Naaj9801 2

BONUS-MALUS SYSTEMS: THE EUROPEAN AND ASIAN APPROACH TO MERIT-RATING
Jean Lemaire*
ABSTRACT
Bonus-malus is a merit-rating technique used in most of Europe and Asia, and some Latin American and African countries. Policyholders from a given risk cell are subdivided into bonus-malus classes. Their claims histories then modify the class upon each renewal. Markov chain theory provides the tools for the design, evaluation, and comparison of these systems. In this article, denitions and examples of bonus-malus systems are provided (Section 2). The main actuarial tools for the study and design of bonus-malus systems are reviewed (Section 3). In the discussions that follow, Krupa Subramanian outlines a model for analyzing market shares in a competitive environment, a crucial research topic given current deregulation trends, and Pierre Lemaire compares actuarial with regulatory approaches to bonus-malus.
1. INTRODUCTION
In most developed countries, insurers use some form of merit-rating, in addition to other classication variables, in automobile third-party liability insurance. In the U.S. and Canada, insurance companies tend to use many a priori variables, such as age, sex, marital status and driving experience of the policyholder, car model, use of car, county of residence, and so on. Compared to other countries, the U.S. uses a posteriori or merit-rating to a limited extent: in many states and provinces, at-fault accidents and moving trafc violations are translated into penalty points and lead to a premium surcharge for three years. In other countries, insurance carriers usually rely on far fewer rating variables. For instance, in Switzerland, until recently insurers were authorized to use only one variable, the power of the engine, with more than 70% of the policies clustered in one single cell. In these countries, insurers rely on a much more sophisticated form of merit-rating, called bonus malus. Bonus-malus systems (BMSs) were introduced in Europe in the early 1960s, following the seminal works of Delaporte (1965), Bichsel (1964), and Bu hlmann (1964). The very rst ASTIN Colloquium, held in La Baule, France, in 1959, was devoted exclusively
*Jean Lemaire, A.S.A., Ph.D., is Chairperson of the Insurance and Risk Management Department at The Wharton School, University of Pennsylvania, Philadelphia, Pennsylvania 19104.
to bonus-malus. According to ASTIN legend, when Ge ne ral De Gaulle became President of France in 1958, he ordered French companies to introduce BMSs in automobile insurance. French actuaries then convened the rst ASTIN meeting. There exists a vast literature on BMSs in actuarial journals, mainly in the ASTIN Bulletin and the Swiss Actuarial Journal. A recent book (Lemaire 1995) summarizes this literature and provides more than 140 references and the complete description of 31 systems. The goal of this paper is to introduce the main ideas underlying the design of a BMS to North American actuaries and to present some recent research developments. BMSs are not commonly used in North America. This may change in the future because: A posteriori rating is a very efcient way of classifying policyholders into cells according to their risk. Several studies have shown that, if insurers are allowed to use only one rating variable, it should be some form of merit-rating (for instance, Lemaire 1985, ch. 7). The best predictor of the number of claims of a driver in the future is not age, car, or the township of residence, but past claims behavior. Several variables commonly used in North America, such as age, sex, and territory, are subject to intense scrutiny. Regulatory authorities may prohibit insurers from using some of these in the future, thereby forcing the insurance industry to rely more on merit-rating. In Massachusetts, where insurers do not use age, sex, and marital status, a Safe Driver
26
BONUS-MALUS SYSTEMS: THE EUROPEAN
AND
ASIAN APPROACH
TO
MERIT-RATING
27
Insurance Plan that is essentially a BMS was introduced in 1990.
2. DEFINITION
AND
EXAMPLES
By denition, an insurance company uses a BMS when The insureds of a given tariff group can be partitioned into a nite number of classes, denoted Ci or simply i (i1, . . . , s), so that the annual premium depends only on the class (s denotes the number of classes) and on the tariff group Policyholders begin their driving career in a specied starting class Ci0 An insureds class for a given period of insurance (usually a year) is determined uniquely by the class for the preceding period and the number of claims reported during the period. Such a system is determined by three elements: The premium scale b(b1, . . . , bs) The initial class Ci0 The transition rulesthe rules that determine the transfer from one class to another when the number of claims is known. These rules can be introduced as transformations Tk, such that Tk(i)j if the policy is transferred from class Ci into class Cj when k claims have been reported. The term Tk can be written in the form of a matrix Tk (ti j(k)), where t(k) i j 1 if Tk(i)j and 0 otherwise. The probability pij() of a policy moving from Ci into Cj in one period, for a policyholder characterized by some parameter (for instance, claim frequency), is equal to pi j()
k0
p () t
k
(k) ij
where pk() is the probability that a driver with claim frequency has k claims in a year. Obviously pij()0 and
j1
p () 1.
s ij
The matrix M () [pi j()]
k0
p () T
k
An insured enters the system, in the initial class, when he or she obtains a driving license. Then, throughout the entire driving lifetime, the transition rules are applied upon each renewal to determine the new class as a function of claims history. If a policyholder decides to switch to a new carrier, he or she has to obtain a certicate from the former company specifying the current BMS class and recent claims that could inuence the new class. Usually only the number of claims at-fault in thirdparty liability is used to penalize policyholders. Trafc violations are not taken into consideration. Korea is the only country where transition rules depend on claim severity, with property-damage losses subdivided into two categories and bodily injury claims into 14.1 Regulatory environments in European and Asian countries are extremely diversied, from total freedom to government-imposed systems, with many intermediate situations. The approach to BMS design depends on regulation. If a tariff is imposed by the government and every insurer has to use it, there is no commercial pressure to match the premiums to the risks by making use of all available relevant information. Supervising authorities may choose, for sociopolitical reasons, to exclude from the tariff structure certain risk factors, even though they may be signicantly correlated to losses. The government may then seek to correct for the inadequacies of the a priori system by using a tough BMS that penalizes claims more heavily. In a free market, carriers need to use a rating structure that matches the premiums to the risks as closely as possible, or at least as closely as the rating structure used by competitors. This entails using virtually every available classication variable correlated to the risks, because failing to do so would mean sacricing the chance to select against competitors and incurring the risk of suffering adverse selection by them. The use of more classication variables is expected in free market countries, which decreases the need for a sophisticated BMS. Some countries have adopted very simple BMSs. In Brazil, for instance, policyholders are subdivided into just seven classes, with premium levels 100, 90, 85, 80, 75, 70, and 65. New policyholders have to start in class 7, at level 100. Each claim-free year results in a one-class discount. Each at-fault claim is penalized by one class. The transition rules are presented in Table 1.
is the transition matrix of this Markov chain.

1
Japan had a rule that each bodily injury loss was penalized as two property-damage claims, but recently abandoned it.
