0% found this document useful (0 votes)
11 views53 pages

Open BMS Model for Ruin Probabilities

The document presents a model for calculating ruin probabilities for insurance portfolios using an open bonus-malus system. The model allows for more flexibility in bonus classes and considers competitive dynamics. Key assumptions include claim severities being homogeneous, claim frequencies following a mixed Poisson distribution, and homogeneous frequencies within each bonus level. Transition probabilities between levels factor in probabilities of exiting the portfolio.

Uploaded by

zahir abdellah
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
11 views53 pages

Open BMS Model for Ruin Probabilities

The document presents a model for calculating ruin probabilities for insurance portfolios using an open bonus-malus system. The model allows for more flexibility in bonus classes and considers competitive dynamics. Key assumptions include claim severities being homogeneous, claim frequencies following a mixed Poisson distribution, and homogeneous frequencies within each bonus level. Transition probabilities between levels factor in probabilities of exiting the portfolio.

Uploaded by

zahir abdellah
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

RUIN PROBABILITIES AND CAPITAL

REQUIREMENT FOR OPEN


AUTOMOBILE PORTFOLIOS WITH A
BONUS-MALUS SYSTEM

Zahir Abdellah

Leibniz University Hannover

January 2024

Zahir Abdellah (Institute) 20/01 1 / 23


INTRODUCTION AND MOTIVATION

The main goal of this presentation is to improve the methodology for


calculating …nite time ruin probabilities for large motor insurance
portfolios by updating the existing Markovian Bonus-Malus System
(BMS) model. While the classical BMS model assumes certain
behaviors and …xed entry bonus classes, the authors argue that an
open BMS, which allows for more ‡exibility in bonus classes and
considers the competitive dynamics of the motor insurance market,
provides a more realistic representation of the business.
The implications of the proposed open BMS model are explored,
particularly in terms of its potential impact on ruin probabilities. The
authors suggest that the introduction of an open model may lead to
signi…cant changes in ruin probability …gures compared to classical
BMS models

Zahir Abdellah (Institute) 20/01 2 / 23


Modeling the Ruin Probability

we consider a risk process over n-year period,


The evolution of the surplus of an insurance company or portfolio, U (t ),
for any time 0 6 t 6 n, is given by
i 1
U (t ) = u + ∑ Pj + (t (i 1))Pi S (t ),
j =1

Zahir Abdellah (Institute) 20/01 3 / 23


Modeling the Ruin Probability

We summarize the basic assumptions of our model formulation as follows:


the portfolio is homogeneous with respect to claim severities

Zahir Abdellah (Institute) 20/01 4 / 23


Modeling the Ruin Probability

We summarize the basic assumptions of our model formulation as follows:


the portfolio is homogeneous with respect to claim severities
the portfolio is heterogeneous with respect to claim frequencies,
following a mixed Poisson distribution;

Zahir Abdellah (Institute) 20/01 4 / 23


Modeling the Ruin Probability

We summarize the basic assumptions of our model formulation as follows:


the portfolio is homogeneous with respect to claim severities
the portfolio is heterogeneous with respect to claim frequencies,
following a mixed Poisson distribution;
we consider a homogeneous claim frequency in each bonus level, that
is, in level j the number of reported claims in one year is Poisson
distributed with parameter λj , j = 1, 2...L where L is the number of
bonus malus levels or classes

Zahir Abdellah (Institute) 20/01 4 / 23


Modeling the Ruin Probability

We summarize the basic assumptions of our model formulation as follows:


the portfolio is homogeneous with respect to claim severities
the portfolio is heterogeneous with respect to claim frequencies,
following a mixed Poisson distribution;
we consider a homogeneous claim frequency in each bonus level, that
is, in level j the number of reported claims in one year is Poisson
distributed with parameter λj , j = 1, 2...L where L is the number of
bonus malus levels or classes
the portfolio is open for incoming and outgoing of policyholders

