Credibility Theory Features of Actuar

Credibility theory features of actuar
Christophe Dutang
Université Paris Dauphine
Vincent Goulet
Université Laval
Xavier Milhaud
Université Claude Bernard Lyon 1
Mathieu Pigeon
Université du Québec à Montréal
1 Introduction
Credibility models are actuarial tools to distribute premiums fairly among a
heterogeneous group of policyholders (henceforth called entities). More gen-
erally, they can be seen as prediction methods applicable in any setting where
repeated measures are made for subjects with different risk levels.
The credibility theory features of actuar consist of matrix hachemeister
containing the famous data set of Hachemeister (1975) and function cm to fit
hierarchical (including Bühlmann, Bühlmann-Straub), regression and linear
Bayes credibility models. Furthermore, function rcomphierarc can simulate
portfolios of data satisfying the assumptions of the aforementioned credibility
models; see the "simulation" vignette for details.
2 Hachemeister data set

The data set of Hachemeister (1975) consists of private passenger bodily injury
insurance average claim amounts, and the corresponding number of claims,
for five U.S. states over 12 quarters between July 1970 and June 1973. The
data set is included in the package in the form of a matrix with 5 rows and
25 columns. The first column contains a state index, columns 2–13 contain the
claim averages and columns 14–25 contain the claim numbers:
1
> data(hachemeister)
> hachemeister
state ratio.1 ratio.2 ratio.3 ratio.4 ratio.5
[1,] 1 1738 1642 1794 2051 2079
[2,] 2 1364 1408 1597 1444 1342
[3,] 3 1759 1685 1479 1763 1674
[4,] 4 1223 1146 1010 1257 1426
[5,] 5 1456 1499 1609 1741 1482
ratio.6 ratio.7 ratio.8 ratio.9 ratio.10 ratio.11
[1,] 2234 2032 2035 2115 2262 2267
[2,] 1675 1470 1448 1464 1831 1612
[3,] 2103 1502 1622 1828 2155 2233
[4,] 1532 1953 1123 1343 1243 1762
[5,] 1572 1606 1735 1607 1573 1613
ratio.12 weight.1 weight.2 weight.3 weight.4
[1,] 2517 7861 9251 8706 8575
[2,] 1471 1622 1742 1523 1515
[3,] 2059 1147 1357 1329 1204
[4,] 1306 407 396 348 341
[5,] 1690 2902 3172 3046 3068
weight.5 weight.6 weight.7 weight.8 weight.9
[1,] 7917 8263 9456 8003 7365
[2,] 1622 1602 1964 1515 1527
[3,] 998 1077 1277 1218 896
[4,] 315 328 352 331 287
[5,] 2693 2910 3275 2697 2663
weight.10 weight.11 weight.12
[1,] 7832 7849 9077
[2,] 1748 1654 1861
[3,] 1003 1108 1121
[4,] 384 321 342
[5,] 3017 3242 3425
3 Hierarchical credibility model

The linear model fitting function of R is lm. Since credibility models are very
close in many respects to linear models, and since the credibility model fit-
ting function of actuar borrows much of its interface from lm, we named the
credibility function cm.
Function cm acts as a unified interface for all credibility models supported
by the package. Currently, these are: the unidimensional models of Bühlmann
(1969) and Bühlmann and Straub (1970); the hierarchical model of Jewell
(1975) (of which the first two are special cases); the regression model of
Hachemeister (1975), optionally with the intercept at the barycenter of time
2
(Bühlmann and Gisler, 2005, Section 8.4); linear Bayes models. The modular
design of cm makes it easy to add new models if desired.
This section concentrates on usage of cm for hierarchical models.
There are some variations in the formulas of the hierarchical model in the
literature. We compute the credibility premiums as given in Bühlmann and
Jewell (1987) or Bühlmann and Gisler (2005), supporting three types of esti-
mators of the between variance structure parameters: the unbiased estimators
of Bühlmann and Gisler (2005) (the default), the slightly different version of
Ohlsson (2005) and the iterative pseudo-estimators as found in Goovaerts and
Hoogstad (1987) or Goulet (1998).
Consider an insurance portfolio where entities are classified into cohorts. In
our terminology, this is a two-level hierarchical classification structure. The
observations are claim amounts Sijt , where index i = 1, . . . , I identifies the
cohort, index j = 1, . . . , Ji identifies the entity within the cohort and index t =
1, . . . , nij identifies the period (usually a year). To each data point corresponds
a weight — or volume — wijt . Then, the best linear prediction for the next
period outcome of a entity based on ratios Xijt = Sijt /wijt is
π̂ij = zij Xijw + (1 − zij )π̂i