28
NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 1
TABLE 1 BRAZILIAN SYSTEM (STARTING
CLASS:
7)
Class After 0 Class 7 6 5 4 3 2 1 Premium 100 90 85 80 75 70 65 Claims 6 5 4 3 2 1 1 1 Claims 7 7 6 5 4 3 2 2 Claims 7 7 7 6 5 4 3 3 Claims 7 7 7 7 6 5 4 4 Claims 7 7 7 7 7 6 5 5 Claims 7 7 7 7 7 7 6 6 Claims 7 7 7 7 7 7 7
Belgium recently adopted a more sophisticated system, with 23 classes. Pleasure-users and commuters enter the system in class 11, at level 85. Businessusers enter in class 14, at level 100. Each claim-free year leads to a one-class discount. The rst claim in any policy year is penalized by four classes. Any subsequent claim in the same year results in a ve-class penalty. Table 2 presents the transition rules of this system. The preceding denition assumes that the BMS forms a Markov chain process. A (rst-order) Markov
chain is a stochastic process in which the future development depends only on the present state but not on the history of the process or the manner in which the present state was reached. It is a process without memory, such that the states of the chain are the different BMS classes. The knowledge of the present class and the number of claims for the year sufce to determine next years class. It is not necessary to know how the policy reached the current class. In fact, the Belgian system does not form a Markov process. In addition to the rules mentioned above,
TABLE 2 BELGIAN SYSTEM
Class After 0 Class 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 Premium 200 160 140 130 123 117 111 105 100 95 90 85 81 77 73 69 66 63 60 57 54 54 54 Claims 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 0 1 Claims 22 22 22 22 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 2 Claims 22 22 22 22 22 22 22 22 22 22 21 20 19 18 17 16 15 14 13 12 11 10 9 3 Claims 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 21 20 19 18 17 16 15 14 4 Claims 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 21 20 19 5 Claims 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22
Starting class: 11 or 14
AND
ASIAN APPROACH
TO
MERIT-RATING
29
Belgian regulatory authorities have added a special transition rule that no policy can be in a class above 14 after four consecutive claim-free years. This last restriction is a concession to youthful operators who have many accidents in their early years and who suddenly improve. Very few policyholders are ever able to take advantage of this rule. Yet it makes the BMS non-Markovian. It forces insurance companies to memorize the claims history of some policyholders for four years, instead of simply the present class, had this restriction not been allowed. Indeed, after a claim-free year, a Belgian customer in class 17 will be
sent to class 14 or 16, depending on the number of consecutive claim-free years earned before. Fortunately, it is possible to modify the presentation of the system into a Markovian way, at the price of an increase of the number of classes. Classes are subdivided by adding an index that counts the number of consecutive claim-free years. In Markov chain terminology, the state variable is augmented with sufcient information so that a Markovian analysis is possible. Table 3 provides this extended presentation. The modied BMS requires 35 classes, up from 23 initially.
TABLE 3 BELGIAN BMS: MARKOVIAN PRESENTATION
Class After 0 Class 22 21.0 21.1 20.0 20.1 20.2 19.0 19.1 19.2 19.3 18.0 18.1 18.2 18.3 17 17.2 17.3 16 16.3 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 Premium 200 160 160 140 140 140 130 130 130 130 123 123 123 123 117 117 117 111 111 105 100 95 90 85 81 77 73 69 66 63 60 57 54 54 54 Claims 21.1 20.1 20.2 19.1 19.2 19.3 18.1 18.2 18.3 14 17 17.2 17.3 14 16 16.3 14 15 14 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 0 1 Claims 22 22 22 22 22 22 22 22 22 22 22 22 22 22 21.0 21.0 21.0 20.0 20.0 19.0 18.0 17 16 15 14 13 12 11 10 9 8 7 6 5 4 2 Claims 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 21.0 20.0 19.0 18.0 17 16 15 14 13 12 11 10 9 3 Claims 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 21.0 20.0 19.0 18.0 17 16 15 14 4 Claims 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 22 21.0 20.0 19.0
30
3. TOOLS FOR THE DESIGN AND EVALUATION OF BONUS-MALUS SYSTEMS

BMSs can be analyzed from the perspective of the policyholder or the insurance carrier. The tools are the same, but assumptions about probability distributions for the number of claims vary. For instance, if the Poisson distribution is acceptable to model the number of losses of an individual policyholder, differences among drivers make it inadequate to represent loss counts for an insurer. Distributions such as the negative binomial and the Poisson-Inverse Gaussian systematically outperform the Poisson to t observed portfolio loss counts. The policyholder is emphasized in this paper. It is assumed that the distribution {pk; k0, 1, 2, . . .} of the number of claims of a specic driver conforms to a Poisson with parameter , where is called the claim frequency of the policyholder and is assumed to be constant over time. pk e
nine classes. While statistically perfectly justied, such penalties seem commercially impossible to enforce. A forecast of the future distribution of policies among the classes, say n years from now, can be obtained easily through simulation or by computing the n-th power of the transition matrix M(). For many purposes, an asymptotic study is sufcient to compare BMSs. Only steady-state results are outlined in this paper. By using the Kemeny and Snell (1960) terminology, a BMS forms a regular Markov chain: all its states are ergodic (it is possible to go from every state to every other state), and the chain is not cyclic. In that case the value 1 is a simple eigenvalue of the transition matrix M(). The corresponding left eigenvector, row vector a () [a1(), a2(), . . . , as()] dened by the equation a () a () M (), and
k!
This section reviews the major tools that actuaries use to design, evaluate, and compare BMSs.