Zahir Abdellah (Institute) 20/01 4 / 23


Modeling the Ruin Probability

Let H (s ) + κs Γ(α, β) (with κ is the translation parameter) a random


variable with a translated Gamma distribution whose …rst three moments
match those of S (s ) and we denote FG (., s ) and fG (., s ) the CDF and
PDF of H (s ), we can approximate the ruin probability in year i as de…ned
in a previous work by
u (i ) u (i ) κu (i ) u (i )
fG (u (i 1 )+(p i κ ))(1 p i ),1 p i ))F G ( pi , pi )
Ψ (u (i 1), 1, u (i )) = fG (u (i 1 )+p i κ u (i ),1 )
1 u (i )
Z
p (i )

u (i )
fG (u (i 1 )+(p i κ )s ,s )( 1 s )fG ((p i κ )(1 s ) u (i ),1 s )
s =0
+ fG (u (i 1 )+p i κ u (i ),1 )
ds
we will use this formula to estimate the probability of ruin for a …nite time,
say n, in a later section

Zahir Abdellah (Institute) 20/01 5 / 23


BMS for Open Portfolios
The main results on BMS for open portfolios

We assume that the exits from the portfolio do not need to be


perfectly compensated by new entries and use an extra absorbing
state L + 1 , This represents the exit from the portfolio, We’ll have
then L + 1 BMS classes, where states 1,...L are transient and L + 1 is
an absorbing state.
the corresponding transition block matrix is given by

KT ,λ qλ
PT ,λ =
0 1

where λ is the claim frequency and T a set of transition rules KT ,λ is


a (L L) matrix representing the one step transition probabilities
among the BMS classes for those policies that remain in the portfolio,
qλ is a column vector of conditional probabilities of an insured leaving
the portfolio at the end of the time period
Zahir Abdellah (Institute) 20/01 6 / 23
BMS for Open Portfolios

in classical BMS models, for a given set of transition rules, the


probability of a randomly chosen policyholder, with a given claim
frequency λ , move from class l to class j is given by

PT ,λ = ∑ Pk (λ)tl ,j (k ) l, j = 1, . . . , L
k =0

where Pk (λ) is the probability of an insured with claim frequency λ


reporting k claims in 1 year, tl ,j (k ) = 1 if he reports k claims leading
the policy to move from class l to class j according to transition rules
T and tl ,j (k ) = 0, otherwise.

Zahir Abdellah (Institute) 20/01 7 / 23


BMS for Open Portfolios
In the open model formulation, we set the transition probabilities
among classes 1, ..., L to be
kT ,λ (l, j ) = pT ,λ (l, j )(1 qλ (l )), l, j = 1, . . . , L

Zahir Abdellah (Institute) 20/01 8 / 23


BMS for Open Portfolios
In the open model formulation, we set the transition probabilities
among classes 1, ..., L to be
kT ,λ (l, j ) = pT ,λ (l, j )(1 qλ (l )), l, j = 1, . . . , L
where kT ,λ is the entry level of the matrix KT ,λ and qλ (l ) is the
probability of existing the company, so , a policyholder with claim
frequency λ moves, at the end of the year, from class l to class j only
if he doesn’t exit the company

Zahir Abdellah (Institute) 20/01 8 / 23


BMS for Open Portfolios
In the open model formulation, we set the transition probabilities
among classes 1, ..., L to be
kT ,λ (l, j ) = pT ,λ (l, j )(1 qλ (l )), l, j = 1, . . . , L
where kT ,λ is the entry level of the matrix KT ,λ and qλ (l ) is the
probability of existing the company, so , a policyholder with claim
frequency λ moves, at the end of the year, from class l to class j only
if he doesn’t exit the company
The n-step transition matrix, n = 1, 2..., is given by
(n ) KTn ,λ qn,λ
PT ,λ =
0 1
n 1
with qn,λ = ∑ KTj ,λ qn,λ , KTj ,λ corresponds to the j-step transition
j =0
probabilities for the policyholders remaining in the portfolio after
these j years
Zahir Abdellah (Institute) 20/01 8 / 23
Assumptions regarding the evolution of the portfolio over
time

