records. Specifically, we analyzed information from two
databases: the policy and claims files. The policy file consists of all policy-
holders with vehicle insurance coverage purchased from a general insurer during
the observation period. Each vehicle is identified with a unique code. This file
provides characteristics of the policyholder and the vehicle insured, such as age
and gender, and type and age of vehicle insured. The claims file provides a
record of each accident claim that has been filed with the insurer during the
observation period and is linked to the policyholder file. For this analysis, we
ignored claims where no payments are made. Unfortunately, there was no infor-
mation in the files as to whether the claim was open or settled.
Some insurers also use a payment file that consists of information on each
payment that has been made during the observation period and is linked to the
claims file. Although it is common to see that a claim will have multiple pay-
ments made, we do not use that information for this paper and consider the
aggregate of all payments that arise from each accident event. See Antonio et
al. (2006) for a recent description of the claims “run-off ” problem.
The policyholder file provides several characteristics to help explain and
predict automobile accident frequency, type and severity. These characteristics
include information about the vehicle, such as type and age, as well as person
level characteristics, such as age, gender and prior driving experience. Table 1
summarizes these characteristics. These characteristics are denoted by the
vector xit and will serve as explanatory variables in our analysis. Person level
characteristics are largely unavailable for commercial use vehicles, so the
explanatory variables of personal characteristics are only used for observations
having non-commercial purposes. We also have the exposure eit , measured in
(a fraction of) years, which provides the length of time throughout the calen-
dar year for which the vehicle had insurance coverage.
With the information in the claims file, we potentially observe a trivariate
claim amount, one claim for each type. For each accident, it is possible to have
more than a single type of claim incurred; for example, an automobile accident
can result in damages to a driver’s own property as well as damages to a third
party who might be involved in the accident.
170 E.W. FREES, P. SHI AND E.A. VALDEZ

TABLE 1
DESCRIPTION OF RISK RATING FACTORS

Covariates Description

Year The year 1-9 corresponding to calendar year 1993-2001.

Vehicle Type The type of vehicle insured, either automobile (A) or others (O).
Vehicle Age The age of vehicle, in years, grouped in six categories.
Vehicle Capacity The cubic capacity of the vehicle.
Gender The gender of the policyholder, either male or female.
Age The age of the policyholder, in years, grouped in to seven categories.
NCD No Claim Discount. This is based on the previous accident record of the
policyholder.
The higher the discount, the better is the prior accident record.

To provide focus, we restrict our considerations to “non-fleet” policies;

these comprise about 90% of the policies for this company. These are policies
issued to customers whose insurance covers a single vehicle. In contrast,
fleet policies are issued to companies that insured several vehicles, for exam-
ple, coverage provided to a taxicab company, where several taxicabs are insured.
See Angers et al. (2006) and Desjardins et al. (2001) for discussions of fleet poli-
cies. The unit of observation in our analysis is therefore a registered vehicle
insured, broken down according to their exposure in each calendar year 1993 to
2001. In order to investigate the full multivariate nature of claims, we further
restrict our consideration to policies that offer comprehensive coverage, not merely
for only third party injury or property damage.
In summary, the hierarchical insurance claims model is based on observable
data consisting of

{eit , xit , Nit , Mitj , yitjk, k = 1, 2, 3, j = 1, …, Nit , t = 1, …, Ti , i = 1, …, n}.

Here, e represents exposure, x are for explanatory variables, N is the number

of claims, M is the type of claim and y represents the claim amount. Appen-
dix A.1 provides descriptive statistics of accident frequency, type of losses and
claim amount.

2.3. Statistical Model

The statistical model, from Frees and Valdez (2008), consists of three compo-
nents – each component uses explanatory variables to account for heterogeneity
among policyholders. The first component uses a negative binomial regression
model to predict accident probability. The second component uses a multinomial
logit model to predict type of losses, either third party injury, own damage, third
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 171

party property or some combination. The third component is for severity; here,
a GB2 distribution is used to fit the marginal distributions and a t-copula is
used to model the dependence of the three types of claims.
It is customary in the actuarial literature to condition on the frequency
component when analyzing the joint frequency and severity distributions. See,
for example, Klugman, Panjer and Willmot (2004). Frees and Valdez (2008)
incorporate an additional claims type layer to handle the many zeros in each
distribution (known as “two-part” data) as well as accounting for the possibility
of multivariate claims. Specifically, conditional on having observed at least one
type of claim, the random variable M describes which of the seven combinations
is observed. Table 2 provides potential values of M.

TABLE 2
VALUE OF M, BY CLAIM TYPE.

Value of M 1 2 3 4 5 6 7
Claim by Combination Observed (y1 ) ( y2 ) (y3 ) ( y1, y2 ) ( y1, y3 ) ( y2, y3 ) (y1, y2, y3 )

We are now in a position to describe the full predictive model. Suppressing the
{i} subscripts, the joint distribution of the dependent variables is:

f (N, M, y) = f (N ) ≈ f (M| N ) ≈ f (y | N, M)
joint = frequency ≈ conditional claim type ≈ conditional severity,

where f (N, M, y) denotes the joint distribution of (N, M, y).

We now discuss each of the three components. The parameter estimates
corresponding to each components are provided in Appendix A.2.

2.3.1. Frequency Component

For our purposes, we use standard count models. For these models, one uses
lit = eit exp(x
l,it bl ) to be the conditional mean parameter for the itth observa-
tional unit. Here, the vector xl,it is a subset of x it , representing the variables
needed for frequency modeling. The amount of exposure, eit, is used as an off-
set variable because a driver may have insurance coverage for only part of the
year. We use the negative binomial distribution with parameters p and r, so that
Pr(N = k) = a k r+-r 1- 1k pr (1 – p) k. Here, s = r – 1 is the dispersion parameter and
p = pit is related to the mean through (1 – pit ) / pit = lit s = eit exp(x
it bl )s.

2.3.2. Claims Type Component

The multinomial logit regression model allows us to incorporate explanatory
variables into our explanations of the claim type. This model is of the form
172 E.W. FREES, P. SHI AND E.A. VALDEZ

Pr(M = m) = exp(Vm ) / {!7s =1 exp(Vs )}, where Vm = Vit, m = x
M, it b M, m . Note that
for our application, the covariates in xM,it do not depend on the accident num-
ber j nor on the claim type m although we allow parameters ( b M, m ) to depend
on m. This portion of the model was proposed by Terza and Wilson (1990).

2.3.3. Severity Component

To accommodate the long-tail nature of claims, we use the GB2 distribution
for each claim type. This has density function

exp ^a 1 zh
f ^ yh = , (1)
y s B ^a 1 , a 2h 61 + exp ]zg @ a1 + a 2

where z = (ln y – m) / s and B(a1, a2 ) = G(a1) G(a2) / G(a1 + a2 ), the usual beta
function. Here, m is a location parameter, s is a scale parameter and a1 and a2
are shape parameters. This distribution is well known in actuarial modeling of
univariate loss distributions (see for example, Klugman, Panjer and Willmot,
2004). With four parameters, the distribution has great flexibility for fitting heavy
tailed data. Many distributions useful for fitting long-tailed distributions can
be written as special or limiting cases of the GB2 distribution; see, for exam-
ple, McDonald and Xu (1995).
We use this distribution but allow scale and shape parameters to vary by
type and thus consider a1k , a2k and sk for k = 1, 2, 3. Despite the prominence
of the GB2 in fitting distributions to univariate data, there are relatively few
applications that use the GB2 in a regression context. Recently, Sun et al.
(2008) used the GB2 in a longitudinal data context to forecast nursing home
utilization.
To accommodate dependencies among claim types, we use a parametric
copula. See Frees and Valdez (1998) for an introduction to copulas. Suppressing
the {i} subscripts, we may write the joint distribution of claims ( y1, y2, y3 ) as

F (y1, y2, y3 ) = H (F1 ( y1 ), F2 (y2 ), F3(y3 )).

Here, the marginal distribution of yk is given by Fk (·) and H(·) is the copula
linking the marginals to the joint distribution. We use a trivariate t-copula with
an unstructured correlation matrix. See Frees and Valdez (2008) for a further
motivation of the use of the t-copula.