i1
a () 1
s i
3.1. The Relative Stationary Average Level

The relative stationary average level (RSAL) measures the position of the average driver, on a scale from zero to one, once the BMS has reached its steady-state condition. It evaluates the degree of concentration of policies in the lower classes of the BMS. An apparently inescapable consequence of the implementation of a BMS is a progressive decrease of the observed average premium level because of a clustering of the policies in the high-discount classes. For instance, the Belgian system penalizes accidents by four or ve classes and awards a one-class discount per claim-free year. The average observed claim frequency in the country is, however, close to 10%. Consequently, in any given year, the total number of bonus classes awarded to claim-free policyholders is much larger than the number of malus classes given to drivers with claims, and the mean premium level decreases. After a few years, the majority of policies is concentrated in the lowest classes of the BMS. The insurance carrier is then forced to compensate for this decrease by increasing the dollar cost of the premium at level 100, which somewhat defeats the purpose of the system. To maintain the BMS in nancial equilibrium without raising the basic premium, it would be necessary to penalize each claim by eight or
is the stationary probability distribution. The term ai() is the limit value for the probability that the policy is in class Ci, when the number of periods tends to innity. It is also the fraction of the time a policyholder with claim frequency spends in class Ci, once stationarity has been reached. As an illustration, let us compute the stationary distribution for the Brazilian system, when 0.1. For this BMS the transition matrix is (with class 7 at the top): 1p0 1p0 1pi 1pi 1pi 1pi 1pi p0 0 p1 p2 p3 p4 p5 0 p0 0 p1 p2 p3 p4 0 0 0 0 0 0 0 0 p0 0 0 0 0 p0 0 0 p1 0 p0 0 p2 p1 0 p0 p3 p2 p1 p0
M ()
where the lower bound of pi in each row is such that all row probabilities add up to 1. For 0.10, the left eigenvector of this matrix has the components: a1 0.88948; a2 0.09355; a3 0.01444; a4 0.00215; a5 0.00032; a6 0.00005; a7 0.00001. It represents the asymptotic distribution of policies among the seven classes of the system. It indicates,
AND
ASIAN APPROACH
TO
MERIT-RATING
31
for instance, that 89% of the policyholders with 0.10 will eventually belong to class 1. Multiplying each probability by the corresponding premium level, the mean asymptotic premium level is found to be 65.65, an extremely low value easily explained by the soft transition rules of that BMS. Figure 1 presents the evolution of the mean premium level for ve systems. For a simple system like the Taiwanese, the premium decreases abruptly in the rst few years, the time it takes for the best policyholders to reach the largest discount. The system then stabilizes rapidly. For the more sophisticated systems, the premium decreases in a much smoother way, and the steady state is not reached until more than 30 years has elapsed. This is a very long time to stabilize, given the short driving lifetime of the average policyholder: because most insureds do not drive for more than 60 years, a period of 30 years to stabilize their premium around their true level seems excessive. Given the wide variety of systems in force, stationary average levels are difcult to compare. Therefore, the relative stationary average level is dened as RSAL stationary average level minimum level . maximum level minimum level
Expressed as a percentage, this index determines the relative position of the average policyholder, when the
lowest premium is set equal to zero and the highest to 100. A low value of RSAL indicates a high clustering of policies in the high-discount BMS classes. A high RSAL suggests a better spread of policies among classes. The RSAL is only a crude measure of the severity of a BMS, because it is inuenced by the premium level of the sparsely populated highest class. Most systems lead to a very low RSAL value. Only Kenya, Spain, and Malaysia-Singapore use BMSs with a RSAL higher than 20%. For the ve systems of Figure 1, the RSAL is less than 10%. Note that the RSAL is measured here assuming no entries or exits. In an open portfolio, the RSAL would be somewhat higher, because exits are at lower levels than entries. A consequence of a low RSAL is that all BMSs carry an implicit penalty for new drivers, since the premium level of the access class is substantially higher than the average stationary premium level. The rstyear surcharge is easily computed as (entry premium stationary premium)/stationary premium. BMSinduced surcharges range from a low of 26.95% to a high of 212.97% in Germany. In addition to these implicit increases, many countries have introduced explicit penalties for inexperienced drivers, for example, a surcharge in France and a deductible after a claim in Belgium and Switzerland.
EVOLUTION
OF
FIGURE 1 MEAN PREMIUM LEVEL
32
3.2. The Coefcient of Variation of the Insureds Premium

Insurance consists of a transfer of risk from the policyholder to the carrier. Without experience rating, the transfer is total (perfect solidarity): the variability of insureds payments is zero. With experience rating, personalized premiums vary from year to year, according to claims history; cooperation between drivers is weakened. Solidarity between insureds can be evaluated by a measure of the variability of annual premiums. The coefcient of variation (standard deviation divided by mean) was selected, because it is a dimensionless parameter. There is thus no need for currency conversions. The Actuarial Institute of the Republic of China kindly provided marketwide observed loss severity distributions for property damage and bodily injury for accident years 1987 to 1989. These distributions are very well represented by a lognormal model (Lemaire 1993). Assuming that the aggregate claims process is compound Poisson with lognormal severities (Bowers et al. 1986, ch. 11), the coefcient of variation of aggregate losses is found to average 6.40. Although loss distributions in other countries of course differ from the Taiwanese experience, the coefcient
of variation is not likely to be affected much. So, without insurance, the policyholder is exposed to a loss process with a coefcient of variation of 6.40. With full insurance and no a posteriori rating, the variability of premiums is zero. With a BMS, the coefcient of variation will be somewhere between 0 and 6.40. The tougher the BMS, the larger this coefcient. Figure 2 shows the evolution of the coefcient of variation with time, for a policyholder with claim frequency 0.10, for the ve selected systems. Typically, the coefcient of variation starts at zero for the rst policy year, increases until the best policyholders reach the maximum discount, and then decreases until stationarity is reached, in some cases after more than 30 years. A common criticism against BMSs is that premiums paid by policyholders are too variable. This argument seems unwarranted. The highest variability for 0.10 is achieved by the new Swiss system, with an asymptotic coefcient of variation of 0.4595. This represents only 7.18% of the variation of the loss process. Even with a severe system such as the Swiss, policyholders are requested to assume only a small part of the risk. Figure 3 shows the coefcient of variation, when stationarity has been reached, as a function of the
EVOLUTION
OF
FIGURE 2 COEFFICIENT
OF
VARIATION
AND
ASIAN APPROACH
TO
MERIT-RATING
33
COEFFICIENT
OF
VARIATION
FIGURE 3 AS A FUNCTION
OF
CLAIM FREQUENCY
claim frequency. The variability is zero for the policyholders who never cause a claim, because they will permanently stay in the best class. Similarly, the coefcient of variation tends to zero as . With the exception of the Taiwanese system, the maximum variation occurs for claim frequencies that are significantly higher than the benchmark 10%.
3.3 The Elasticity of the Mean Stationary Premium with Respect to the Claim Frequency
The main goal of a BMS is to reduce the premium for good drivers and to increase it for bad risks. The random variables number of claims and claim amount are often assumed to be independent. This assumption essentially states that the cost of an accident is for the most part beyond the control of a policyholder. The degree of care exercised by a driver mostly inuences the number of accidents, but in a much lesser way the cost of these accidents. The validity of the independence assumption has been questioned [for instance, by Lemaire (1985, ch. 5)]. Nevertheless, the fact that nearly all BMSs in force around the world penalize the number of claims, independently of their amount, is an indication that
insurers and regulators accept it, at least as an approximation. It has the important implication that the risk of each driver can be measured by individual claim frequency . The elasticity of a BMS measures the response of the system to a change in the claim frequency. Obviously, for any BMS, the lifetime premiums paid by policyholders have to be an increasing function of . Ideally, this dependence should be linear. A relative increase of the claim frequency should produce the same relative increase of the premium. To provide an intuitive description of the concept, consider two policyholders, one with a claim frequency of 0.10, the other with a of 0.11. Over a long time, the second driver should pay 10% more premiums than the rst. A BMS with this property is called perfectly elastic. In practice, however, the mean premium increase in most cases is much lower than 10%. If the increase is, say, 2% instead of 10%, the systems elasticity is said to be 20%. More rigorously, denote P() the mean stationary premium associated with a claim frequency . Ideally, an increment d/ of the claim frequency should lead to an equal change, dP()/P(), of the premium. A BMS is called perfectly elastic if
34
d / 1. dP ()/ P () As a general rule, however, the change in premium is less than the change in . The elasticity () of the BMS is dened as () dP ()/P () d ln P () . d / d ln
cross-subsidization between the risk classes. For all systems in force, the good risks still subsidize the poor drivers.