the arrival of new policies at the portfolio happens at the beginning of


each time period

chra7

Zahir Abdellah (Institute) 20/01 9 / 23


Assumptions regarding the evolution of the portfolio over
time

the arrival of new policies at the portfolio happens at the beginning of


each time period
new policies entering the portfolio are allocated to a BMS level
according to the probability vector ci = [ci (l )]1 L , i 2 N

chra7

Zahir Abdellah (Institute) 20/01 9 / 23


Assumptions regarding the evolution of the portfolio over
time

the arrival of new policies at the portfolio happens at the beginning of


each time period
new policies entering the portfolio are allocated to a BMS level
according to the probability vector ci = [ci (l )]1 L , i 2 N
the number of new policies entering the portfolio at time period i are
independentrandom variables with mean value ϑi , and are denoted as
Ei .
chra7

Zahir Abdellah (Institute) 20/01 9 / 23


Assumptions regarding the evolution of the portfolio over
time

In order to evaluate and compare the ruin probabilities over di¤erent


hypotheses on portfolio evolution, we consider two di¤erent models
for the mean value ϑi , namely:
Scenario 1— Exponential Model:

ϑ i = τ (1 e δi
), i 2 N, τ, δ 2 R+

Scenario 2— Sigmoid Model:

ϑi = (a + be θi
) 1
, (a, b, θ ) 2 Θ, i 2 N

with Θ = (a, b, θ ) : a 2 R+ , b, θ 2 R, a + be θi 0, i 2 N .

Zahir Abdellah (Institute) 20/01 10 / 23


THE PORTFOLIO EVOLUTION AND LIMITING
RESULTS

Theorem
When the mean number of new policies entering the portfolio in period i is
modeled by ϑi and the insurer already has an existing portfolio with
policies distributed over BMS classes according to the row vector
ϑ0 ’= [ϑ0 (j )] , j = 1, ..., L the row vector of expected number of
policyholders in the BMS classes in time period i for a given λ will be
given by :
0 i
∑ ϑk ck Kλi
0 0
ϑi+,λ = ϑ0 Kλi + k
, i 2N
k =1

Zahir Abdellah (Institute) 20/01 11 / 23


THE PORTFOLIO EVOLUTION AND LIMITING
RESULTS
the proportion of policyholders, with claim frequency λ, belonging to class
j in year i, is given by
ϑi+,λ (j )
π i ,λ (j ) = L
, j = 1, ..., L, i 2 N
0
∑ ϑi+,λ (j )
j =1

with an open portfolio approach, the asymptotic properties of a Markov


chain do not apply
Given λ, the limiting state probability for a policyholder belonging to
bonus class j is
ϑ+
∞,λ (j )
π ∞,λ (j ) = L
, j = 1, ..., L,
0
∑ ϑ+
∞,λ (j )
j =1

with ϑ+ +
∞,λ = limi !∞ ϑi ,λ
Zahir Abdellah (Institute) 20/01 12 / 23
THE PORTFOLIO EVOLUTION AND LIMITING
RESULTS
the heterogeneity present in the insurance portfolio with respect to claim
frequencies is modeled by treating claim frequency as a random variable
that can vary across individuals. This randomness is captured by
considering claim frequency λ as an outcome of a random variable Λ with
a certain distribution VΛ (.).
As widely set in the BMS literature, the unconditional probability of an
insured belonging to class j, after i steps, and the long run distribution, for
a policyholder chosen at random from the portfolio, is assumed as the
expectation with respect to Λ,
Z∞
π i (j ) = π i ,λ (j )dV (λ), j = 1, ..., L
0
and
Z∞
π ∞ (j ) = π ∞,λ (j )dV (λ), j = 1, ..., L
Zahir Abdellah (Institute)
0 20/01 13 / 23
The portfolio evolution and limiting results
the portfolio dimension in year i, measured by the number of policies, is
denoted as NPoli , is given by
L
NPoli = ∑ NPoli (j ), i 2N
j =1