3. PREDICTIVE MEANS FOR INDIVIDUAL RISK RATING

Given a set of risk rating factors such as in Table 1, one basic task is to arrive
at a fair price for a contract. In setting prices, often the actuary is called upon
to quantify the effects of certain policy modifications. As a basis of fair pricing
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 173

is the predictive mean, this section calculates predictive means for several alter-
native policy designs that may be of interest.
For alternative designs, we consider four random variables:
• individuals losses, yijk
• the sum of losses from a type, Si, k = yi,1, k + … + yi,Ni, k
• the sum of losses from a specific accident, SACC, i, j = yi, j,1 + yi, j, 2 + yi, j,3 , and
• an overall loss per policy, Si = Si,1 + Si,2 + Si,3 = SACC, i,1 + ... + SACC, i, Ni .

Our database is from comprehensive policies with premiums based on the

fourth random variable. The other random variables represent different ways
of “unbundling” this coverage, similar to decomposing a financial contract
into primitive components for risk analysis. The first random variable can be
thought of as claims arising from a policy that covers losses from a single acci-
dent of a certain type. The second represents claims from a policy that covers
all losses within a year of a certain type. The third variable corresponds to a
policy that covers all types of losses from a single accident.
We also examine the effect of standard coverage modifications that consist
of (1) deductibles d, (2) coverage limits u and (3) coinsurance percentages a.
As in Klugman et al. (2004), we define the function
Z0 y<d
]]
g ^ y; a, d, uh = [ a ^ y - d h d # y < u .
]]
a ]u - d g y $ u
\

From the conditional severity model, we define mik = E( yijk | Ni , Ki = k). The
random variable Ki indicates the type, for Ki = 1, 2, 3. Then, basic probability
calculations show that:

E( yijk ) = Pr(Ni = 1) Pr(Ki = k) mik , (2)

3
E(Si, k ) = mik Pr(Ki = k) ! n Pr(Ni = n), (3)
n =1
3
E(SACC, i, j ) = Pr(Ni = 1) ! mi k Pr(Ki = k), and (4)
k =1

E(Si ) = E(Si,1) + E(Si,2 ) + E(Si,3 ). (5)

Thus, to compute means, we only need to calculate the probability of number

and type of loss, as well as the expected loss given the type of loss in case of
accident. Appendix A.2 provides the necessary coefficients and probabilities
estimates.
To provide baseline comparisons, we emphasize the simplest situation, a pol-
icy without deductible, coverage limits and coinsurance modifications. In this
 174 E.W. FREES, P. SHI AND E.A. VALDEZ

case, from the severity distribution, we have an analytic expression for the con-
ditional mean of the form
B ^a 1 k + s k , a 2 k - s k h
mik = exp (x
ik bk) , (6)
B ^ a 1 k , a 1 kh

where bk, ajk, sk are parameters of the GB2 severity distribution (see Sec-
tion 2.3.3). With policy modifications, we approximate mik via simulation (see
Section A.3).
The predictive means in equations (2)-(5), by level of no claims discount (NCD)
and insured’s age, are shown in Tables 3 and 4, respectively. The calculation is
based on a randomly selected observation from the 2001 portfolio. The policy-
holder is a 50-year old female driver who owns a Toyota Corolla manufactured
in year 2000 with a 1332 cubic inch capacity. For losses based on a coverage
type, we chose own damage because the risk factors NCD and age turned out
to be statistically significant for this coverage type. (Appendix A.2 shows
that both NCD and age are statistically significant variables in at least one
component of the predictive model, either frequency, type or severity.) The
point of this section is to understand their economic significance.
Using equation (6), Table 3 shows that the insured who enjoys a higher no
claim discount has a lower expected loss; this result holds for all four random

TABLE 3
PREDICTIVE MEAN BY LEVEL OF NCD

Level of NCD
Type of Random Variable
0 10 20 30 40 50

Individual Loss (Own Damage) 330.67 305.07 267.86 263.44 247.15 221.76
Sum of Losses from a Type (Own Damage) 436.09 391.53 339.33 332.11 306.18 267.63
Sum of Losses from a Specific Event 495.63 457.25 413.68 406.85 381.70 342.48
Overall Loss per Policy 653.63 586.85 524.05 512.90 472.86 413.31

TABLE 4
PREDICTIVE MEAN BY INSURED’S AGE

Insured’s Age
Type of Random Variable
# 21 22-25 26-35 36-45 46-55 56-65 $ 66

Individual Loss (Own Damage) 258.41 238.03 198.87 182.04 221.76 236.23 238.33
Sum of Losses from a Type (Own Damage) 346.08 309.48 247.67 221.72 267.63 281.59 284.62
Sum of Losses from a Specific Event 479.46 441.66 375.35 343.59 342.48 350.20 353.31
Overall Loss per Policy 642.14 574.24 467.45 418.47 413.31 417.44 421.93
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 175

variables. This is consistent with our intuition and, as shown in the Appendix,
with the statistical model fitting results.
Table 4 presents a more complex nonlinear pattern for insured’s age. For
each random variable, predictive means are at their highest at the youngest
age group, decrease as age increases and remain relatively stable thereafter.
Figure 1 presents a graphical display in the lower left-hand corner.
When different coverage modifications are incorporated, we need to simulate
the amount of losses to calculate predictive means. Tables 5 and 6 show the
effects of different policy designs under various combinations of NCD and
insured’s age. The first row under each of the four random variables corresponds
to the predictive mean for the policy without any coverage modification;
here, readers will notice a slight difference between these entries and the cor-
responding entries in Tables 3 and 4. This is due to simulation error. We used
5,000 simulated values.
To understand the simulation error, Figure 1 compares the analytic and sim-
ulated predictive means. Here, we consider the case of no deductible, policy
limits and coinsurance, so that the analytic result in equation (6) is available.
The upper two panels shows the relationship between predictive mean and
NCD, whereas the lower two panels are for insured’s age. The two panels on
the left are for the analytic result, whereas the two panels on the right are for
the simulation results. For the simulated results, the lines provide the 95%
confidence intervals. The width of these lines show that the simulation error
is negligible for our purposes – for other purposes, one can always reduce the
simulation error by increasing the simulation size.
Table 5 shows the simulated predictive mean at different levels of NCD
under various coverage modifications. As expected, any of a greater deductible,
lower policy limit or smaller coinsurance results in a lower predictive mean.
Deductibles and policy limits change the predictive mean nonlinearly, whereas
coinsurance changes the predictive mean linearly. For example, the predictive
mean decreases less when the deductible increases from 250 to 500, compared
to the decrease when deductible increases from 0 to 250. This pattern applies
to all the four loss variables at all NCD levels. The effect of policy limit depends
on the expected loss. For random variables with a small expected loss (e.g.
individual loss and sum of losses from a type), there is little difference in pre-
dictive means between policies with a 50,000 limit and no limit. In contrast,
for random variables with large expected losses (e.g. sum of losses from a specific
event and overall losses), the difference in predictive means can be greater when
limit increases from 25,000 to 50,000 than an increase from 50,000 to no limit.
Table 6 shows the effect of coverage modifications on the predictive mean for
the insured at different age categories. As with Table 5, any of a higher deductible,
lower coverage limit or lower coinsurance percentage results in a lower pre-
dictive mean. The combined effect of three kinds of coverage modifications
can be derived from the three marginal effects. For example, when the insured’s
age is in the 26-35 category, the predictive mean of individual loss with
deductible 250, coverage limit 25,000 and coinsurance 0.75 can be calculated
176 E.W. FREES, P. SHI AND E.A. VALDEZ

700

700
600

600
Simulated Mean
Analytic Mean

500

500
400

400
300

300
200

200
0 10 20 30 40 50 0 10 20 30 40 50

NCD NCD
100 200 300 400 500 600 700

100 200 300 400 500 600 700

Simulated Mean
Analytic Mean

<21 22−25 26−35 36−45 46−55 56−65 >65 <21 22−25 26−35 36−45 46− 55 56−65 >65

Age Category Age Category

FIGURE 1: Analytic and Simulated Predictive Means.