3.4 The Average Optimal Retention

A side effect of BMSs is a tendency of policyholders to pay small claims themselves and not report them to their carrier to avoid future premium increases. In some countries the existence of this phenomenon, called the hunger for bonus, has been explicitly recognized by regulators. In Germany, for instance, the policy wording species that, if the insured reimburses the carrier for the cost of the claim, the penalty will not be applied. Moreover, if the claim amount does not exceed DM 1,000, the company is required to inform the policyholder of the right to reimburse the loss. If the claim amounts borne by policyholders are reasonable, the hunger for the bonus effect amounts to a small deductible and reduces administrative expenses. However, if a BMS induces drivers to pay claims in excess of, say, $5,000, it denitely penalizes claims excessively and may create hit-and-run behaviors. The main objective of a BMS is to achieve a
It is the elasticity of the mean stationary premium with respect to the claim frequency. This concept was introduced in actuarial science, under the name of efciency, by Loimaranta (1972). Ideally, the elasticity should be close to 1 for the most common values of . Figure 4 shows the elasticity of the selected systems as a function of . For the most common values of (0.15), () is very low. The elasticity reaches the neighborhood of 1 only for unusual values of . The Swiss system presents a rare case of overelasticity: ()1 for [0.150.28]. Only the new Swiss and Finnish systems exhibit an elasticity of more than 0.40 for 0.10. The low observed elasticities, along with the low coefcients of variation, indicate that the implementation of a BMS decreases, but does not eliminate,
FIGURE 4 ELASTICITY
AND
ASIAN APPROACH
TO
MERIT-RATING
35
better separation of the good and the bad risks; it is not to transfer most claims from the insurer to the insureds. The determination of the optimal strategy of the policyholder is an interesting decision problem that has close links with innite-horizon dynamic programming under uncertainty. For each class of the system, the following algorithm computes the optimal retention levelthe level under which it is in the policyholders interest not to report a claim. To simplify the presentation, retentions are computed here under an innite-horizon assumption. Policyholders are supposed to drive forever. While this is obviously not the case in practice, calculations show that the effect of this assumption for long driving careers is minuscule. The decision problem can be formulated as follows. Step 1. Evaluation of Given Strategies Dene a strategy for the policyholder as a vector x (x1, . . . , xi), where xi is the retention limit for class Ci. The cost of any accident of amount less than or equal to xi is borne by the policyholder. Claims of higher amounts are reported. Consider a policyholder who has just caused an accident of amount x, at time t, with time units such that 0t1. Denote f(x) the density function of the random variable X representing the cost of a claim. Given a specic strategy, the probability pi of a claim not being reported by a policyholder in class Ci is pi P (X xi)
The expectation of the total cost for this period is E(xi) bi 1/2 Ei(X) ( i), introducing a discount factor and assuming all claims take place in the middle of the period. The vector v () [v1(), . . . , vs()] of the discounted expectation of all the payments by the policyholder satises the set of s equations with s unknowns vi() (i1, . . . , s): vi() E(xi)
k0
p () v
i k
Tk(i)
().
This system always has a unique solution. The vector v () of solutions to the system provides a numerical evaluation of the cost of every single possible strategy. Step 2. Determination of the Optimal Strategy The policyholder who causes a claim of amount x at time t has two possible courses of action: (1) if he or she does not report the accident, the expectation of total cost, discounted at the time of the claim, is t E(xi) x 1t
k0
p [(1 t)] v
i k
Tkm(i)
(),
xi
f (x) dx.
where m is the number of claims already reported during the period; or (2) if the accident is reported to the company, the expectation is equal to t E(xi) 1t
The probability p ik() of reporting k claims during one period equals p ik()
hk
k0
p [(1 t)] v
i k
Tkm1(i)
().
(1 p ) p () h k
h i
p ihk.
The retention limit xi is the claim cost x for which the two actions are equivalent: xi 1t
The expectation of the number of reported claims is i
k0
k p ().
i k
k0
p [(1 t)] [v
i k
Tkm1(i)
() vTkm(i)()].
The expected cost of a nonreported claim is equal to E i(X) 1 pi
xi
xf (x) dx.
Assuming independence between the number and the amount of claims, the policyholder pays, on average, for the compensation of nonreported claims E i(X) ( i).
These equations constitute a set of s equations with s unknowns xi, as the xi appear in an implicit way in the p ik[(1t)]. The system provides a new strategy vector x , corresponding to a cost vector v (). Summarizing, the rst system provides v () [v1(), . . . , v s()], for given x (x1, . . . , xs). The second system provides x , for given v (). Taken together, the two systems consist in 2s equations with 2s unknowns, with, as solution, the optimal strategy x *(x* *(). The 1 , . . . , x* s ) and its associated cost v system is best solved by successive approximations.
36
Application The optimal strategy depends on the class Ci, the discount factor , the claim frequency , the time of the accident t, and the number of claims already reported m. The optimal strategy is an increasing function of t: the retention increases, as policy renewal (and the resulting likely discount) nears. However, the dependence of x on t is quite small, compared to the inuences of Ci, , and . Setting t0 (and consequently m0) simplies all calculations and hardly reduces optimal retentions. The algorithm described above was applied to the Belgian BMS, rst assuming a claim frequency 0.10 and a discount coefcient 0.90. The 1989 observed property damage loss distribution in Taiwan was used, with some allowance for ination and for the fact that Taiwanese losses may be slightly below world averages.2 Table 4 presents the optimal retention, for all classes, as a percentage of the average premium. Remember that some classes had to be subdivided to obtain a Markovian presentation of the system (Table 3). Given the toughness of the BMS, optimal retentions are very high. With the exception of a handful of classes (the most populated, however), it is always in a policyholders interest to bear the cost of an accident that is more expensive than the average premium.