Due to the fact that the portfolio is open, NPoli changes over time
considering the new annual entries and exits.
For a given year i and known involved quantities, total premium in the
presence of a BMS can be computed as follow
L
Pi = (1 + ξ )NPoli ∑ E (S (1))πi (j )bj , i = 1, ..., n
j =1

where ζ 0 is the safety loading parameter and bj is the relativity of level


j, that is, the proportion of a priori premium to apply in level j
we consider E (S (1)) to be dependent on Class j and for BMS based only
on claim frequency there is an implicit assumption that average individual
claim size is constant across BMS classes
Zahir Abdellah (Institute) 20/01 14 / 23
SIMULATION AND ESTIMATION PROCEDURE

In this section, we will summarize the steps of the simulation (and


calculation for ruin probabilities) procedure using real data (historical
data). This model is targeted for large portfolios, we need to calculate
annual aggregate claims, the authors suggest the approximation by a
translated gamma with parameters α, β and κ .
The steps are outlined and listed below.
Estimation of expected claim frequency λj , j = 1, ..., L. From
historical Data

Zahir Abdellah (Institute) 20/01 15 / 23


SIMULATION AND ESTIMATION PROCEDURE

In this section, we will summarize the steps of the simulation (and


calculation for ruin probabilities) procedure using real data (historical
data). This model is targeted for large portfolios, we need to calculate
annual aggregate claims, the authors suggest the approximation by a
translated gamma with parameters α, β and κ .
The steps are outlined and listed below.
Estimation of expected claim frequency λj , j = 1, ..., L. From
historical Data
Estimation of the mean value of new annual policies ϑi , i = 1, ..., L.
Estimate the mean number of new annual contracts arriving to the
portfolio using regression techniques or time series models.

Zahir Abdellah (Institute) 20/01 15 / 23


SIMULATION AND ESTIMATION PROCEDURE

Estimation of allocation probabilities ci (j ), i = 1, ..., n.j = 1, ..., L


Using MLE and observations from real data in m past year periods,
estimate the probability of a new contract to be allocated at level j in
E
year i by bci (j ) = Ei i,j . Ei refers to the number of new contracts in
year i and Ei ,j to the number of contracts that, in year i, were
allocated to level j. In particular, in stable portfolios (with respect to
allocation probabilities), we may set ci = c and, in this way
m
∑ Ei ,j
i =1
b
c (j ) = m
∑ Ei
i =1

Zahir Abdellah (Institute) 20/01 16 / 23


SIMULATION AND ESTIMATION PROCEDURE
Estimation of exit probabilities qλ (j ), Using real data from m past
year periods, estimate the exit probabilities for each BMS level j using
m
∑ A i ,j
i =1
bλ (j ) =
q m
∑ C i ,j
i =1

where Ai ,j is the number of annulments from class j in year i and Ci ,j the


number of policies that, in year i, were allocated to class j.

Zahir Abdellah (Institute) 20/01 17 / 23


SIMULATION AND ESTIMATION PROCEDURE
Estimation of exit probabilities qλ (j ), Using real data from m past
year periods, estimate the exit probabilities for each BMS level j using
m
∑ A i ,j
i =1
bλ (j ) =
q m
∑ C i ,j
i =1

where Ai ,j is the number of annulments from class j in year i and Ci ,j the


number of policies that, in year i, were allocated to class j.
Estimation of the expected number of claims for the portfolio
L
E (Ni ) = NPoli ∑ λj π i (j )
j =1

