This article was downloaded by: [University of Toronto Libraries]

On: 30 December 2014, At: 21:02

Publisher: Routledge
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office:
Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Accounting and Business Research

Publication details, including instructions for authors and subscription
information:
http://www.tandfonline.com/loi/rabr20

Markov Chain Models for Motor Insurance

Robert Croasdale
Published online: 27 Feb 2012.

To cite this article: Robert Croasdale (1972) Markov Chain Models for Motor Insurance, Accounting and
Business Research, 2:7, 177-181, DOI: 10.1080/00014788.1972.9728595

To link to this article: http://dx.doi.org/10.1080/00014788.1972.9728595

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”)
contained in the publications on our platform. However, Taylor & Francis, our agents, and our
licensors make no representations or warranties whatsoever as to the accuracy, completeness, or
suitability for any purpose of the Content. Any opinions and views expressed in this publication
are the opinions and views of the authors, and are not the views of or endorsed by Taylor &
Francis. The accuracy of the Content should not be relied upon and should be independently
verified with primary sources of information. Taylor and Francis shall not be liable for any
losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities
whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or
arising out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any substantial
or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or
distribution in any form to anyone is expressly forbidden. Terms & Conditions of access and use
can be found at http://www.tandfonline.com/page/terms-and-conditions
 SUMMER 1972
Markov Chain Models for Motor Insurance
Robert Croasdale
Summary one of a number of states. Transitions from one state
An introductory account of Markov Chain Theory* to another can take place at fixed times and the prob-
is interwoven with a discussion of how under certain abilities of transitions between given states depend on
conditions the theory can be used in the study of the current state of the system and not directly on its
motor insurance policies which offer no-claim
Downloaded by [University of Toronto Libraries] at 21:02 30 December 2014
discounts.
Introduction
Many British motorists have comprehensive car
insurance policies offering a discount or ‘no-claim
bonus’ on terms such as the following:
Consecutiveyears Discount as % of
without a claim whole premium
I 30
2 40
3 50
4 60
The policyholder will revert to payment of the
whole premium in the event of a single claim after
one or two years claim-free. He will revert to
30%/40% discount if a single claim is made or
arises in the year following the allowance of a
discount of 50 %/6o% then, assuming the company
agrees and continues insurance at the next subse-
quent renewal date, entitlement to discount will
apply as if there had been one yearltwo years free
of claims,
Now the chance that a given policyholder will have
an accident is a complicated function of numerous
factors such as his health, the driving habits of him-
self and others, weather conditions, traffic volume,
vehicle speeds and degrees of roadworthiness. In the
rather unlikely event that (a) such factors act so that
the chances remain the same every year that any
policyholder will have a claim-free year and if (b) the
terms of the policy allow for the annual premium to
be determined solely by the man’s premium and
claims record for the previous year, then the con-
ditions are favourable for studying the operation of
that policy as a Markov chain.
Markov chains
A Markov chain describes a system which can be in
For more advanced theory see Feller, W.,A n Intro- or, symbolically
duction to Probability Theory and its Applications, Vol. I
(Third Edition), New York. Wiley (1968).
previous history. We shall consider such systems for
which transition probabilities are independent of time,
i.e. homogeneous Markov chains.
A 2-state Markov chain
Let us consider a homogeneous Markov chain with
2 states, state I and state 2 and let plz denote the
probability that an individual member of a system will
be in state 2 at time n + I given that the individual is
in state I at time n. We may note that for a Markov
chain no knowledge of the states of the individual at
times n- I ,n-2, . . .,etc. is needed in order to express
the transition probabilities for time n. We define
Piu P21 and P22 similarly.
It is usual to arrange these probabilities in the form
of a matrix thus
Next state
I 2
with the convention that rows refer to ‘original’ states
and the columns to ‘final’ states.
This belongs to a family of matrices known as
stochastic matrices, which have the property that the
s u m of the elements of any row is unity.
If, at time n, the system consisted of a proportion
fi(n) of individuals in state I and the remainder
f2(n)=l-fl(n) in state 2, then the expected proportion
of individuals in state I at time n+I would be given
by :
(Proportion in state I at time n) x (Probability of
being in state I at time n+r given in state I at
time n)
+(Proportion in state 2 at time n)x(Probability
of being in state I at time n + given
~ in state 2 at
time n)
fl(n+l)=f 1(n)p11+f,’n’ pzl .......... . . (1)
 178 A C C O U N T I N G A N D B U S I N E S S RESEARCH