TABLE 4 OPTIMAL RETENTIONS: BELGIAN BMS
Optimal Class 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 16.3 17 Retention 38.41% 56.50 76.59 98.26 117.80 137.34 156.05 174.03 190.40 208.83 224.98 239.38 254.56 273.65 285.46 269.02 254.05 305.99 252.17 Class 17.2 17.3 18.0 18.1 18.2 18.3 19.0 19.1 19.2 19.3 20.0 20.1 20.2 21.0 21.1 22
Optimal Retention 296.85% 360.03 288.16 326.98 382.01 457.52 257.56 304.64 369.69 457.52 228.29 283.74 359.28 196.03 260.11 147.31
The system has a special rule, that no policyholder can be in the malus zone after four consecutive claimfree years. The impact of the special transition rule on optimal retentions is evidenced in Table 4; a driver in class 18.0 (who had an accident last year) has an optimal retention of 288.16% of the average premium. This retention increases to 457.52% for an insured in class 18 with three claim-free years. The retentions in Table 4 depend on the values taken by two crucial parameters: , the claim frequency, and , the discount factor. Because the values taken by these parameters are difcult to estimate accurately by the policyholder, it is of crucial importance to study the variability of x* in terms of or . Figure 5 shows, for ve different classes of the Belgian BMS, the optimal xi as a function of , with given discount factor 0.9. For values of the claim frequency in excess of 1, x* i rapidly drops to zero, because it does not pay to indemnify claims for someone who expects several claims per year. Note the low slope of these curves: the optimal strategy is not inuenced much by a change in . A slight error in the estimation of has only minor consequences. Figure 6 shows the variation of the optimal retention with the discount factor, for constant 0.10, for ve selected classes. Obviously, all curves are increasing. A policyholder who discounts future payments a great deal does not have much interest in paying claims. For large discount factors, the slopes of the curves are rather large. A 1% error in the determination of the discount factor has a much larger impact than a 1% error in the claim frequency.
3.5 The Rate of Convergence of BonusMalus Systems

The rst part of this section demonstrates that simple BMSs like the Taiwanese reach stationarity much faster than sophisticated systems like the Swiss. An evaluation of the rate of convergence of BMSs to their steady-state condition is of great importance, because many of the tools dened here assume that stationarity has been reached. Let pij() be the transition probabilities of the Markov chain associated with each BMS, and pn ij () the nstep transition probabilities. The term pn ij () is the probability of moving from class Ci to class Cj in exactly n transitions. These probabilities are obtained by computing the n-th power of the transition matrix M(). Let aj(), j1, . . . , n, be the stationary distribution. Consider a specic BMS class Ci (usually the
Other applications of the algorithm have shown optimal strategies to be quite insensitive to the loss distribution.
AND
ASIAN APPROACH
TO
MERIT-RATING
37
OPTIMAL RETENTION
AS A
FIGURE 5 FUNCTION
OF
CLAIM FREQUENCY
OPTIMAL RETENTION
AS A
FIGURE 6 FUNCTION OF
THE
DISCOUNT FACTOR
38
starting class of the system), a given , and a given number of steps n. Following Bonsdorff (1992), dene the total variation (TV )n
REFERENCES
BICHSEL, F. 1964. Erfahrungs-Tarierung in der Motorfahrzeughalfplichtversicherung, Mitteilungen der Vereinigung Schweizerischer Versicherungsmathematiker, 119129. BONSDORFF, H. 1992. On the Convergence Rate of Bonus-Malus Systems, ASTIN Bulletin 22:217223. BOWERS, N., HICKMAN, J., GERBER, H., JONES, D., AND NESBITT, C. 1986. Actuarial Mathematics. Itasca, Ill.: Society of Actuaries. HLMANN, H. 1964. Optimale Pra BU mienstufensysteme, Mitteilungen der Vereinigung Schweizerischer Versicherungsmathematiker, 193213. DELAPORTE, P. 1965. Tarication du risque individuel daccidents par la prime modele e sur le risque, ASTIN Bulletin 3:251271. KEMENY, J.G., AND SNELL, J.L. 1960. Finite Markov Chains. New York: Van Nostrand Reinhold. LEMAIRE, J. 1985. Automobile Insurance: Actuarial Models. Boston: Kluwer. LEMAIRE, J. 1993. Selecting a Fitting Distribution for Taiwanese Automobile Losses, unpublished manuscript. LEMAIRE, J. 1995. Bonus-Malus Systems in Automobile Insurance. Boston: Kluwer. LOIMARANTA, K. 1972. Some Asymptotic Properties of Bonus Systems, ASTIN Bulletin 6:233245.
j1
| p () a () |
s n ij j
as a measure of the degree of convergence of the system after n transitions. For any two probability distributions, the total variation is always between 0 and 2. The term (TV)n can prove to be a very useful tool for selecting factors that affect the pace to stationarity, such as the number of classes, the starting class, or the transition rules. Table 5 shows the total variation for four systems, when 0.10 and Ci is the starting class of the BMS, as a function of time. It shows that sophisticated BMSs converge extremely slowly. While the Taiwanese system reaches full stationarity after only three years, there still exists signicant variability after 30 years for other systems. The total variation for the Belgian BMS is still more than 20% of its original value. Even after 60 years, it still has not fully stabilized! This is a drawback of bonus-malus rating. The main objective of bonus-malus rating is to correct the inadequacies of a priori rating by separating the good from the bad drivers. This separation process should proceed as fast as possible. A 30-year separation phasealthough justied by the low overall claim frequency and the resulting inherent variability of eventscan be considered as excessive, because it exceeds half the driving lifetime of the majority of policyholders. It also exceeds the life expectancy of all BMSs. In fact, no single BMS has in the past been allowed to survive 30 years and to reach stationarity. All the countries that adopted bonus-malus rating in the late 1950s or early 1960s have switched to a second-generation BMS.
TABLE 5 TOTAL VARIATION FOR FOUR SYSTEMS Years 0 10 20 30 60 Belgium 1.9913 1.7769 0.9120 0.4209 0.0382 Japan 1.9950 1.1551 0.3217 0.0529 0.0007 Taiwan 2.000 0 0 0 0 Switzerland 1.9742 1.0124 0.3541 0.1348 0.0061
DISCUSSIONS KRUPA SUBRAMANIAN* BONUS-MALUS SYSTEMS IN A COMPETITIVE ENVIRONMENT

In this discussion, a simple model is developed to analyze the evolution of market shares and the nancial stability of two insurers when one of them adopts an aggressive competitive behavior by modifying its bonus-malus system. The two scenarios that are developed show that rating freedom encourages insurers to adopt tougher systems. Until a few years ago, in all but a handful of countries, all companies had to use the same BMS.1 This situation is changing very fast. Deregulation ideas are gaining ground in Asia. European Economic Community directives have introduced complete rating
*Krupa Subramanian, A.S.A., is a doctoral student at The Wharton School, University of Pennsylvania, Philadelphia, Pennsylvania 19104. 1 A notable exception is the United Kingdom, where complete rating freedom has existed for many years. Drivers have always been able to choose among many different (but similar) systems.