The left-hand panels provide analytical means, the right-hand panels are based on simulated values.
From top to bottom, the curves represent individual loss, sum of losses from a type, sum of losses
from a specific accident, and overall loss per policy, respectively.

from predictive mean of individual loss with deductible 250 and predictive
mean of individual loss with coverage limit 25,000, that is (170.54 + 189.64 –
191.13) * 0.75 = 126.79. Similar results can be derived for all the four random
variables under different NCD or insured’s age values.
Comparing Tables 5 and 6, the effect of NCD has a greater effect on the
predictive mean than that of insured’s age, in the sense that the range of predictive
means is greater under alternative NCD levels compared to age levels. For
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 177

TABLE 5
SIMULATED PREDICTIVE MEAN BY LEVEL OF NCD AND COVERAGE MODIFICATIONS

Coverage Modification Level of NCD

Deductible Limits Coinsurance 0 10 20 30 40 50

INDIVIDUAL LOSS (OWN DAMAGE)

0 none 1 339.78 300.78 263.28 254.40 237.10 227.57
250 none 1 308.24 271.72 235.53 227.11 211.45 204.54
500 none 1 280.19 246.14 211.32 203.43 188.94 184.39
0 25,000 1 331.55 295.08 260.77 250.53 235.42 225.03
0 50,000 1 337.00 300.00 263.28 254.36 237.10 227.27
0 none 0.75 254.84 225.59 197.46 190.80 177.82 170.68
0 none 0.5 169.89 150.39 131.64 127.20 118.55 113.78
250 25,000 0.75 225.00 199.51 174.76 167.43 157.33 151.50
500 50,000 0.75 208.05 184.02 158.49 152.54 141.70 138.07
SUM OF LOSSES FROM A TYPE (OWN DAMAGE)
0 none 1 445.81 386.04 334.05 322.09 294.09 273.82
250 none 1 409.38 352.94 302.65 291.29 265.41 248.43
500 none 1 376.47 323.36 274.82 264.12 239.90 225.93
0 25,000 1 434.86 378.55 330.50 316.57 291.78 270.39
0 50,000 1 442.35 385.05 333.98 321.87 294.07 273.40
0 none 0.75 334.36 289.53 250.54 241.56 220.56 205.37
0 none 0.5 222.91 193.02 167.03 161.04 147.04 136.91
250 25,000 0.75 298.82 259.09 224.32 214.33 197.33 183.75
500 50,000 0.75 279.75 241.77 206.06 197.94 179.91 169.13
SUM OF LOSSES FROM A SPECIFIC EVENT
0 none 1 512.74 444.50 407.84 390.87 376.92 350.65
250 none 1 475.56 410.12 374.90 358.54 346.58 323.41
500 none 1 439.84 377.11 343.33 327.64 317.47 297.37
0 25,000 1 483.88 433.28 394.80 380.54 359.31 340.67
0 50,000 1 494.20 442.06 401.99 388.21 367.02 348.79
0 none 0.75 384.55 333.38 305.88 293.15 282.69 262.98
0 none 0.5 256.37 222.25 203.92 195.44 188.46 175.32
250 25,000 0.75 335.02 299.17 271.39 261.15 246.73 235.08
500 50,000 0.75 315.98 281.00 253.11 243.74 230.68 221.64
OVERALL LOSS PER POLICY
0 none 1 672.68 572.51 516.77 493.93 466.26 421.10
250 none 1 629.88 533.50 479.64 457.56 432.43 391.14
500 none 1 588.55 495.85 443.87 422.63 399.85 362.37
0 25,000 1 634.81 555.90 499.72 479.90 445.04 408.81
0 50,000 1 649.67 568.30 509.52 490.46 454.84 418.92
0 none 0.75 504.51 429.39 387.58 370.45 349.69 315.82
0 none 0.5 336.34 286.26 258.39 246.96 233.13 210.55
250 25,000 0.75 444.01 387.67 346.94 332.65 308.41 284.14
500 50,000 0.75 424.16 368.72 327.46 314.37 291.32 270.15
178 E.W. FREES, P. SHI AND E.A. VALDEZ

TABLE 6
SIMULATED PREDICTIVE MEAN BY INSURED’S AGE AND COVERAGE MODIFICATIONS

Coverage Modification Level of Insured’s Age

Deductible Limits Coinsurance ≥ 21 22-25 26-35 36-45 46-55 56-65 ≥ 66

INDIVIDUAL LOSSES (OWN DAMAGE)

0 none 1 252.87 242.94 191.13 179.52 220.59 233.58 235.44
250 none 1 226.93 219.16 170.54 160.61 197.57 211.76 213.42
500 none 1 204.13 198.39 152.52 144.00 177.44 192.24 193.78
0 25,000 1 246.94 238.24 189.64 178.33 217.14 230.52 232.35
0 50,000 1 250.64 242.62 191.13 179.46 219.32 233.38 235.44
0 none 0.75 189.65 182.21 143.35 134.64 165.44 175.19 176.58
0 none 0.5 126.43 121.47 95.57 89.76 110.29 116.79 117.72
250 25,000 0.75 165.75 160.84 126.79 119.57 145.60 156.52 157.75
500 50,000 0.75 151.42 148.56 114.39 107.95 132.12 144.03 145.34
SUM OF LOSSES FROM A TYPE (OWN DAMAGE)
0 none 1 339.05 314.08 239.04 219.34 266.34 278.61 280.74
250 none 1 308.86 286.80 215.95 198.39 240.96 254.71 256.59
500 none 1 281.82 262.57 195.44 179.74 218.47 233.12 234.84
0 25,000 1 331.01 307.77 236.54 217.53 262.13 274.59 276.51
0 50,000 1 336.33 313.60 238.89 219.16 264.92 278.29 280.67
0 none 0.75 254.29 235.56 179.28 164.50 199.75 208.96 210.55
0 none 0.5 169.53 157.04 119.52 109.67 133.17 139.31 140.37
250 25,000 0.75 225.61 210.37 160.08 147.43 177.56 188.02 189.27
500 50,000 0.75 209.33 196.57 146.47 134.67 162.79 174.60 176.08
SUM OF LOSSES FROM A SPECIFIC EVENT
0 none 1 480.49 452.84 360.72 336.00 339.24 341.88 355.91
250 none 1 441.68 417.13 329.75 307.68 312.02 316.15 329.97
500 none 1 404.35 382.86 300.06 280.46 285.91 291.37 305.06
0 25,000 1 461.26 434.27 356.68 329.88 326.36 335.92 341.76
0 50,000 1 471.44 444.84 360.30 333.98 331.88 341.66 351.95
0 none 0.75 360.37 339.63 270.54 252.00 254.43 256.41 266.93
0 none 0.5 240.24 226.42 180.36 168.00 169.62 170.94 177.95
250 25,000 0.75 316.83 298.92 244.28 226.17 224.35 232.65 236.87
500 50,000 0.75 296.48 281.14 224.73 208.83 208.91 218.37 225.83
OVERALL LOSS PER POLICY
0 none 1 641.63 585.21 450.69 410.37 410.93 408.05 423.90
250 none 1 596.61 544.40 416.07 379.07 380.98 379.93 395.52
500 none 1 553.07 505.04 382.74 348.87 352.15 352.76 368.17
0 25,000 1 616.34 561.58 444.58 402.51 394.26 399.93 406.63
0 50,000 1 630.29 575.81 449.98 407.74 401.61 407.27 419.34
0 none 0.75 481.22 438.91 338.02 307.78 308.20 306.04 317.92
0 none 0.5 320.82 292.60 225.34 205.19 205.46 204.03 211.95
250 25,000 0.75 428.49 390.58 307.48 278.41 273.23 278.86 283.69
500 50,000 0.75 406.30 371.73 286.52 259.68 257.13 263.98 272.71
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 179

example, for a policy with a 500 deductible, 50,000 policy limit and 0.75 coin-
surance, the predictive mean of individual losses is 138.07 from Table 5 and
132.12 from Table 6 (the difference is due to simulation error). For this policy, the
predictive mean varies between 138.07 and 208.05 under various NCD levels, and
varies between 107.95 and 151.42 under alterative insured’s age categories.
Under other coverage modifications we observe similar results.