Zahir Abdellah (Institute) 20/01 17 / 23


SIMULATION AND ESTIMATION PROCEDURE
Estimation of exit probabilities qλ (j ), Using real data from m past
year periods, estimate the exit probabilities for each BMS level j using
m
∑ A i ,j
i =1
bλ (j ) =
q m
∑ C i ,j
i =1

where Ai ,j is the number of annulments from class j in year i and Ci ,j the


number of policies that, in year i, were allocated to class j.
Estimation of the expected number of claims for the portfolio
L
E (Ni ) = NPoli ∑ λj π i (j )
j =1
Simulation of the aggregate claim amount for each year i, fYi gni=1 ,
Yi is the aggregate claim amount in a given year i, assumed to have a
translated Gamma distribution

Zahir Abdellah (Institute) 20/01 17 / 23


SIMULATION AND ESTIMATION PROCEDURE
Estimation of exit probabilities qλ (j ), Using real data from m past
year periods, estimate the exit probabilities for each BMS level j using
m
∑ A i ,j
i =1
bλ (j ) =
q m
∑ C i ,j
i =1

where Ai ,j is the number of annulments from class j in year i and Ci ,j the


number of policies that, in year i, were allocated to class j.
Estimation of the expected number of claims for the portfolio
L
E (Ni ) = NPoli ∑ λj π i (j )
j =1
Simulation of the aggregate claim amount for each year i, fYi gni=1 ,
Yi is the aggregate claim amount in a given year i, assumed to have a
translated Gamma distribution
Estimation of the premium collected in each year i, Pi and estimate
the total amount of premium collected in year i For a given bonus
scale
Zahir Abdellah (Institute) 20/01 17 / 23
SIMULATION AND ESTIMATION PROCEDURE

Estimation of the ruin probability in year n , Ψ(u, n).this step is the


…nal step and performed as follow :

Zahir Abdellah (Institute) 20/01 18 / 23


SIMULATION AND ESTIMATION PROCEDURE

Estimation of the ruin probability in year n , Ψ(u, n).this step is the


…nal step and performed as follow :
1 we simulate values of fYi gni=1 and calculate the surplus at the end of
each year:u (i ) = u (i 1) + pi yi , i = 1, ..., n

Zahir Abdellah (Institute) 20/01 18 / 23


SIMULATION AND ESTIMATION PROCEDURE

Estimation of the ruin probability in year n , Ψ(u, n).this step is the


…nal step and performed as follow :
1 we simulate values of fYi gni=1 and calculate the surplus at the end of
each year:u (i ) = u (i 1) + pi yi , i = 1, ..., n
2 if u (i ) < 0 we say that Ψm (u, n) = 1 and start sumilatation in the
run m + 1 and if u (i ) > 0 we calculate the approximation for run m
Ψm (u (i 1), 1, u (i ))

Zahir Abdellah (Institute) 20/01 18 / 23


SIMULATION AND ESTIMATION PROCEDURE

Estimation of the ruin probability in year n , Ψ(u, n).this step is the


…nal step and performed as follow :
1 we simulate values of fYi gni=1 and calculate the surplus at the end of
each year:u (i ) = u (i 1) + pi yi , i = 1, ..., n
2 if u (i ) < 0 we say that Ψm (u, n) = 1 and start sumilatation in the
run m + 1 and if u (i ) > 0 we calculate the approximation for run m
Ψm (u (i 1), 1, u (i ))
3 Calculate the …nite time ruin probability estimate in run m , as follow
n h i
Ψb m (u, n) = 1 ∏ 1 Ψ b m (u (i 1), 1, u (i ))
i =1

Zahir Abdellah (Institute) 20/01 18 / 23


SIMULATION AND ESTIMATION PROCEDURE

Estimation of the ruin probability in year n , Ψ(u, n).this step is the


…nal step and performed as follow :
1 we simulate values of fYi gni=1 and calculate the surplus at the end of
each year:u (i ) = u (i 1) + pi yi , i = 1, ..., n
2 if u (i ) < 0 we say that Ψm (u, n) = 1 and start sumilatation in the
run m + 1 and if u (i ) > 0 we calculate the approximation for run m
Ψm (u (i 1), 1, u (i ))
3 Calculate the …nite time ruin probability estimate in run m , as follow
n h i
Ψb m (u, n) = 1 ∏ 1 Ψ b m (u (i 1), 1, u (i ))
i =1