AND
ASIAN APPROACH
TO
MERIT-RATING
39
freedom in all 15 member nations, as of July 1, 1994. Insurers are free to set up their own rates, select their own classication variables, and design their own BMS. In most countries, companies have taken advantage of this freedom by introducing more rating variables. An increased segmentation of portfolios has resulted. Competition based on aggressive design of BMS seems to have been limited to Italy and Portugal; Belgium, France, Finland, Sweden, and Greece maintain a uniform BMS. In several countries, the insurance market is challenging the idea that full rating freedom means freedom for companies to design their own BMS and is attempting to maintain a unique BMS. A probable reason for this choice of weapons in competition is policyholder information, or lack thereof, in automobile insurance. The well-known assertion by Joskow (1973), There is no other product for which consumer ignorance is so prevalent, is still applicable today. While most European drivers know that their premium depends on their car model and that they are subject to some form of merit-rating, only a minuscule minority knows its BMS premium level and the number of penalty classes in case of an accident at-fault. It is much easier for insurance carriers to advertise immediate discounts for suburban female drivers than to explain that a more efcient BMS will, in the long run, lead to decreased premiums for better drivers. Another possible reason for the lack of creativity in BMS design in most countries is insurer uncertainty about the outcome of unbridled competition. Insurers do not know whether the introduction of a more efcient BMS will attract good or bad drivers or result in an increased or decreased market share. The results of the simple model presented here seem to indicate that rating freedom should induce insurers to adopt severe BMSs. Two scenarios are developed. In each, the carrier using the mildest system is progressively eliminated from the market.
TABLE 1 INITIAL BONUS-MALUS SYSTEM Premium Level 200 160 130 100 90 80 70 60 50
Class 9 8 7 6 5 4 3 2 1
Rules Starting Class: 6 Discount per claim-free year: 1 class Penalty per claim: 3 classes
The distribution of the number of claims of each policyholder follows a Poisson with parameter . The distribution of in the portfolio is assumed to be a gamma; hence the portfolio distribution of loss counts is a negative binomial. Following Lemaire (1995), assume that the mean of the negative binomial is 0.10 and its variance 0.107, corresponding to a typical accident pattern of European drivers. For computational simplicity, the gamma has been discretized. Ten values of , presented in Table 2, have been selected so that their mean and variance correspond to 0.10 and 0.107, respectively. Further assume that the mean cost of an accident is $10,000 and that all expense and prot loadings have been incorporated in this gure.
TABLE 2 SELECTED VALUES
OF
Group 1 2 3 4 5 6 7 8 9 10
0.01081 0.02536 0.03918 0.05363 0.06952 0.08775 0.10972 0.13815 0.17982 0.28608
Percentile in the Gamma Distribution 5.01% 15.04 25.08 35.14 45.19 55.24 65.27 75.27 85.20 96.21
Basic Assumptions
Assume that only two insurance companies, A and B, operate in a given country. They both apply the same BMS, a nine-class system described in Table 1. They have used this system for such a long time that the steady-state condition has been attained. There are no entries or exits. The two insurers apply the same rates, so that each has a 50% market share. Each company has 10,000 policyholders and an initial surplus of $10,000,000.
Surveys of policyholders have consistently demonstrated some reluctance to switch insurers. In a survey of 2,462 policyholders by Cummins et al. (1974), 54% of respondents confessed never to have shopped around for auto insurance prices. To the question Which is the most important factor in your decision to buy insurance?, 40% responded the company, 29%
40
the agent, and only 27% the premium. A similar survey of 2,004 Germans (Schlesinger and von der Schulenburg 1993) indicated that, despite the fact that 67% of those responding knew that considerable price differences exist between automobile insurers, only 35% chose their carrier on the basis of their favorable premium. Therefore we will assume that, given the opportunity to switch for a reduced premium, one-third of the policyholders will do so. Scenario 1: Company B Introduces Another Discount Class; Company A Does Not React Assume that, at time 0, rating freedom is introduced in the country and that all policies are renewed annually, at times 0, 1, 2, . . . . A few weeks before time 1, company B announces that it is introducing a new class, class 0, with a premium level of 40. All other levels and the transition rules remain unchanged. Company A decides not to change its system. Neither company changes its basic premiumthe dollar amount it charges at level 100at that time. Some of the class 1 policyholders of company A switch to
company B to take advantage of the extra discount. This results in an overall premium decrease in the market, and both companies lose money, B because of the discounted premiums, A because the deterioration of the quality of its portfolio (many of its best drivers have switched to B). Basic premiums are assumed to be adapted retrospectively, at time 2, such that the nancial balance is restored for the current portfolio. The rst-year loss is never recouped. The same process is repeated annually: some policyholders switch, resulting in modied market shares and in a gain or loss for each company. These oneyear gains or losses are never recouped, but each company adjusts premiums one year later, taking into account the new mean premium level and the new claim frequency. Income stability is, however, never achieved, because some insureds immediately take advantage of the new rate information to switch carriers. Table 3 summarizes the evolution of the portfolios over the rst 10 years. This scenario leads to a complete elimination of company A in a few years, as evidenced by Figure 1,
EVOLUTION
OF
PORTFOLIOS Market Avg. Level 63.86 63.86 59.58 59.58 58.60 58.60 57.86 57.86 57.30 57.30 56.91 56.91 56.52 56.52 56.15 56.15 55.83 55.83 55.56 55.56 55.35 55.35