and, in the same way, previous year. The matrix of transition probabilities
fz(n+l)=fl(n) p 12+f 2(n) pzz . . . . . . . . . . (2) for a holder of this type of policy, whose chances of
If we introduce a row vector whose elements are the making a claim are unaffected by his enjoying or not
proportions of the population in the different states at enjoying a discount, is
time n (the 'state vector' for time n), say f ( n ) = Next premium
(fl(n),f,(n)) Y 0.8Y
then we can write equations (I) and (2) in matrix form
f (n+l)=f(n). P . . . . . . . . . . . . . . . * (3)
the dot denoting matrix multiplication (see Ayres
(1962) Chapter I). The equations contained in (3) are
If the award of a discount served not just as a
known as the Chapman-Kolmogorov Equations for
reward for a claim-free year, but also as an induce-
the system. Just as we have derived equation (3)
ment towards further claim-free years, the appro-
which expresses the proportions in the different states
priate matrix might then be of the form:
at time n+ I in terms of
Downloaded by [University of Toronto Libraries] at 21:02 30 December 2014

(i) the corresponding proportions at time n, and

(ii) the transition matrix P (which is independent of
P,= (Fi '-")where
-Pz
1 1>p1>p2>o
time), and with the same convention as to the states associated
so we can express the corresponding state vector for with rows and columns.
time n in terms of the state vector for time n-1
and P. The distribution of policyholders
Thus f(n)=f(n-l). P. . . . . . . . . . . . . (4) between premium groups
Similarly f(n--l)=f(n--$). P . . . . . . . . . . . (5) Although in practice it may be difficult for a company
By using (4)repeatedly for different values of n, we to identify a homogeneous group of policyholders, let
obtain us suppose by way of illustration that, in a particular
f(n),f(n-Z). p2 year, an insurance company had a number of policy-
-
-f(n-k). p k holders characterised by P,, of whom a proportion
=f(0).Pn.. . . . . . . . . . . . . . . . . . (6) fl(n) were paying full premium and the rest were
where f ( O ) is the initial state vector, i.e. the vector receiving discount. Then, in the following year the
composed of the proportions in the respective states proportions (assuming no new customers and no
at time 0. cancellations of policies) on N1 and reduced pre-
The computation of successive powers of P for use miums, fl(n+l)and fz(n+l)=I-fI(n+l) respectively,
as in equation (6) is often facilitated by spectral de- would be given by
composition of the matrix (see Ayes (1962) Chapter f ( n + l ) =f ( n ) . P, . . . . . . . . . . . . . . . . . (7)
21). which is equivalent to equations (I) and (2) with
p11=1 -P21=P1
A motor insurance policy with one and PlZ" 1--Pzz=pz
level of discount
Consider a simplified car insurance policy which offers Equilibrium (or steady-state)
the following discount terms : solution
'The policyholder will pay a premium of EY p.a. in If an equilibrium situation exists in which the pro-
the first year. If he has no claim during that year he portions of the policyholders in the two categories
will be allowed a 20% discount in the second year. eventually become constant, these equilibrium pro-
If there is at least one claim in the first year he will portions
continue to pay the full premium in the second year. f(m),(fl(m), fi(m))
If there is no claim then in the second year he will say, can be determined by letting n increase without
be allowed a 20% discount. The discount rate limit in equation (3). Then we have
continues Until there is a claim, whereupon he f(a)=f(m).P . . . . . . . . . . . . . . . . . . (8)
reverts to the full premium in the following year.' It should be noted that the proportions in different
The policy can thus be in one of two states only: states may be constant even though individuals may
Stare I. Pays full premium LY. still be changing states, if the rates of change in each
State 2 Pays full premium less 200/6, i.e. he pays direction are equal. This phenomenon, known as
E0-8Y. dynamic equilibrium, can also occur in chemical
The process is Markovian since the probabilities reactions or among an electorate in which some voters
associated with the different levels of premium in a switch allegiances between parties at successive
given year depend only on the premium paid in the elections.
 SUMMER 1972 I79

A numerical example these probabilities may differ according to whether

In the case characterised by P2, suppose we have a full or reduced premiums were paid in the last year.
company starting in year o with all its holders of this We further suppose that after a cancellation the ex-
type of policy paying full premium. policyholder concerned is lost to the company for ever.
i.e. f1(O)=I and f2(0)=o This provision introduces a third state, which is an
Further, let us suppose that, for all these policy- absorbing state. Let the appropriate matrix be
holders,
p,=0*02 and P,=O.OI Next premium
Then in Year I the proportions in the two states of 0.8Y ‘0’ Y
premium payable will be given by c Y (0.018 r980 0.002
P,= g ‘3 0-8Y 0-oog 0.990 0.001
3O Q&) ‘0’ 0 I