4. PREDICTIVE DISTRIBUTIONS FOR PORTFOLIOS

4.1. Predictive Distribution

Actuaries are trained to look beyond the mean – to manage risk, one should
understand a risk’s entire distribution. This section examines the predictive
distribution for a portfolio of risks.
In contrast, Section 3 dealt with a single illustrative contract. For a single
contract, there is a large mass at zero (about 92% for many combinations
of risk rating factors) and thus each of the random variables introduced in Sec-
tion 3 had a large discrete component as well as continuous severity component.
For a Bernoulli random variable, it is known that the mean determines the
distribution. Because of the analogy between Bernoulli random variables and
the Section 3 random variables, analysts have historically tended to focus on
the mean as the first step in understanding the distribution.
One cannot make that case for a portfolio of risks. As noted in Section 1,
the portfolio could be selected by geographic region, sales agent type or some
other method of clustering. For illustrative purposes, we have randomly selected
1,000 policies from our 2001 sample. If the concern is with overall losses, we
wish to predict the distribution of S = S1 + … + S1000 . Clearly the predictive
mean provides only one summary measure.
Studying the distribution of S is a well-known problem in probability that
receives substantial attention in actuarial science, see for example, Klugman et
al. (Chapter 6, 2004). The problem is to analyze the convolution of distribution
functions. Unlike textbook treatments, for our application the distribution
functions for S1, …, S1000 are nonidentical, each having a discrete and highly
non-normal continuous component. Thus, although analytic methods are
feasible, we prefer to use simulation methods to compute the predictive dis-
tribution of S. For each policy, using known risk rating factors and estimated
parameters, we simulated the event of an accident and type of loss, as well as
the severity of losses. These components are then used to develop simulated
values of each of the four types of random variables introduced in Section 3.
As in Section 3, we report results based on 5,000 replications. Further details
of the simulation procedure are in Appendix A.3.
Figure 2 summarizes the results for the overall loss, S. This figure shows
that by summing over 1,000 policies, the discrete component is no longer evi-
dent. It is also interesting to see that the portfolio distribution is still long-tail.
Elementary statistics texts, citing the central limit theorem, typically state that
180 E.W. FREES, P. SHI AND E.A. VALDEZ

8e−06
Density

4e− 06
0e+00

0 250000 500000

Portfolio Losses

FIGURE 2: Simulated Predictive Distribution for a Randomly Selected Portfolio

of 1,000 Policies.
4e−05

third party injury

Density

own damage
2e−05

third party property

0e+00

0 200000 400000 600000 800000

Predicted Losses

FIGURE 3: Simulated Density of Losses for Third Party Injury, Own Damage and
Third Party Property of a Randomly Selected Portfolio.

the normal is a good approximation for the distribution of the sum based on 30
to 50 i.i.d. draws. The distribution of the sum shown in Figure 2 is not approx-
imately normal; this is because (1) the policies are not identical, (2) have discrete
and continuous components and (3) have long-tailed continuous components.
As described in Section 3, there may be situations in which the analyst is
interested in the claims from each type of coverage instead of the total claims
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 181

in a comprehensive policy. In this case, one should consider the random variables
Sk = !1000
i =1 Si, k , where k = 1, 2, 3 corresponds to third party injury, own damage
and third party property claims.
As shown in Figure 3, the distributions for the three types of losses are quite
different in terms of their skewness and kurtosis as well as other properties.
The density for third party injury and own damage have higher peaks and are
more positively skewed than third party property. The density of third party
injury has a heavier tail. This type of analysis can provide useful information
about the risk profile of different coverages within a comprehensive policy, as
well as the risk profile of different lines of business. For example, the analyst
may consider which type of coverage should be incorporated and which type
should be eliminated, so that the product can be tuned to meet the company’s
risk management requirement, regulatory policy or other strategic goals.

4.2. Risk Measures

Graphical displays of distributions help analysts understand patterns, linear and
nonlinear; numerical summary measures complement these displays by pro-
viding precise information of selected aspects of the distribution. This section
examines the Value-at-Risk (VaR) and Conditional Tail Expectation (CTE), two
numerical risk measures focusing on the tail of the distribution that have been
widely used in both actuarial and financial work. The VaR is simply a quantile
or percentile; Var(a) gives the 100(1 – a) percentile of the distribution. The CTE(a)
is the expected value conditional on exceeding the Var(a). See, for example,
Hardy (2003) for further background on these and related risk measures.
In addition to the effects of coverage modifications on predictive mean
investigated in Section 3, we are also interested in their effects on the distribu-
tion of losses, S. In this section we focus on VaR and CTE, shown in Tables 7
and 8, respectively. The calculations are based on the randomly selected port-
folio of policies as investigated above. The results in first row of two tables are
corresponding to a policy without deductibles and limits.
Table 7 shows the VaRs at different quantiles and coverage modifications, with
a corresponding 95% confidence interval. The results are consistent with expec-
tations. First, larger deductibles and smaller policy limits decrease the VaR in
a nonlinear way. The marginal effect of the deductible on VaR decreases as the
deductible increases; for example, the VaR difference between deductibles 0 and
250 is larger than the VaR difference between deductibles 250 and 500. Similarly,
the marginal effect of policy limits also decreases as the policy limit increases.
Second, under each combination of deductible and policy limit, the confidence
interval becomes wider as the VaR percentile increases. This result is in part
because of the heavy tailed nature of the losses. Third, policy limits exert a
greater effect than deductibles on the tail of the distribution. This can be seen
by comparing the VaRs in the last three rows in Table 7. The three policy designs
consist of an increasing deductible (which decreases VaR) and an increasing
policy limit (which increases VaR); the overall results show an increasing effect
 TABLE 7
VAR BY PERCENTILE AND COVERAGE MODIFICATION WITH A CORRESPONDING CONFIDENCE INTERVAL
182

Coverage Modification Lower Upper Lower Upper Lower Upper

VaR (90%) VaR (95%) VaR (99%)
Deductible Limit Bound Bound Bound Bound Bound Bound
0 none 258,644 253,016 264,359 324,611 311,796 341,434 763,042 625,029 944,508
250 none 245,105 239,679 250,991 312,305 298,000 329,689 749,814 612,818 929,997
500 none 233,265 227,363 238,797 301,547 284,813 317,886 737,883 601,448 916,310
1,000 none 210,989 206,251 217,216 281,032 263,939 296,124 716,955 581,867 894,080
0 25,000 206,990 205,134 209,000 222,989 220,372 225,454 253,775 250,045 256,666
0 50,000 224,715 222,862 227,128 245,715 243,107 249,331 286,848 282,736 289,953
0 100,000 244,158 241,753 247,653 272,317 267,652 277,673 336,844 326,873 345,324
250 25,000 193,313 191,364 195,381 208,590 206,092 211,389 239,486 235,754 241,836
500 50,000 199,109 196,603 201,513 219,328 216,395 222,725 259,436 255,931 263,516
1,000 100,000 197,534 194,501 201,685 224,145 220,410 229,925 287,555 278,601 297,575

TABLE 8
CTE BY PERCENTILE AND COVERAGE MODIFICATION WITH A CORRESPONDING STANDARD DEVIATION

Coverage Modification Standard Standard Standard

CTE (90%) CTE (95%) CTE (99%)
Deductible Limit Deviation Deviation Deviation
E.W. FREES, P. SHI AND E.A. VALDEZ

0 none 468,850 22,166 652,821 41,182 1,537,692 149,371

250 none 455,700 22,170 639,762 41,188 1,524,650 149,398
500 none 443,634 22,173 627,782 41,191 1,512,635 149,417
1,000 none 422,587 22,180 606,902 41,200 1,491,767 149,457
0 25,000 228,169 808 242,130 983 266,428 1,787
0 50,000 252,564 1,082 270,589 1,388 304,941 2,762
0 100,000 283,270 1,597 309,661 2,091 364,183 3,332
250 25,000 213,974 797 227,742 973 251,820 1,796
500 50,000 225,937 1,066 243,608 1,378 277,883 2,701
1,000 100,000 235,678 1,562 261,431 2,055 315,229 3,239
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 183

on VaR. Fourth, the policy limit exerts a greater effect than a deductible on
the confidence interval capturing the VaR.
Table 8 shows CTEs by percentile and coverage modification with a corre-
sponding standard deviation. The results are consistent with Table 7. Either
a larger deductible or a smaller policy limit results in a lower CTE and the effect
is nonlinear. The range of policy limits explored has a greater influence on
CTE than the range of deductibles. The sparsity of data combined with the
heavy tailed nature of the distribution result in the greater standard deviations
at higher percentiles. Finally, the decrease in the policy limits reduces the stan-
dard deviation of the CTE substantially whereas changes in the deductible
have little influence.