FOR
TABLE 3 SCENARIO 1
OVER
FIRST TEN YEARS Next Premium Increase Premium Income (000s) $10,000 10,000 2.12% 11.27 4.07 0.17 3.69 (0.81) 3.06 (1.13) 2.41 0.82 1.65 1.21 1.13 1.10 0.88 0.86 0.66 0.68 0.24 (0.62) 8,324 10,335 8,969 10,648 9,537 10,193 9,956 9,851 6,679 13,053 4,486 15,256 3,006 16,775 2,009 17,820 1,342 18,522 896 19,003 Annual Prot (000s)
Time 0 1 2 3 4 5 6 7 8 9 10 Cumulated Prot
Company A B A B A B A B A B A B A B A B A B A B A B A B
No. of Policies 10,000 10,000 7,859 12,141 7,625 12,375 7,469 12,531 7,365 12,635 4,910 15,090 3,273 16,727 2,182 17,818 1,455 18,545 970 19,030 647 19,353 $(1,495) $(1,626)
Avg. Premium Level 63.86 63.86 67.63 54.36 73.55 49.39 76.73 46.61 78.34 45.04 76.49 50.54 75.25 52.85 74.41 53.91 73.77 54.42 73.28 54.65 72.91 54.76
Expected No. of Claims 1,000 1,000 850 1,150 933 1,067 989 1,011 1,026 974 684 1,316 456 1,544 304 1,696 203 1,797 135 1,865 90 1,910
Claim Frequency 10.00% 10.00 10.81 9.47 12.24 8.62 13.24 8.07 13.94 7.71 13.93 8.72 13.93 9.23 13.93 9.52 13.93 9.69 13.93 9.80 13.93 9.87
Claims (000s) $10,000 10,000 8,500 11,500 9,334 10,666 9,889 10,111 10,260 9,740 6,840 13,160 4,560 15,440 3,040 16,960 2,027 17,973 1,351 18,649 898 18,886
$ (176) (1,165) (365) (18) (352) 82 (304) 111 (161) (107) (74) (184) (34) (185) (18) (153) (9) (127) (2) 117
AND
ASIAN APPROACH
TO
MERIT-RATING
41
EVOLUTION
OF
FIGURE 1 MARKET SHARES
FOR
SCENARIO 1
which shows the evolution of the market shares in this scenario. At time 1, one-third of the claim-free drivers from As class 1 switch to Bs class 0. Company B gains 2,140 new policyholders in the process. Moreover, these are better-than-average drivers, because Bs claim frequency for year 2 drops from 10% to 9.47%. However, Bs average premium level decreases from the steady-state 63.86 to 54.36 because of the extra discount. Company B has to increase its basic premium rate by 11.27%. Company As average premium level has increased to 67.63, but its claim frequency is now 10.81%. Company As basic premium needs to be increased by 2.12%. As a result of these differentiated premium increases, Bs rates at time 2 are higher than As in all classes, except of course the new class 0. Consequently, while some claim-free policies from As class 1 switch to Bs class 0, all other switches go the other way: Company A receives new policyholders from Bs other classes. As a result, market shares change negligibly, but a selection process is beginning to take place, with the best policyholders concentrating in Bs portfolio. Company Bs claim frequency has now dropped to 8.62%, and As has increased to 12.24%. Company A has to increase its basic rate by 4.07%, while Bs rate hardly has to change.
A similar pattern of switches takes place at times 3 and 4. Company B attracts many of As good drivers and sends away some of its worst drivers. Market shares do not change signicantly, but the quality of As portfolio progressively deteriorates. Company A needs important annual price increases. The improvement in Bs portfolio allows a slight decrease in its premiums. At time 5, As premiums become higher than Bs in all classes. From that time on, one-third of As policyholders switch to B every year, and A progressively disappears from the market. At time 10, its market share is less than 3.5%. Company B has taken over the entire market and only has to pay a reasonable price, because its surplus has been depleted by only $1,626,000. Company As surplus has been reduced by nearly as much. Figure 2 shows the evolution of the claim frequencies in the two portfolios and further illustrates the two phases of this scenario. During phase 1 (up to time 4), B progressively improves its portfolio by attracting As best drivers and eliminating some of its worst drivers. Company A has to increase its rates so much that it becomes more expensive than B in all classes from time 5 on. Company A is then eliminated from the market during this second phase.
42
EVOLUTION
OF
FIGURE 2 CLAIM FREQUENCIES
FOR
SCENARIO 1
Scenario 2: Company A Reacts by Creating a Super-Discount Class Under scenario 1, company A passively watches its portfolio being taken over by B and never uses its surplus to engage in a competitive battle with B. In scenario 2, we assume that, at time t5, A creates a super-discount class 0, with a premium level 35. Table 4 provides the evolution of the two portfolios from time 5 on. Figure 3 shows that A manages not only to recover its former market shares but also eventually succeeds in eliminating B from the market. Recall that at time 5 under scenario 1, As premiums are higher than Bs across all classes. Introduction of a cheaper class 0 at that time allows A to attract just the class 0 drivers from B and retain its class 1 claim-free drivers, while one-third of all other drivers shift to B. Because of this super-discount class, As average premium level decreases to 53.99, necessitating a premium increase of 11.36% at time 6. Such an increase coupled with the negligible change in Bs premiums results in As rates being more expensive than Bs across all classes, including class 0. Thus, at time 6, B attracts one-third of As drivers from all classes. At time 7, due to a slight deterioration in Bs portfolio, its class 0 premium becomes higher than As,
while its premiums in the other 9 classes are cheaper. Thus, similar to time 5, A gains the better-than-average drivers, while B entices the poorer drivers away from A. This shift of policyholders is repeated each period until time 12. As As claim frequency steadily decreases, as illustrated in Figure 4, its premiums correspondingly decrease. From time 12 on, As rates are cheaper than Bs across all classes, which results in one-third of Bs policyholders switching companies each year, leading to an erosion of Bs market share. Conclusions Under the two scenarios, the most aggressive insurer eventually manages to drive its competitor out of the market. Table 5 provides the elasticity and the premium coefcient of variation of all three systems, when 0.10. The most severe BMS wins in all cases. The two scenarios seem to indicate that complete rating freedom will eventually result in the adoption of tougher systems and less solidarity among policyholders.