The equilibrium proportions are given by solving

In Year 2 the proportions in the two states will be f(oo)=f(m)* p
Downloaded by [University of Toronto Libraries] at 21:02 30 December 2014

given by
(fi(2),f2(2))=(0.02y0.98)
and in Year 3 we expect
(fl(,), f2(3))=(0.0~02,0.9898)
and so on.
In the long run :
3
for f(m),by which means we obtain
fl(m)=f2(m)=0 f (OO)=I
Y 3
i.e. if this trend continued the company would
ultimately lose all its policyholders, though this
(o‘0102’ 0*9898) would only happen very slowly. Even after 64 years
only about one in 16 of the original policyholders
would have been lost as Table I shows.
0.02 0.98 -
(0.01 0.99)-
(0-010012,0.989988)
TABLE 1
.002 .003 .005 .009 .017 .033 a0635
p z ~ .001 *002 so04 .008 .016 -031 -0626

(f1(W),f2(m))=(f1(m),f 2 9(0.02
0’01
0.98)
0.99
. . . . (9)
fi(m)=0.02f,’m)+001 fi(m) . . . . . . . (Io) p13 denotes the proportion of clients initially paying
f2(m)=o.g8 f,(m)+o.gg fJm). . . . . . . (11) full premium that will have been lost to the company
in the following n years.
giving fi(m)=fl(m):98:1
p23is the corresponding propomon for clients who had
Hence, in the long run only an average of one in 99 of discount initially.
these policyholders will be paying full premium; the
rest will be enjoying discount. Motor insurance policies with
In general, for the P, situation, the equilibrium several levels of discount
proportion paying full premium is P ~ K I-P1+PJ. SO Until a few years ago car insurance policies of the
if, for example, the probabilities of claims increased following type were common in the United Kingdom,
tenfold, thus p:=0’2J p ~ = O * Ithen
, in the long run (a) The premium paid in the fitst year is Y. If no
one in nine policyholders (as compared with one in 99 claim is made in the first year, a discount of
previously) would pay full premium. IOOayo is allowed in the second year, (so the
Looked at from the insurance company’s point of premium charged is (~--cc)Y). If no claim is made
view, in the equilibrium state, with a 20% discount for two consecutive years the discount is doubled,
offered, the average premium received per policy is the resulting premium being (1-2a)Y, and so on,
with a limit on the maximum discount allowable,
of the form: no further reductions after N con-
=80.2Y0 of full premium when p1=2p2=o-02 secutive claim-free years. In the event of a claim at
or 82.2y0 of full premium when p1=2p2=0*20 any stage, the policyholder reverts to full premium
in the following year.
Failure to renew the policy e.g. Premium payable in year n f I
Let us suppose now that there are constant probabili- =(I-na)Y, oGn<N
ties that the policy will not be renewed (either because =(I -Na)Y, n>N
the company refuses to renew the policy or the policy- The transition matrix for such a policy, neglecting
holder decides to take his business elsewhere) and that cancellations, would be:
 180 A C C O U N T I N G A N D B U S I N E S S RESEARCH
Next premium
Y (I-a)Y (I-2a)Y (I-3a)Y ...... (I-Na)Y
P 9 0 0 ...... 0
0 ...... 0

P4= a
(I -2a)Y 0 q ...... 0
..........
0 0 ......
0 0 ......

A variation of this type of policy was one which P P4 Pq2 qS

(
Downloaded by [University of Toronto Libraries] at 21:02 30 December 2014

awarded the no-claim bonus as a fixed percentage of

the previous year's premium instead of making a P P9 Pq2 C13
flat percentage reduction on the basic premium = pq Pq2
P P9 pq2 q3
q3)
thus :
(b) 'The premium paid in the first year is Y. If no and, generalising,
claim is made in the first year, the premium in the
..... 0 0
second year is BY where B is a constant (o<B<r).
If no claim is made for two consecutive years, the
p q o

p 0 .0 .....
~ . q ; )~N +~I .
rows
1 ~ ~ ~ ~
premium in the next year is P Y and so on, so that
in general, if no claim is made in n consecutive
years, then:
p 0 0 ..... J-
premium payable in year (n+z)=BnY, o < n<N
- (N+I) columns -
=BNY, n>N MN(N+l)=MN+l(N+l)=...
(i.e. the lowest premium ever payable is BNY).
If a claim is made, the premium for the following p pq pq2..... pq'-1 ..... pqN-l qN
..... pq'-1 ..... pqN-' qN
year reverts to Y and the whole procedure starts
again.'
In the case of a policyholder for whom the prob-
ability of a claim in any one year is a constant, p,
=( p pq pq2
.............................
p pq pq2 ..... pq'-1 ..... PqN-'
Thus, with N levels of discount in addition to the
qN

the transition matrix for annual premiums paid takes

full premium level, after N years, the probabilities
the form of P4 but the states differ, a premium of
associated with the different states are independent of
(1-ra)Y being replaced by PY.
the initial states.
We observe that Pahas its last two rows identical.
Alternatively the vector f(m)of the ultimate pro-
Let us write M(N+1)to denote P4w i t h N + ~rows and
portions in the various states could be determined
N +I columns. Then our matrix has the property that
directly from the equilibrium equations
the rows of Mm(N+I)(i.e. the mth power of M(N+l)) f(m)=f(m). M (N+1)
for m> N are identical.
Thus for a 3-state process we have: e.g. withN=2