4.3. Dependence
One way to examine the role of dependence is to decompose the comprehensive
coverage into more “primitive” coverages for the three types of claims (third
party injury, own damage and third party property). As in derivative securities,
we call this “unbundling” of coverages. We are able to calculate risk measures
for each unbundled coverage, as if separate financial institutions owned each
coverage, and compare them to risk measures for the bundled coverage that the
insurance company is responsible for. The results are shown in Table 9. For the
bundled (comprehensive) coverage, the VaR and CTE are from the first row of
Tables 7 and 8, respectively, for policies without coverage modifications.
In general, risk measures such as VaR need not be subadditive and so there
is no guarantee that bundling diversifies risk. However, for our application,
our results show that the risk measures for bundled coverages are smaller than
the sum of unbundled coverages, for both risk measures and all percentiles.
One of the important purposes of risk measures is to determine economic cap-
ital which is the amount of capital that banks and insurance companies set
aside as a buffer against potential losses from extreme risk event. The impli-
cation of this example is that by bundling different types of coverage into one
comprehensive policy, the insurers can reduce the economic capital for risk
management or regulatory purpose. Another perspective is that this example
simply demonstrates the effectiveness of economies of scales; three small
financially independent institutions (one for each coverage) require in total
more capital than a single combined institution (one for the bundled coverage).
The interesting thing is that this is true even though the dependencies among
the three coverages are positive, as shown in Appendix A.2.
The statistical model described in Section 2.3 with parameter estimates in
Appendix A.2 show strong significance evidence of positive relations among
the three coverage types, third party injury, own damage and third party prop-
erty. However, the model is complex, using copulas to assess this nonlinear
dependence. How important is the dependence for the financial risk measures?
To quantify this issue, Table 10 shows the VaR and CTE for different copula
models. The independence copula comes from treating the three lines as unrelated,
184 E.W. FREES, P. SHI AND E.A. VALDEZ

TABLE 9
VAR AND CTE BY PERCENTILE FOR UNBUNDLED AND BUNDLED COVERAGES

VaR CTE
Unbundled Coverages
90% 95% 99% 90% 95% 99%

Third party injury 82,080 161,213 557,825 277,101 440,537 1,143,825

Own damage 49,350 59,890 85,403 65,627 77,357 110,334
Third party property 161,903 178,530 214,252 186,679 204,609 257,657
Sum of Unbundled Coverages 293,333 399,633 857,480 529,407 722,503 1,511,816
Bundled (Comprehensive) Coverage 258,644 324,611 763,042 468,850 652,821 1,537,692

the normal copula captures the correlation among different coverages within
a comprehensive policy compared with independence copula, whereas the t-cop-
ula captures the heavy-tail features of the risk compared to the normal copula.
As a sensitivity analysis, we incorporated the copulas in two ways. In the top
portion of the table, we assumed that the specified copula was consistently
used for the estimation and prediction portions. For the bottom portion, we
assumed that the (correct but more complex) t-copula was used for estimation
with the specified copula used for prediction. The idea behind the bottom
portion was that a statistical analysis unit of a company may perform a more
rigorous analysis using a t-copula and another unit within a company may
wish to use this output for quicker calculations about their financial impact.
Table 10 shows that the copula effect is large and increases with the per-
centile. Of course, the upper percentiles are the most important to the actuary
for many financial implications.
Table 10 demonstrates a large difference between assuming independence
among coverages and using a t-copula to quantify the dependence. We found,
when re-estimating the full model under alternative copulas, that the marginal
parameters changed to produce significant differences in the risk measures.
Intuitively, one can think of estimation and prediction under the independence
copula to be similar to “unbundled” coverages in Table 9, where we imagine
separate financial institutions accepting responsibility for each coverage. In one
sense, the results for the independence copula are somewhat counterintuitive.
For most portfolios, with positive correlations among claims, one typically
needs to go out further in the tail to achieve a desired percentile, suggesting
that the VaR should be larger for the t-copula than the independence copula.
To reinforce these findings, the bottom half of the table reports results
when the marginal distributions are unchanged, yet the copula differs. Table 10
shows that the VaR is not affected by the choice of copula; the differences in
Table 10 are due to simulation error. In contrast, for the CTEs, the normal and
t-copula give higher values than the independence copula. This result is due
to the higher losses in the tail under the normal and t-copula models. Although
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 185

TABLE 10
VAR AND CTE FOR BUNDLED COVERAGE BY COPULA

VaR CTE
Copula
90% 95% 99% 90% 95% 99%

EFFECTS OF RE-ESTIMATING THE FULL MODEL

Independence 359,937 490,541 1,377,053 778,744 1,146,709 2,838,762
Normal 282,040 396,463 988,528 639,140 948,404 2,474,151
t 258,644 324,611 763,042 468,850 652,821 1,537,692
EFFECTS OF CHANGING ONLY THE DEPENDENCE STRUCTURE
Independence 259,848 328,852 701,681 445,234 602,035 1,270,212
Normal 257,401 331,696 685,612 461,331 634,433 1,450,816
t 258,644 324,611 763,042 468,850 652,821 1,537,692

not displayed here, we re-ran this portion of the analysis with 50,000 simulation
replications (in lieu of 5,000) and verified these patterns.
When applied to economic capital, these results indicate that the independence
copula leads to more conservative risk measures while the t-copula leads to more
aggressive risk measures. To determine which model to be used to calculate the
economic capital may depend on the purpose of the capital, either for interval
estimation, risk management or regulatory use. It may also depend on the trade-
off among model simplicity, estimation accuracy and computational complexity.

5. PREDICTIVE DISTRIBUTIONS FOR REINSURANCE

Reinsurance, an important risk management tool for property and casualty insur-
ers, is another area of application where predictive distributions can be utilized.
We examine two types of reinsurance agreements: quota share reinsurance and
excess-of-loss reinsurance. In addition, we investigate a simple portfolio rein-
surance. We simulate the number of accident, type of losses and severity of
losses, and then allocate the losses between insurer and reinsurer according to
different reinsurance agreements.
Quota share reinsurance is a form of proportional reinsurance which
specifies that a fixed percentage of each policy written will be transferred to
the reinsurer. The effect of different quota shares on the retained claims for the
ceding company is examined and presented in Figure 4. The distributions of
retained claims are derived assuming the insurer retains 25%, 50%, 75% and
100% of its business. In Figure 4 a quota of 0.25 means the insurer retains 25%
of losses and cedes 75% to the reinsurer. The curve corresponding to quota
of 1 represent the losses of insurer without a reinsurance agreement, as in
Figure 2. As we can see, the quota share reinsurance does not change the shape
186 E.W. FREES, P. SHI AND E.A. VALDEZ

of the retained losses, only the location and scale. For example, if the insurer
ceded 75% of losses, the retained losses will shift left and the variance of
retained losses will be 1/16 times the variance of original losses. We did not
present the distribution of losses for reinsurer, because under quota share rein-
surance, the insurer and reinsurer share the losses proportionally.
Excess-of-loss is a nonproportional reinsurance under which the reinsurer
will pay the ceding company for all the losses above a specified dollar amount,
the retention limit. The retention limit is similar to the deductible in a primary
policy, the reinsurer will assume all the losses above it. Figure 5 shows the
effect of different retention limits on the losses of insurer and reinsurer. Losses
are simulated and limits of 5,000, 10,000 and 20,000 per policy are imposed.
Unlike quota share arrangements, the retention limit changes the shape of the
distribution for both the insurer and reinsurer. The lower the retention limit,
the more business the insurer cedes so that losses for insurer become less skewed
with thinner tails because the losses in the tail of distribution become the
responsibility of reinsurer. Correspondingly, as the retention limit decreases,
the distribution of losses for the reinsurer exhibits fatter tails. This is because
the reinsurer retains a larger portion of the claim.
Figures 4 and 5 provide insights on how the various types of reinsurance
agreement will affect their risk profile of the insurer and reinsurer. Through
such analyses, the insurer can choose proper forms of reinsurance to manage
its risk portfolio, and the reinsurer can decide upon the amount of risk to be
underwritten. In practice, there are many other types of reinsurance contracts.
The above analysis focused on the reinsurance agreements where reim-
bursements are based on losses for each policy. As another example, we
4e−05
3e−05