(1)
π̂i = zi Xizw + (1 − zi )m,
with the credibility factors

nij
wijΣ
zij =
wijΣ + s2 /a
, wijΣ = ∑ wijt
t =1
Ji
zi Σ
zi =
ziΣ + a/b
, zi Σ = ∑ zij
j =1
and the weighted averages

nij
wijt
Xijw = ∑ wijΣ Xijt
t =1
Ji zij
Xizw = ∑ ziΣ Xijw .
j =1
The estimator of s2 is
I Ji nij
1
2
ŝ = J ∑ ∑ ∑ wijt (Xijt − Xijw )2 . (2)
∑iI=1 ∑ j=
i
1 ( nij − 1) i =1 j =1 t =1
The three types of estimators for the variance components a and b are the
3
following. First, let
Ji Ji wij2 Σ
Ai = ∑ wijΣ (Xijw − Xiww )2 − ( Ji − 1)s2 ci = wiΣΣ − ∑ wiΣΣ
j =1 j =1
I Iz2iΣ
B= ∑ ziΣ (Xizw − X̄zzw )2 − ( I − 1)a d = zΣΣ − ∑
z
,
i =1 i =1 ΣΣ
with
I
z
X̄zzw = ∑ zΣΣ
iΣ
Xizw . (3)
i =1
(Hence, E[ Ai ] = ci a and E[ B] = db.) Then, the Bühlmann–Gisler estimators

are
1 I

Ai
â = ∑ max ,0 (4)
I i =1 ci

B
b̂ = max ,0 , (5)
d
the Ohlsson estimators are
∑iI=1 Ai
â0 = (6)
∑iI=1 ci
B
b̂0 = (7)
d
and the iterative (pseudo-)estimators are
I Ji
1
ã = I
∑i=1 ( Ji − 1)
∑ ∑ zij (Xijw − Xizw )2 (8)
i =1 j =1
I
1
b̃ = ∑ z ( X − Xzzw )2 ,
I − 1 i=1 i izw
(9)
where
I
z
Xzzw = ∑ zΣi Xizw . (10)
i =1
Note the difference between the two weighted averages (3) and (10). See Bel-
hadj et al. (2009) for further discussion on this topic.
Finally, the estimator of the collective mean m is m̂ = Xzzw .
The credibility modeling function cm assumes that data is available in the
format most practical applications would use, namely a rectangular array (ma-
trix or data frame) with entity observations in the rows and with one or more
classification index columns (numeric or character). One will recognize the
output format of rcomphierarc and its summary methods.
4
Then, function cm works much the same as lm. It takes in argument: a
formula of the form ˜ terms describing the hierarchical interactions in a data
set; the data set containing the variables referenced in the formula; the names
of the columns where the ratios and the weights are to be found in the data
set. The latter should contain at least two nodes in each level and more than
one period of experience for at least one entity. Missing values are represented
by NAs. There can be entities with no experience (complete lines of NAs).
In order to give an easily reproducible example, we group states 1 and 3 of
the Hachemeister data set into one cohort and states 2, 4 and 5 into another.
This shows that data does not have to be sorted by level. The fitted model
below uses the iterative estimators of the variance components.
> X <- cbind(cohort = c(1, 2, 1, 2, 2), hachemeister)
> fit <- cm(~cohort + cohort:state, data = X,
+ ratios = ratio.1:ratio.12,
+ weights = weight.1:weight.12,
+ method = "iterative")
> fit
Call:
cm(formula = ~cohort + cohort:state, data = X, ratios = ratio.1:ratio.12,
weights = weight.1:weight.12, method = "iterative")
Structure Parameters Estimators
Collective premium: 1746
Between cohort variance: 88981