M(3,=
(1 :I.
p o q

P P9 q2
M(3)2=M(3)3=M(3)4= ...
=(p, f1(m)q, [I -fl(m)]

since fl(m)+f2(m)+f3(m)=
q)

=(p pq q2) . . . . . . . . . . . . . (13) and p+q=I

I

P P9 q2 Hence fi(OD)=p, f2(m)=pq, f3(m)=q3

which agrees with the result in equation (13).
and for the 4-state process
(c) Policies of the type characterised by the matrix
P, had the unpleasant effect of depriving even
someone who had enjoyed the maximum discount
for a number of years, of all his no-claim bonus if
he had even a single claim in any year. Policies
 SUMMER 1972 I81

such as the one cited at the beginning of this paper State 2 means ‘Pay premium Yz next year having paid
(see matrix P,below) were a little gentler, allowing Y, this year’
the claimant to retain part of the bonus he had State 3 means ‘Pay premium Y2next year having paid
earned. Yzthis year’
Premium next year The matrix P, now represents a Markov chain.
Y -7Y *6Y -5Y -4Y
m y P q o o o Testing models
3 -7Y In order to examine whether the operation of a
P,= 2 -6Y certain type of policy behaves according to a Markov
*g 2 -5y
& g-4y 0 0
chain, it is necessary to collect data, preferably for a
given group or cohort of policyholders and to follow
the progress of their premiums over successiveyears.
Solving the equations Estimates of the transition probabilities can then be
Downloaded by [University of Toronto Libraries] at 21:02 30 December 2014

f(m),f(m). p
made and these can be used to estimate the numbers
gives
that would be expected in the various categories in
f(m)=(I-2pq2)-’ (P2[ISpql, p2q[1+ s l y pq2, succeeding years, on the assumption of a Markov
Ps3Y s3 model.
which implies an average per policyholder of
Two obvious departures from the conditions
EY[p2(I+2PQ)+0’7 P2q(I+q)+O*6Pq2+O’5Pq3+ assumed in this paper would be (a) transition
o*4q41/(I-2Pq2) probabilities that were not constant but changed with
An insurance company will want to take note of the
time, (b) a population of policyholders that did not
falling revenue from policies in succeeding years and
consist of one or more homogeneous groups. Indeed
of its ultimate income in order to ensure that claims
a typical population has been likened to a spectrum
can be met. The diminished value of &I paid in several
that has to be forced into classes. However, it might
years’ time compared with EI nowt and the increasing
still be possible sometimes to identify a few well-
sums involved in settlement of claims imply that either
defmed categories such as policyholders who never
premiums must be revised upwards every few years
claim or those who claim frequently and pay full
or that - less likely - lower discounts will have to be
premium year after year.
offered.
Conclusion
A Non-Markovian policy Markov chain models thus offer a promising approach
Consider next a policy which offers discounts as in the to the study of the performance of motor insurance
case of P,but requires that the 50% discount rate be policies. Whilst, as with any mathematical or statis-
earned for two consecutiveyears before the maximum
tical techniques, these should not be applied in-
of 60% discount is allowed. discriminately but only advisedly in the light of sound
At first sight this appears to violate the Markovian
business experience, there is a growing volume of
property since the probability of being on a premium
evidence (see e.g. Johnson et. al. (1972)) that, given
of 0.4Y in year n+x is dependent not only on the
appropriate safeguards, there are real benefits to be
premium paid in year n but also on that for year n-I.
achieved from this approach.
However this obstacle can be overcome by re-defining
the states, thus :
Next premium yo YI y 2 yz y3
Next state 0 1 2 3 4

t A ~ ‘ $ 6p.a. inflation rate will reduce the value of EI by
half in about fourteen years. A 1006 rate will do so in seven
years.
References
W. Feller (1968), An Introduction to Probability Theory
and its Applications Vol. I. (Third Edition) New York:
Wiley.
F. Ayres (1962). Theory and Problems of Matrices. New
York: Schaum.
P. D. Johnson and G. B. Hey (1972).A review of the
scope for using claims histories of individual policies in risk
assessment. 1. Inst. Actuaries.