Quota=0.25
Density

2e−05

Quota=0.5
Quota=0.75
1e−05

Quota=1
0e+00

0 250000 500000

Retained Claims

FIGURE 4: Distribution of Retained Claims for the Insurer under Quota Share Reinsurance.
The insurer retains 25%, 50%, 75% and 100% of losses, respectively.
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 187

4e−05
Density

2e−05 Retention=5,000
Retention=10,000
Retention=20,000
0e+00

0 250000 500000

Retained Claims
4e−05

Retention=20,000

Retention=10,000
Density

2e−05

Retention=5,000
0e+00

0 250000 500000

Ceded Claims

FIGURE 5: Distribution of Losses for the Insurer and Reinsurer under Excess-of-Loss Reinsurance.
The losses are simulated under different primary company retention limits.
The upper panel is for the insurer and lower panel is for the reinsurer.

consider a case where both policy reinsurance and portfolio reinsurance are
involved. Portfolio reinsurance generally refers to the situation where the rein-
surer is insuring all risks in a portfolio of policies, such as a particular line of
business. For the purpose of illustration, we still consider the randomly selected
portfolio which consists of 1000 policies from held-out-sample in year 2001.
 TABLE 11
188

PERCENTILES OF LOSSES FOR INSURER AND REINSURER UNDER REINSURANCE AGREEMENT

PERCENTILE FOR INSURER

Quota Policy Retention Portfolio Retention 1% 5% 10% 25% 50% 75% 90% 95% 99%

0.25 none 100,000 22,518 26,598 29,093 34,196 40,943 50,657 64,819 83,500 100,000
0.5 none 100,000 45,036 53,197 58,187 68,393 81,885 100,000 100,000 100,000 100,000
0.75 none 100,000 67,553 79,795 87,280 100,000 100,000 100,000 100,000 100,000 100,000
1 10,000 100,000 86,083 99,747 100,000 100,000 100,000 100,000 100,000 100,000 100,000
1 10,000 200,000 86,083 99,747 108,345 122,927 140,910 159,449 177,013 188,813 200,000
1 20,000 200,000 89,605 105,578 114,512 132,145 154,858 177,985 200,000 200,000 200,000
0.25 10,000 100,000 21,521 24,937 27,086 30,732 35,228 39,862 44,253 47,203 53,352
0.5 20,000 100,000 44,803 52,789 57,256 66,072 77,429 88,993 100,000 100,000 100,000
0.75 10,000 200,000 64,562 74,810 81,259 92,195 105,683 119,586 132,760 141,610 160,056
1 20,000 200,000 89,605 105,578 114,512 132,145 154,858 177,985 200,000 200,000 200,000

PERCENTILE FOR REINSURER

Quota Policy Retention Portfolio Retention 1% 5% 10% 25% 50% 75% 90% 95% 99%

0.25 none 100,000 67,553 79,795 87,280 102,589 122,828 151,972 194,458 250,499 486,743
E.W. FREES, P. SHI AND E.A. VALDEZ

0.5 none 100,000 45,036 53,197 58,187 68,393 81,885 102,630 159,277 233,998 486,743
0.75 none 100,000 22,518 26,598 29,093 36,785 63,771 102,630 159,277 233,998 486,743
1 10,000 100,000 0 8,066 16,747 36,888 63,781 102,630 159,277 233,998 486,743
1 10,000 200,000 0 0 992 5,878 18,060 43,434 97,587 171,377 426,367
1 20,000 200,000 0 0 0 0 2,482 24,199 78,839 151,321 412,817
0.25 10,000 100,000 68,075 80,695 88,555 104,557 127,652 161,743 215,407 292,216 541,818
0.5 20,000 100,000 45,132 53,298 58,383 68,909 84,474 111,269 167,106 245,101 491,501
0.75 10,000 200,000 23,536 28,055 31,434 39,746 54,268 81,443 135,853 209,406 462,321
1 20,000 200,000 0 0 0 0 2,482 24,199 78,839 151,321 412,817
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 189

We now assume that there is a limit for the entire portfolio of business, the
reinsurer will assume all the losses exceeding the portfolio limit. In addition,
both quota share and policy limit have been applied to the individual policy
in the portfolio. The combined effect of these coverage modifications on the
risk profile for insurer and reinsurer are investigated and the results are pre-
sented in Table 11.
Table 11 presents percentiles of losses for the insurer and reinsurer under
various reinsurance arrangements. As before, we see the long tail nature in the
total losses of the portfolio and the interactive effect of coverage modifications
on the claim distribution. In the first three rows where there is no policy reten-
tion limit, the insurer and reinsurer share the losses proportionally before the
total losses reach the portfolio limit; that is why the 25th percentile for the
insurer is equal to the 75th percentile for the reinsurer. When the total losses
reach portfolio limit, the amount above it should be ceded to reinsurer. The
second three rows for insurer and reinsurer shows the interactive effect of policy
and portfolio retention limits. Both limits reduce the heavy tailed nature of
losses for the insurer. Under a policy retention limit of 10,000, a greater port-
folio retention limit has less effect on changing the tail behavior of losses for
insurer. However the effect of portfolio retention limit (say 200,000) depends
on the policy limit in the sense that the effect will be bigger under a greater
policy limit (20,000). The last four rows of percentiles for both insurer and
reinsurer shows the combined effect of coinsurance, policy limit and portfolio
limit.
Table 11 provides useful information about the distribution of potential
losses for insurer and reinsurer. This can help the ceding company understand
the risk characteristics of retained business. The insurer can choose appropriate
reinsurance agreements including setting proper policy retention limit, select-
ing the right quota and deciding the aggregate stop loss, to manage the risk
more efficiently. The results also can be helpful in determine the reinsurance
premium and to help reinsurer to assess the risk of the business they assumed
from ceding company.

6. SUMMARY AND CONCLUSIONS

This paper follows our prior work (Frees and Valdez (2008)) where a statisti-
cal hierarchical model was introduced using detailed, micro-level automobile
insurance records. In this paper, we demonstrate the financial implications of
the statistical modeling.
We examined three types of applications that commonly concern actuaries.
The first was individual risk rating. We examined the effect of coverage modifi-
cations including deductibles, coverage limits and coinsurance. We showed how
to apply our techniques to rating “unbundled” coverages, very much akin to
financial derivatives. We examined both analytic and simulated means.
Our second type of application dealt with estimating financial risk measures
for portfolios of policies. In this paper we focused on the value at risk, VaR,
190 E.W. FREES, P. SHI AND E.A. VALDEZ

and conditional tail expectation, CTE, although our approach could be easily
extended to other risk measures. We assessed the effects of some of statistical
assumptions, such as the copula, on these measures.
The third type of application was to provide predictive distributions for
reinsurance. We examined the combined effect of coverage modifications for
policies and portfolios on the claim distributions of the insurer and reinsurer,
by examining the tail summary for the losses of portfolio generated using
Monte Carlo simulation.
This paper demonstrates some of the interesting financial analyses of concern
to actuaries that can be accomplished with detailed micro-level data and advanced
statistical models. With simulation, we can easily calculate predictive distrib-
utions for several financial risk measures. A limitation of this paper is that we
did not explicitly incorporate estimation error into our predictive distributions
(see, for example, Cairns (2000)). On one hand, one might argue that we had
many observations available for estimation and that the resulting standard errors
for the statistical model were inherently small. On the other hand, one could say
that that the statistical model incorporates many parameters whose joint uncertainty
should be accounted for. We view this as an interesting area for future research.