Within cohort/Between state variance: 10952
Within state variance: 139120026
The function returns a fitted model object of class "cm" containing the
estimators of the structure parameters. To compute the credibility premiums,
one calls a method of predict for this class.
> predict(fit)
$cohort
[1] 1949 1543
$state
[1] 2048 1524 1875 1497 1585
One can also obtain a nicely formatted view of the most important results
with a call to summary.
> summary(fit)
Call:
5

Within cohort/Between state variance: 10952
Detailed premiums
Level: cohort
cohort Indiv. mean Weight Cred. factor Cred. premium
1 1967 1.407 0.9196 1949
2 1528 1.596 0.9284 1543
Level: state
cohort state Indiv. mean Weight Cred. factor
1 1 2061 100155 0.8874
2 2 1511 19895 0.6103
1 3 1806 13735 0.5195
2 4 1353 4152 0.2463
2 5 1600 36110 0.7398
Cred. premium
2048
1524
1875
1497
1585
The methods of predict and summary can both report for a subset of the
levels by means of an argument levels.
> summary(fit, levels = "cohort")
Call:

Within cohort variance: 10952
Detailed premiums
6
cohort Indiv. mean Weight Cred. factor Cred. premium
1 1967 1.407 0.9196 1949
2 1528 1.596 0.9284 1543
> predict(fit, levels = "cohort")
$cohort
[1] 1949 1543
4 Bühlmann and Bühlmann–Straub models

As mentioned above, the Bühlmann and Bühlmann–Straub models are simply
one-level hierarchical models. In this case, the Bühlmann–Gisler and Ohlsson
estimators of the between variance parameters are both identical to the usual
Bühlmann and Straub (1970) estimator
!
I
wΣΣ
â = 2 ∑ wiΣ (Xiw − Xww ) − ( I − 1)ŝ ,
wΣΣ − ∑iI=1 wi2Σ i=1
2 2
(11)
and the iterative estimator

I
1
ã = ∑ z ( X − Xzw )2
I − 1 i=1 i iw
(12)
is better known as the Bichsel–Straub estimator.

To fit the Bühlmann model using cm, one simply does not specify any
weights.
> cm(~state, hachemeister, ratios = ratio.1:ratio.12)
Call:
cm(formula = ~state, data = hachemeister, ratios = ratio.1:ratio.12)
Between state variance: 72310

When weights are specified together with a one-level model, cm automat-
ically fits the Bühlmann–Straub model to the data. In the example below, we
use the Bichsel–Straub estimator for the between variance.
> cm(~state, hachemeister, ratios = ratio.1:ratio.12,
+ weights = weight.1:weight.12)
Call:
cm(formula = ~state, data = hachemeister, ratios = ratio.1:ratio.12,
weights = weight.1:weight.12)
7
Between state variance: 89639

5 Regression model of Hachemeister

The credibility regression model of Hachemeister (1975) is a generalization
of the Bühlmann–Straub model. If data shows a systematic trend, the latter
model will typically under- or over-estimate the true premium of an entity.
The idea of Hachemeister was to fit to the data a regression model where
the parameters are a credibility weighted average of an entity’s regression
parameters and the group’s parameters.
In order to use cm to fit a credibility regression model to a data set, one
simply has to supply as additional arguments regformula and regdata. The
first one is a formula of the form ˜ terms describing the regression model,
and the second is a data frame of regressors. That is, arguments regformula
and regdata are in every respect equivalent to arguments formula and data
of lm, with the minor difference that regformula does not need to have a left
hand side (and is ignored if present). Below, we fit the model
Xit = β 0 + β 1 t + ε t , t = 1, . . . , 12
to the original data set of Hachemeister (1975).

> fit <- cm(~state, hachemeister, regformula = ~ time,
+ regdata = data.frame(time = 1:12),
> fit
Call:
weights = weight.1:weight.12, regformula = ~time, regdata = data.frame(time = 1:12))
Collective premium: 1469 32.05
Between state variance: 24154 2700.0

2700 301.8
To compute the credibility premiums, one has to provide the “future” val-
ues of the regressors as in predict.lm.
8
2000
1800
1600
1400
1200
collective
individual
1000
credibility
2 4 6 8 10 12
Figure 1: Collective, individual and credibility regression lines for State 4 of

the Hachemeister data set. The point indicates the credibility premium.
> predict(fit, newdata = data.frame(time = 13))