4 The estimate for the continuous and …nite time ruin probability
b (u, n), is set by the mean of the estimates obtained from each
Ψ
simulation Ψ b m (u, n), m = 1, ..., M
Zahir Abdellah (Institute) 20/01 18 / 23
RUIN PROBABILITIES IN AN OPEN PORTFOLIO WITH
A BMS
The e¤ect of a BMS in the probability of ruin of a motor portfolio

In this section we discuss the e¤ect of a BMS with open model in the
probability of ruin of a motor portfolio and We compare results using
the same automobile portfolio and the BMS utilized in the classical
model to get clear conclusions.

Zahir Abdellah (Institute) 20/01 19 / 23


RUIN PROBABILITIES IN AN OPEN PORTFOLIO WITH
A BMS
The e¤ect of a BMS in the probability of ruin of a motor portfolio

In this section we discuss the e¤ect of a BMS with open model in the
probability of ruin of a motor portfolio and We compare results using
the same automobile portfolio and the BMS utilized in the classical
model to get clear conclusions.
The insurer’s commercial scale has 18 premium entries (bonus-malus
levels), with transition Matrix, which establishes Level 10 as the entry
class

Zahir Abdellah (Institute) 20/01 19 / 23


RUIN PROBABILITIES IN AN OPEN PORTFOLIO WITH
A BMS
The e¤ect of a BMS in the probability of ruin of a motor portfolio

In this section we discuss the e¤ect of a BMS with open model in the
probability of ruin of a motor portfolio and We compare results using
the same automobile portfolio and the BMS utilized in the classical
model to get clear conclusions.
The insurer’s commercial scale has 18 premium entries (bonus-malus
levels), with transition Matrix, which establishes Level 10 as the entry
class
Number of claims reported by a randomly chosen insured follows a
mixed Poisson distribution, where the random parameter Λ follows an
Inverse Gaussian distribution with parameter estimates µ̂ = 0.082401
and η̂ = 0.130271

Zahir Abdellah (Institute) 20/01 19 / 23


the number of annual claims follows a Poisson distribution with
parameter λj

Zahir Abdellah (Institute) 20/01 20 / 23


the number of annual claims follows a Poisson distribution with
parameter λj
The number of existing policies in each level j, at the evaluation date,
ϑ0 , was known .

Zahir Abdellah (Institute) 20/01 20 / 23


the number of annual claims follows a Poisson distribution with
parameter λj
The number of existing policies in each level j, at the evaluation date,
ϑ0 , was known .
From the history of the insurer’s portfolio claim amounts, we
estimated a mean value of 1, 766.31, a variance of 71, 097, 953.5 and
a third central moment of 21, 068, 298, 856, 615. They were then
used to get estimates for the parameters α, β and κ of the translated
Gamma approximation

Zahir Abdellah (Institute) 20/01 20 / 23


Ruin Probability
Estimates for the probability of ruin

the calculated total premium without BMS for the portfolio with 442, 490
policies is P = (1 + 0.8)E (S (1)) = 115, 838, 792
for the calculation of the ruin probability Ψ(u, n), we …xed the initial
surplus at u = 2, 000, 000 and we used M = 50, 000 runs

Zahir Abdellah (Institute) 20/01 21 / 23


Ruin Probability
Changing Allocations and Annulment Probabilities

To evaluate the impact of allocations and annulments on ruin probabilities,


we considered the three scenarios for the allocation and exit probabilities
,working only with Open Sigmoid Model,

Zahir Abdellah (Institute) 20/01 22 / 23


Zahir Abdellah (Institute) 20/01 23 / 23

You might also like