ACKNOWLEDGMENTS
The first author thanks the National Science Foundation (Grant Number SES-
0436274) and the Assurant Health Insurance Professorship for funding to sup-
port this research. We also thank Richard Derrig, seminar participants at the
Risk Theory Society, Université de Montreal and the Universitat de Barcelona
as well as anonymous reviewers for comments that helped improve the paper.
The third author wishes to acknowledge hospitality from hosts Louis Doray,
Montserratt Guillen, Carmen Ribas and Oriol Roch.

REFERENCES
ANGERS, J.-F., DESJARDINS, D., DIONNE, G. and GUERTIN, F. (2006) Vehicle and fleet random
effects in a model of insurance rating for fleets of vehicles. ASTIN Bulletin 36(1), 25-77.
ANTONIO, K., BEIRLANT, J., HOEDEMAKERS, T. and VERLAAK, R. (2006) Lognormal mixed models
for reported claims reserves. North American Actuarial Journal 10(1), 30-48.
BOUCHER, J.-Ph. and DENUIT, M. (2006) Fixed versus random effects in Poisson regression models
for claim counts: A case study with motor insurance. ASTIN Bulletin 36(1), 285-301.
BROCKMAN, M.J. and WRIGHT, T.S. (1992) Statistical motor rating: making effective use of your
data. Journal of the Institute of Actuaries 119, 457-543.
CAIRNS, A.J.G. (2000) A discussion of parameter and model uncertainty in insurance. Insurance:
Mathematics and Economics 27(3), 313-330.
COUTTS, S.M. (1984) Motor insurance rating, an actuarial approach. Journal of the Institute of
Actuaries 111, 87-148.
DESJARDINS, D., DIONNE, G. and PINQUET, J. (2001) Experience rating schemes for fleets of vehi-
cles. ASTIN Bulletin 31(1), 81-105.
FRANGOS, N.E. and VRONTOS, S.D. (2001) Design of optimal bonus-malus systems with a frequency
and a severity component on an individual basis in automobile insurance. ASTIN Bulletin 31(1),
1-22.
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 191

FREES, E.W. and VALDEZ, E.A. (1998) Understanding relationships using copulas. North American
Actuarial Journal 2(1), 1-25.
FREES, E.W. and VALDEZ, E.A. (2008) Hierarchical insurance claims modeling. Journal of the
American Statistical Association 103(484), 1457-1469.
GOURIEROUX, Ch. and JASIAK, J. (2007) The Econometrics of Individual Risk: Credit, Insurance,
and Marketing. Princeton University Press, Princeton, NJ.
HARDY, M. (2003) Investment Guarantees: Modeling and Risk Management for Equity-Linked
Life Insurance. John Wiley & Sons, New York.
HSIAO, C., KIM, C. and TAYLOR, G. (1990) A statistical perspective on insurance rate-making.
Journal of Econometrics 44, 5-24.
KAHANE, Y. and HAIM Levy, H. (1975) Regulation in the insurance industry: determination of
premiums in automobile insurance. Journal of Risk and Insurance 42, 117-132.
KLUGMAN, S., PANJER, H. and WILLMOT, G. (2004) Loss Models: From Data to Decisions (Second
Edition), Wiley, New York.
MCDONALD, J.B. and XU, Y.J. (1995) A generalization of the beta distribution with applications.
Journal of Econometrics 66, 133-152.
PINQUET, J. (1997) Allowance for cost of claims in bonus-malus systems. ASTIN Bulletin 27(1),
33-57.
PINQUET, J. (1998) Designing optimal bonus-malus systems from different types of claims. ASTIN
Bulletin 28(2), 205-229.
RENSHAW, A.E. (1994) Modeling the claims process in the presence of covariates. ASTIN Bulletin
24(2), 265-285.
SUN, J., FREES, E.W. and ROSENBERG, M.A. (2008) Heavy-tailed longitudinal data modeling
using copulas. Insurance: Mathematics and Economics, 42(2), 817-830.
TERZA, J.V. and WILSON, P.W. (1990) Analyzing frequencies of several types of events: A mixed
multinomial-Poisson approach. The Review of Economics and Statistics 108-115.
WEISBERG, H.I. and TOMBERLIN, T.J. (1982) A statistical perspective on actuarial methods for
estimating pure premiums from cross-classified data. Journal of Risk and Insurance 49,
539-563.
WEISBERG, H.I., TOMBERLIN, T.J. and CHATTERJEE, S. (1984) Predicting insurance losses under
cross-classification: A comparison of alternative approaches. Journal of Business & Economic
Statistics 2(2), 170-178.

EDWARD W. FREES
School of Business
University of Wisconsin
Madison, Wisconsin 53706 USA
E-Mail: [email protected]

PENG SHI
School of Business
University of Wisconsin
Madison, Wisconsin 53706 USA
E-Mail: [email protected]

EMILIANO A. VALDEZ
Department of Mathematics
College of Liberal Arts and Sciences
University of Connecticut
Storrs, Connecticut 06269-3009 USA
E-Mail: [email protected]
192 E.W. FREES, P. SHI AND E.A. VALDEZ

A. APPENDIX – THE PREDICTIVE MODEL

The summary statistics and parameter estimates corresponding to each com-

ponent of the hierarchical model are provided in Appendix A.1 and A.2, respec-
tively. These results are not directly comparable with Frees and Valdez (2008)
because we selected a different Singaporean company. By examining data from
a different company, we provide further validity of the model’s robustness.
Appendix A.3 describes the simulation procedure, on which the simulations
throughout this work are based.

A.1. SUMMARY STATISTICS

To provide readers with a feel for the data, Table A.1 describes the frequency
of claims, Tables A.2 and A.3 describe the claim frequency relationship with
covariates and Table A.4 displays the distribution by type of claim. Figure 6
gives a density of losses by type.

TABLE A.1
FREQUENCY OF CLAIMS

Count 0 1 2 3 4 Total

Number 118,062 12,092 1,108 77 7 131,346

Percentage 89.89 9.21 0.84 0.06 0.01 100.00

TABLE A.2
NUMBER AND PERCENTAGES OF CLAIMS, BY VEHICLE TYPE AND AGE

Count = 0 Number Percent of Total

VEHICLE TYPE
Automobile 90.06 121,249 92.31
Other 87.75 10,097 7.69
VEHICLE AGE (IN YEARS)
0 93.26 7,330 5.58
1 to 2 89.11 25,621 19.51
3 to 5 89.76 48,964 37.28
6 to 10 89.87 48,226 36.72
11 to 15 92.02 1,103 0.84
16 and older 87.25 102 0.08
Total Number 131,346 100
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 193

TABLE A.3
NUMBER AND PERCENTAGES OF GENDER, AGE AND NCD FOR AUTOMOBILE POLICIES

Count = 0 Number Percent of Total

GENDER
Female 90.77 23,783 19.62
Male 89.89 97,466 80.38
PERSON AGE (IN YEARS)
21 and yonger 85.19 27 0.02
22-25 87.24 948 0.78
26-35 89.28 29,060 23.97
36-45 90.21 44,494 36.7
46-55 90.34 30,737 25.35
56-65 90.76 13,209 10.89
66 and over 90.56 2,774 2.29
Total Number 121,249 100

TABLE A.4
DISTRIBUTION OF CLAIMS, BY CLAIM TYPE OBSERVED

Value of M 1 2 3 4 5 6 7 Total
Claim Type (y1) (y2) (y3) (y1, y2) (y1, y3) (y2, y3) (y1, y2, y3 )
Number 160 9,928 1,660 184 30 2,513 92 14,567
Percentage 1.1 68.15 11.4 1.26 0.21 17.25 0.63 100
0.30

third party injury

own damage
third party property
0.20
Density

0.10
0.00

0 10 20 30 40 50

FIGURE 6: Density by Loss Type. Claims are in thousands of dollars.