[1] 2437 1651 2073 1507 1759
It is well known that the basic regression model has a major drawback:
there is no guarantee that the credibility regression line will lie between the
collective and individual ones. This may lead to grossly inadequate premi-
ums, as Figure 1 shows.
The solution proposed by Bühlmann and Gisler (1997) is simply to po-
sition the intercept not at time origin, but instead at the barycenter of time
(see also Bühlmann and Gisler, 2005, Section 8.4). In mathematical terms, this
essentially amounts to using an orthogonal design matrix. By setting the argu-
ment adj.intercept to TRUE in the call, cm will automatically fit the credibility
regression model with the intercept at the barycenter of time. The resulting
regression coefficients have little meaning, but the predictions are sensible.
> fit2 <- cm(~state, hachemeister, regformula = ~ time,
+ regdata = data.frame(time = 1:12),
+ adj.intercept = TRUE,
9
> summary(fit2, newdata = data.frame(time = 13))
Call:
weights = weight.1:weight.12, regformula = ~time, regdata = data.frame(time = 1:12),
adj.intercept = TRUE)
Collective premium: -1675 117.1
Between state variance: 93783 0

0 8046
Detailed premiums
state Indiv. coef. Cred. matrix Adj. coef.

1 -2062.46 0.9947 0.0000 -2060.41
216.97 0.0000 0.9413 211.10
2 -1509.28 0.9740 0.0000 -1513.59
59.60 0.0000 0.7630 73.23
3 -1813.41 0.9627 0.0000 -1808.25
150.60 0.0000 0.6885 140.16
4 -1356.75 0.8865 0.0000 -1392.88
96.70 0.0000 0.4080 108.77
5 -1598.79 0.9855 0.0000 -1599.89
41.29 0.0000 0.8559 52.22
Cred. premium
2457
1651
2071
1597
1698
Figure 2 shows the beneficient effect of the intercept adjustment on the
premium of State 4.
10
2000
1800
1600
1400
1200
collective
individual
1000
credibility
2 4 6 8 10 12
Figure 2: Collective, individual and credibility regression lines for State 4 of

the Hachemeister data set when the intercept is positioned at the barycenter
of time. The point indicates the credibility premium.
6 Linear Bayes model

In the pure bayesian approach to the ratemaking problem, we assume that
the observations Xt , t = 1, . . . , n, of an entity depend on its risk level θ, and
that this risk level is a realization of an unobservable random variable Θ. The
best (in the mean square sense) approximation to the unknown risk premium
µ(θ ) = E[ Xt |Θ = θ ] based on observations X1 , . . . , Xn is the Bayesian premium
Bn+1 = E[µ(Θ)| X1 , . . . , Xn ].
It is then well known (Bühlmann and Gisler, 2005; Klugman et al., 2012) that
for some combinaisons of distributions, the Bayesian premium is linear and
can written as a credibility premium
Bn+1 = z X̄ + (1 − z)m,
where m = E[µ(Θ)] and z = n/(n + K ) for some constant K.

The combinations of distributions yielding a linear Bayes premium involve
11
members of the univariate exponential family for the distribution of X |Θ = θ
and their natural conjugate for the distribution of Θ:
• X |Θ = θ ∼ Poisson(θ ), Θ ∼ Gamma(α, λ);
• X |Θ = θ ∼ Exponential(θ ), Θ ∼ Gamma(α, λ);
• X |Θ = θ ∼ Normal(θ, σ22 ), Θ ∼ Normal(µ, σ12 );
• X |Θ = θ ∼ Bernoulli(θ ), Θ ∼ Beta( a, b);
• X |Θ = θ ∼ Geometric(θ ), Θ ∼ Beta( a, b);
and the convolutions
• X |Θ = θ ∼ Gamma(τ, θ ), Θ ∼ Gamma(α, λ);
• X |Θ = θ ∼ Binomial(ν, θ ), Θ ∼ Beta( a, b);
• X |Θ = θ ∼ Negative Binomial(r, θ ) and Θ ∼ Beta( a, b).
Appendix A provides the complete formulas for the above combinations of
distributions.
In addition, Bühlmann and Gisler (2005, section 2.6) show that if X |Θ =
θ ∼ Single Parameter Pareto(θ, x0 ) and Θ ∼ Gamma(α, λ), then the Bayesian
estimator of parameter θ — not of the risk premium! — is
α
Θ̂ = η θ̂ MLE + (1 − η ) ,
λ
where
n
θ̂ MLE =
∑in=1 ln( Xi /x0 )
is the maximum likelihood estimator of θ and
∑in=1 ln( Xi /x0 )
η=
λ + ∑in=1 ln( Xi /x0 )
is a weight not restricted to (0, 1). (See the "distributions" package vignette
for details on the Single Parameter Pareto distribution.)
When argument formula is "bayes", function cm computes pure Bayesian
premiums — or estimator in the Pareto/Gamma case — for the combinations
of distributions above. We identify which by means of argument likelihood
that must be one of "poisson", "exponential", "gamma", "normal", "bernoulli",
"binomial", "geometric", "negative binomial" or "pareto". The parameters
of the distribution of X |Θ = θ, if any, and those of the distribution of Θ are
specified using the argument names (and default values) of dgamma, dnorm,
dbeta, dbinom, dnbinom or dpareto1, as appropriate.
Consider the case where
X |Θ = θ ∼ Poisson(θ )
Θ ∼ Gamma(α, λ).
12
The posterior distribution of Θ is
!
n
Θ| X1 , . . . , Xn ∼ Gamma α + ∑ Xt , λ + n .
t =1
Therefore, the Bayesian premium is
Bn+1 = E[µ(Θ)| X1 , . . . , Xn ]
= E [ Θ | X1 , . . . , X n ]
α + ∑nt=1 Xt
=
λ+n
n λ α
= X̄ +
n+λ n+λλ
= z X̄ + (1 − z)m,
with m = E[µ(Θ)] = E[Θ] = α/λ and

n
z= , K = λ.
n+K
One may easily check that if α = λ = 3 and X1 = 5, X2 = 3, X3 = 0, X4 =
1, X5 = 1, then B6 = 1.625. We obtain the same result using cm.
> x <- c(5, 3, 0, 1, 1)
> fit <- cm("bayes", x, likelihood = "poisson",
+ shape = 3, rate = 3)
> fit
Call:
cm(formula = "bayes", data = x, likelihood = "poisson", shape = 3,
rate = 3)
Between variance: 0.3333

Within variance: 1
> predict(fit)
[1] 1.625
> summary(fit)
Call:
cm(formula = "bayes", data = x, likelihood = "poisson", shape = 3,
rate = 3)
13
Between variance: 0.3333

Within variance: 1
Detailed premiums
Indiv. mean Weight Cred. factor Bayes premium

2 5 0.625 1.625
A Linear Bayes formulas

This appendix provides the main linear Bayes credibility results for combi-
nations of a likelihood function member of the univariate exponential family
with its natural conjugate. For each combination, we provide, other than the
names of the distributions of X |Θ = θ and Θ:
• the posterior distribution Θ| X1 = x1 , . . . , Xn = xn , always of the same
type as the prior, only with updated parameters;
• the risk premium µ(θ ) = E[ X |Θ = θ ];
• the collective premium m = E[µ(Θ)];

• the Bayesian premium Bn+1 = E[µ(Θ)| X1 , . . . , Xn ], always equal to the
collective premium evaluated at the parameters of the posterior distri-
bution;
• the credibility factor when the Bayesian premium is expressed as a cred-

ibility premium.
A.1 Bernoulli/beta case

X |Θ = θ ∼ Bernoulli(θ )
Θ ∼ Beta( a, b)
Θ| X1 = x1 , . . . , Xn = xn ∼ Beta( ã, b̃)
n
ã = a + ∑ xt
t =1
n
b̃ = b + n − ∑ xt
t =1
Risk premium
µ(θ ) = θ
14
Collective premium
a
m=
a+b
Bayesian premium
a + ∑nt=1 Xt
Bn+1 =
a+b+n
Credibility factor
n
z=
n+a+b
A.2 Binomial/beta case