194 E.W. FREES, P. SHI AND E.A. VALDEZ

A.2. PARAMETER ESTIMATES OF HIERARCHICAL MODEL

The parameter estimates for hierarchical predictive model are presented in

Tables A.5, A.6 and A.8 for the frequency, type and severity components,
respectively. The risk factors are based on those described in Table 1. For exam-
ple, “vehicle age 3-5” in Table A.5 means that the insured’s vehicle is 3, 4 or 5
years old, while the “vehicle age % 5” in Table A.8 means that the age is less
than or equal to 5. A “*” signifies the interaction between the two risk factors.
For example, “automobile * NCD 10” in Table A.5 indicates an insured who
owns a private car with NCD equal to 10, and “automobile * NCD 0-10” in
Table A.8 refers to a policyholder who has a private car with NCD equal to 0
or 10.
In Table A.6 we did not provide the standard errors to show the statistical
significance of each parameter. Instead, we provide chi-square tests in Table A.7.
In the multilogit model, one is concerned whether a covariate significantly differ-
entiates the claim type probability across each category of M, not the significance
of parameters within each category. In Table A.7, “vehicle age” is divided
into two categories, less than 5 and equal or greater than 6. “automobile*age”
represents three categories, non private car, private car with insured’s age less
than 45, private car with insured’s age equal or greater than 46. Table A.7
shows that all variables are statistically significant.

TABLE A.5
FITTED NEGATIVE BINOMIAL MODEL

Parameter Estimate StdError Parameter Estimate StdError

intercept –2.275 0.730 automobile*NCD 0 0.748 0.027

year 0.043 0.004 automobile*NCD 10 0.640 0.032
automobile –1.635 0.082 automobile*NCD 20 0.585 0.029
vehicle age 0 0.273 0.739 automobile*NCD 30 0.563 0.030
vehicle age 1-2 0.670 0.732 automobile*NCD 40 0.482 0.032
vehicle age 3-5 0.482 0.732 automobile*NCD 50 0.347 0.021
vehicle age 6-10 0.223 0.732 automobile*age %21 0.955 0.431
vehicle age 11-15 0.084 0.772 automobile*age 22-25 0.843 0.105
automobile*vehicle age 0 0.613 0.167 automobile*age 26-35 0.657 0.070
automobile*vehicle age 1-2 0.258 0.139 automobile*age 36-45 0.546 0.070
automobile*vehicle age 3-5 0.386 0.138 automobile*age 46-55 0.497 0.071
automobile*vehicle age 6-10 0.608 0.138 automobile*age 56-65 0.427 0.073
automobile*vehicle age 11-15 0.569 0.265 automobile*age &66 0.438 0.087
automobile*vehicle age &16 0.930 0.677 automobile*male –0.252 0.042
vehicle capacity 0.116 0.018 automobile*female –0.383 0.043
r 2.167 0.195
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 195

TABLE A.6
FITTED MULTI LOGIT MODEL

Parameter Estimates

Category(M) intercept year vehicle age &6 non-automobile automobile*age &46

1 1.194 –0.142 0.084 0.262 0.128

2 4.707 –0.024 –0.024 – 0.153 0.082
3 3.281 –0.036 0.252 0.716 – 0.201
4 1.052 –0.129 0.037 – 0.349 0.338
5 –1.628 0.132 0.132 – 0.008 0.330
6 3.551 –0.089 0.032 – 0.259 0.203

TABLE A.7
MAXIMUM LIKELIHOOD ANALYSIS OF VARIANCE

Source DF Chi-Square Pr > ChiSq

intercept 6 2225.37 <.0001

year 6 59.80 <.0001
vehicle age 6 101.72 <.0001
automobile*age 12 444.04 <.0001
Likelihood Ratio 258 268.30 0.3168

TABLE A.8
FITTED SEVERITY MODEL BY COPULAS

Types of Copula

Parameter Independence Normal Copula t-Copula

Standard Standard Standard

Estimate Estimate Estimate
Error Error Error

THIRD PARTY INJURY

s1 0.225 0.020 0.224 0.044 0.232 0.079
a11 69.958 28.772 69.944 63.267 69.772 105.245
a21 392.362 145.055 392.372 129.664 392.496 204.730
intercept 34.269 8.144 34.094 7.883 31.915 5.606
196 E.W. FREES, P. SHI AND E.A. VALDEZ

Types of Copula

Parameter Independence Normal Copula t-Copula

Standard Standard Standard

Estimate Estimate Estimate
Error Error Error

OWN DAMAGE
s2 0.671 0.007 0.670 0.002 0.660 0.004
a12 5.570 0.151 5.541 0.144 5.758 0.103
a22 12.383 0.628 12.555 0.277 13.933 0.750
intercept 1.987 0.115 2.005 0.094 2.183 0.112
year –0.016 0.006 –0.015 0.006 –0.013 0.006
vehicle capacity 0.116 0.031 0.129 0.022 0.144 0.012
vehicle age %5 0.107 0.034 0.106 0.031 0.107 0.003
automobile*NCD 0-10 0.102 0.029 0.099 0.039 0.087 0.031
automobile*age 26-55 –0.047 0.027 –0.042 0.044 –0.037 0.005
automobile*age &56 0.101 0.050 0.080 0.018 0.084 0.050
THIRD PARTY PROPERTY
s3 1.320 0.068 1.309 0.066 1.349 0.068
a13 0.677 0.088 0.615 0.080 0.617 0.079
a23 1.383 0.253 1.528 0.271 1.324 0.217
intercept 1.071 0.134 1.035 0.132 0.841 0.120
vehicle age 1-10 –0.008 0.098 –0.054 0.094 –0.036 0.092
vehicle age &11 –0.022 0.198 0.030 0.194 0.078 0.193
year 0.031 0.007 0.043 0.007 0.046 0.007
COPULA
r12 – – 0.250 0.049 0.241 0.054
r13 – – 0.163 0.063 0.169 0.074
r23 – – 0.310 0.017 0.330 0.019
n – – – – 6.013 0.688

A.3. SIMULATION PROCEDURE

The simulation procedure used in this paper can be described in terms the three
component hierarchical model.
We start the simulation for the frequency component of the hierarchical pre-
dictive model. The number of accidents Ni for policyholder i follows negative
binomial distribution described in Section 2.3.1. We generate Ni using the prob-
abilities Pr(Ni = k) = a k r+-r 1- 1k pir (1 – pi ) k.
Then we generate the type of losses given an accident by simulating claim
type variable Mi for each policyholder from the distribution Pr(Mi = m) =
exp ]Vm g
7 , which is described in Section 2.3.2.
!s = 1exp ]Vs g
 APPLICATIONS OF A HIERARCHICAL INSURANCE CLAIMS MODEL 197

Finally, we simulate the trivariate GB2 distribution, as follows:

• Generate (t1, t2, t3 ) from a trivariate t-distribution using t = s / W , where s
has a multivariate normal distribution with variance-covariance matrix S
and W, independent of s, follows a chi-square distribution with n degrees of
freedom.
• Generate (u1, u2, u3) from the t-copula using uk = Gn (tk ), k = 1, 2, 3 where Gn
is the distribution function for a t-distribution with n degrees of freedom.
• Calculate qk, the ukth percentile of a beta distribution with parameters a1k
and a2k, k = 1, 2, 3. Here, a1k and a2k represent the shape parameters for a
type k loss.
• Generate a realization of the trivariate GB2 distribution using

zk = exp( mk) (qk / (1 – qk ))sk, k = 1, 2, 3

where mk is the location parameter for a type k loss, defined by equation (6),
and sk represents the scale parameter for a type k loss.

After generating the three components of the predictive model, we can calculate
the simulated losses of the three type from j th accident for policy holder i:

yij1 = zij1 (1(Mi = 1) + 1(Mi = 1) + 1(Mi = 5) + 1(Mi = 7))

yij2 = zij2 (1(Mi = 2) + 1(Mi = 1) + 1(Mi = 6) + 1(Mi = 7))
yij3 = zij3 (1(Mi = 3) + 1(Mi = 1) + 1(Mi = 6) + 1(Mi = 7))

Incorporating the number of accidents, we have the claims of each type from
a single policy:
3
Si, k = yijk ! 1 (Ni > n), k = 1, 2, 3
n=0

and the total losses from a policy:

Si = Si,1 + Si,2 + Si,3

Also we can calculate the losses from a portfolio of policies:

m
S = ! Si
i =1

where m represent the size of this portfolio.