X |Θ = θ ∼ Binomial(ν, θ )
Θ ∼ Beta( a, b)
Θ| X1 = x1 , . . . , Xn = xn ∼ Beta( ã, b̃)
n
ã = a + ∑ xt
t =1
n
b̃ = b + nν − ∑ xt
t =1
Risk premium
µ(θ ) = νθ
Collective premium
νa
m=
a+b
Bayesian premium
ν( a + ∑nt=1 Xt )
Bn+1 =
a + b + nν
Credibility factor
n
z=
n + ( a + b)/ν
A.3 Geometric/Beta case

X |Θ = θ ∼ Geometric(θ )
Θ ∼ Beta( a, b)
Θ| X1 = x1 , . . . , Xn = xn ∼ Beta( ã, b̃)
ã = a + n
n
b̃ = b + ∑ xt
t =1
Risk premium
1−θ
µ(θ ) =
θ
15
Collective premium
b
m=
a−1
Bayesian premium
b + ∑nt=1 Xt
Bn+1 =
a+n−1
Credibility factor
n
z=
n+a−1
A.4 Negative binomial/Beta case

X |Θ = θ ∼ Negative binomial(r, θ )
Θ ∼ Beta( a, b)
Θ| X1 = x1 , . . . , Xn = xn ∼ Beta( ã, b̃)
ã = a + nr
n
b̃ = b + ∑ xt
t =1
Risk premium
r (1 − θ )
µ(θ ) =
θ
Collective premium
rb
m=
a−1
Bayesian premium
r (b + ∑nt=1 Xt )
Bn+1 =
a + nr − 1
Credibility factor
n
z=
n + ( a − 1)/r
A.5 Poisson/Gamma case

X |Θ = θ ∼ Poisson(θ )
Θ ∼ Gamma(α, λ)
Θ| X1 = x1 , . . . , Xn = xn ∼ Gamma(α̃, λ̃)
n
α̃ = α + ∑ xt
t =1
λ̃ = λ + n
Risk premium
µ(θ ) = θ
16
Collective premium
α
m=
λ
Bayesian premium
α + ∑nt=1 Xt
Bn+1 =
λ+n
Credibility factor
n
z=
n+λ
A.6 Exponential/Gamma case

X |Θ = θ ∼ Exponential(θ )
Θ| X1 = x1 , . . . , Xn = xn ∼ Gamma(α̃, λ̃)
α̃ = α + n
n
λ̃ = λ + ∑ xt
t =1
Risk premium
1
µ(θ ) =
θ
Collective premium
λ
m=
α−1
Bayesian premium
λ + ∑nt=1 Xt
Bn+1 =
α+n−1
Credibility factor
n
z=
n+α−1
A.7 Gamma/Gamma case

X |Θ = θ ∼ Gamma(τ, θ )
Θ| X1 = x1 , . . . , Xn = xn ∼ Gamma(α̃, λ̃)
α̃ = α + nτ
n
λ̃ = λ + ∑ xt
t =1
Risk premium
τ
µ(θ ) =
θ
17
Collective premium
τλ
m=
α−1
Bayesian premium
τ (λ + ∑nt=1 Xt )
Bn+1 =
α + nτ − 1
Credibility factor
n
z=
n + (α − 1)/τ
A.8 Normal/Normal case

X |Θ = θ ∼ Normal(θ, σ22 )
Θ ∼ Normal(µ, σ12 )
Θ| X1 = x1 , . . . , Xn = xn ∼ Normal(µ̃, σ̃12 )
σ12 ∑nt=1 xt + σ22 µ

µ̃ =
nσ12 + σ22
σ12 σ22
σ̃12 =
nσ12 + σ22
Risk premium
µ(θ ) = θ
Collective premium
m=µ
Bayesian premium
σ12 ∑nt=1 Xt + σ22 µ
Bn+1 =
nσ12 + σ22
Credibility factor
n
z=
n + σ22 /σ12
References
H. Belhadj, V. Goulet, and T. Ouellet. On parameter estimation in hierarchical
credibility. ASTIN Bulletin, 39(2), 2009.
H. Bühlmann. Experience rating and credibility. ASTIN Bulletin, 5:157–165,
1969.
H. Bühlmann and A. Gisler. Credibility in the regression case revisited. ASTIN
Bulletin, 27:83–98, 1997.
18
H. Bühlmann and A. Gisler. A course in credibility theory and its applications.
Springer, 2005. ISBN 3-5402575-3-5.
H. Bühlmann and W. S. Jewell. Hierarchical credibility revisited. Bulletin of
the Swiss Association of Actuaries, 87:35–54, 1987.
H. Bühlmann and E. Straub. Glaubgwürdigkeit für Schadensätze. Bulletin of

the Swiss Association of Actuaries, 70:111–133, 1970.
M. J. Goovaerts and W. J. Hoogstad. Credibility theory. Number 4 in Surveys
of actuarial studies. Nationale-Nederlanden N.V., Netherlands, 1987.
V. Goulet. Principles and application of credibility theory. Journal of Actuarial
Practice, 6:5–62, 1998.
C. A. Hachemeister. Credibility for regression models with application to
trend. In Credibility, theory and applications, Proceedings of the berkeley Ac-
tuarial Research Conference on Credibility, pages 129–163, New York, 1975.
Academic Press.
W. S. Jewell. The use of collateral data in credibility theory: a hierarchical

model. Giornale dell’Istituto Italiano degli Attuari, 38:1–16, 1975.
S. A. Klugman, H. H. Panjer, and G. Willmot. Loss Models: From Data to Deci-
sions. Wiley, New York, 4 edition, 2012. ISBN 978-1-118-31532-3.
E. Ohlsson. Simplified estimation of structure parameters in hierarchical

credibility. Presented at the Zurich ASTIN Colloquium, 2005. URL http:
//www.actuaries.org/ASTIN/Colloquia/Zurich/Ohlsson.pdf.
19

Credibility Theory Features of Actuar

Uploaded by

Copyright:

Available Formats

Credibility Theory Features of Actuar

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Credibility Theory Features of Actuar

Uploaded by

Copyright:

Available Formats

Credibility theory features of actuar

2 Hachemeister data set

3 Hierarchical credibility model

π̂ij = zij Xijw + (1 − zij )π̂i

with the credibility factors

and the weighted averages

(Hence, E[ Ai ] = ci a and E[ B] = db.) Then, the Bühlmann–Gisler estimators

the Ohlsson estimators are

Structure Parameters Estimators

Collective premium: 1746

Between cohort variance: 88981

Collective premium: 1746

Between cohort variance: 88981

Structure Parameters Estimators

Collective premium: 1746

Between cohort variance: 88981

4 Bühlmann and Bühlmann–Straub models

and the iterative estimator

is better known as the Bichsel–Straub estimator.

Structure Parameters Estimators

Collective premium: 1671

Between state variance: 72310

Collective premium: 1684

Between state variance: 89639

5 Regression model of Hachemeister

to the original data set of Hachemeister (1975).

Structure Parameters Estimators

Collective premium: 1469 32.05

Between state variance: 24154 2700.0

Figure 1: Collective, individual and credibility regression lines for State 4 of

> predict(fit, newdata = data.frame(time = 13))

Structure Parameters Estimators

Collective premium: -1675 117.1

Between state variance: 93783 0

state Indiv. coef. Cred. matrix Adj. coef.

Figure 2: Collective, individual and credibility regression lines for State 4 of

6 Linear Bayes model

where m = E[µ(Θ)] and z = n/(n + K ) for some constant K.

Therefore, the Bayesian premium is

with m = E[µ(Θ)] = E[Θ] = α/λ and

Structure Parameters Estimators

Between variance: 0.3333

Structure Parameters Estimators

Between variance: 0.3333

Indiv. mean Weight Cred. factor Bayes premium

A Linear Bayes formulas

• the collective premium m = E[µ(Θ)];

• the credibility factor when the Bayesian premium is expressed as a cred-

A.1 Bernoulli/beta case

A.2 Binomial/beta case

A.3 Geometric/Beta case

A.4 Negative binomial/Beta case

A.5 Poisson/Gamma case

A.6 Exponential/Gamma case

A.7 Gamma/Gamma case

A.8 Normal/Normal case

σ12 ∑nt=1 xt + σ22 µ

H. Bühlmann and E. Straub. Glaubgwürdigkeit für Schadensätze. Bulletin of

W. S. Jewell. The use of collateral data in credibility theory: a hierarchical