AND
ASIAN APPROACH
TO
MERIT-RATING
43
EVOLUTION
OF
FIGURE 3 MARKET SHARES
FOR
SCENARIO 2
EVOLUTION
OF
FIGURE 4 CLAIM FREQUENCIES
FOR
SCENARIO 2
44
EVOLUTION
OF
PORTFOLIOS Market Avg. Level 63.86 63.86 59.58 59.58 58.60 58.60 57.86 57.86 57.30 57.30 54.47 54.47 54.70 54.70 53.39 53.39 53.04 53.04 52.54 52.54 52.25 52.25 52.03 52.03 51.87 51.87 51.69 51.69 51.54 51.54 51.41 51.41
FOR
TABLE 4 SCENARIO 2
FROM
TIME 5
TO
TIME 15 Premium Income (000s) $10,000 10,000 Annual Prot (000s)
Time 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Cumulated Prot
Company A B A B A B A B A B A B A B A B A B A B A B A B A B A B A B A B A B
No. of Policies 10,000 10,000 7,859 12,141 7,625 12,375 7,469 12,531 7,365 12,635 9,322 10,678 6,215 13,785 8,585 11,415 10,174 9,826 11,296 8,704 12,049 7,951 12,553 7,447 15,035 4,965 16,690 3,310 17,793 2,207 18,529 1,471 $(2,041) $(2,846)
Avg. Premium Level 63.86 63.86 67.63 54.36 73.55 49.39 76.73 46.61 78.34 45.04 53.99 54.89 53.81 55.10 44.05 60.94 40.81 65.70 39.42 69.57 38.83 72.58 38.54 74.76 45.18 72.13 48.02 70.19 49.39 68.87 50.10 67.94
Expected No. of Claims 1,000 1,000 850 1,150 933 1,067 989 1,011 1,026 974 997 1,003 664 1,336 758 1,242 822 1,178 871 1,129 905 1,095 927 1,073 1,285 715 1,523 477 1,682 318 1,788 212
Claim Frequency 10.00% 10.00 10.81 9.47 12.24 8.62 13.24 8.07 13.94 7.71 10.70 9.39 10.68 9.69 8.83 10.88 8.08 11.99 7.71 12.97 7.51 13.77 7.38 14.41 8.55 14.40 9.13 14.41 9.45 14.41 9.65 14.41
Next Premium Increase
Claims (000s) $10,000 10,000 8,500 11,500 9,334 10,666 9,889 10,111 10,260 9,740 9,967 10,032 6,645 13,355 7,584 12,416 8,219 11,781 8,713 11,287 9,048 10,952 9,274 10,726 12,849 7,151 15,233 4,767 16,822 3,178 17,881 2,119
2.12 11.27 4.07 0.17 3.69 (0.81) 3.06 (1.13) 11.36 0.00 0.32 2.72 0.96 1.51 (1.31) 2.25 (1.14) 2.14 (1.18) 1.82 (0.88) 1.51 (1.34) 3.65 0.49 2.75 0.71 1.92 0.64 1.36
8,324 10,335 8,969 10,648 9,537 10,193 9,956 9,851 8,950 10,032 6,624 13,001 7,513 12,231 8,328 11,522 8,814 11,050 9,156 10,756 9,356 10,566 13,023 6,899 15,158 4,639 16,703 3,118 17,768 2,090
$(176) (1,165) (365) (18) (352) 82 (304) 111 (1,017) (21) (354) (72) (185) 109 (259) 100 (237) 108 (196) 82 (160) 174 (252) (75) (128) (119) (60) (113) (29)
ELASTICITY
AND
VARIABILITY
OF
TABLE 5 PREMIUMS
FOR THE
THREE SYSTEMS
REFERENCES
CUMMINS, D., MCGILL, D., WINKLEVOSS, H., AND ZELTEN, H. 1974. Consumer Attitudes Toward Auto and Homeowners Insurance. Philadelphia, Pa.: Department of Insurance, Wharton School. JOSKOW, P. 1973. Cartels, Competition and Regulation in the Property-Liability Insurance Industry, Bell Journal of Economics 4:375427. LEMAIRE, J. 1995. Bonus-Malus Systems in Automobile Insurance. Boston: Kluwer. SCHLESINGER, H., AND VON DER SCHULENBURG, J. M. GRAF 1993. Consumer Information and Decisions to Switch Insurers, Journal of Risk and Insurance LV, no. 4:591615.
System Initial B-revised A-revised
Elasticity 0.2736 0.3346 0.3948
Premium Coefcient of Variation 0.3627 0.4222 0.4889
AND
ASIAN APPROACH
TO
MERIT-RATING
45
PIERRE LEMAIRE*
Actuarial versus Regulatory Perspective
Jean Lemaire (1995) and Venter (1991) have provided actuarial toolssummarized in Section 3for the design and evaluation of BMSs. Ferreira (1974), in a study for the U.S. Department of Transportation, provided the perspective of an American regulator and proposed other ways to assess merit-rating plans, using drastically simplifying assumptions. His appraisal is based on cost redistribution, deterrence, nancial responsibility, and public acceptance. Ferreiras tools are generalized here to be applied to BMSs and are used to evaluate the systems in force in Denmark, Thailand, and Switzerland. They illustrate the profound differences between the actuarial and regulatory approaches to BMS design. While actuarial tools induce insurance carriers to develop severe BMSs, the regulatory approach promotes milder systems. This may explain why the design of a BMS always leads to delicate, lengthy negotiations between insurers and supervising authorities, and why all BMSs currently in force lead to some degree of subsidization of high-risk insureds by good drivers. Tools to Evaluate Bonus-Malus Systems Actuarial Tools Five major tools have been presented in the actuarial literature to evaluate BMSs: 1. Relative stationary average level 2. Coefcient of variation of the insureds premium 3. Elasticity of the mean stationary premium with respect to the claim frequency 4. Average optimal retention (all presented in Jean Lemaires paper) 5. Predictive accuracy, which is the mean square rating error. If the random variables number of losses and claim amount are independent, the risk of a policyholder is some multiple of claim frequency. The rating error can then be measured as the difference between that drivers mean stationary premium and the multiple of the claim frequency. Averaging the squared errors over the entire portfolio of the company provides the mean square error, a summary measure of
the overall rating (in)adequacy, which Venter calls the predictive accuracy of the BMS. Regulatory Tools Cost Redistribution. Ferreiras measure is essentially Venters predictive accuracy. This is the only concern that actuaries and regulators seem to share. Deterrence. This is a main objective of merit-rating. The threat of premium surcharges for accident involvement is expected to deter motorists from hazardous driving. While in many countries the introduction of a BMS has led to a decrease in overall claim frequencies, measuring this effect is an arduous task and has not been attempted. It is thus assumed that BMSs have a deterrent effect, but that the sensitivity of deterrence to a specic system is of second order. Financial Responsibility. Because uninsured motorists are a major problem in the U.S., Ferreira is concerned that BMSs provide incentives to forego insurance. If the premium charged following the application of the BMS transition rules is too high, some policyholders may be tempted to cancel their policy. Ferreira uses a very crude cancellation threshold model to measure nancial responsibility: he assumes that all policyholders drive uninsured if their premium exceeds $300. A better measure of the incentive to cancel is dened here as the percentage of drivers who are requested to pay more than 300% of the mean stationary premium. Public Acceptance. To measure the attractiveness of a merit-rating plan to the public, Ferreira uses the number of years that elapse before a particular accident no longer affects the premium. For most BMSs in force, any accident inuences the premium for the entire lifetime of the policyholder, albeit in a smaller and smaller way, as time goes on. For practical purposes, we assume that the effect of an accident has disappeared when the premium differential has become small. For each policyholder, the i-th row of the n-th power of the transition matrix provides the distribution in the classes of the BMS n years after being in class Ci. This property makes it easy to compute the average premium with and without the shock of an accident. The number of transitions until the premium difference is less than one premium level is used to measure the duration of the effect of an accident. Policyholders are then averaged by using the stationary distribution. This approach has the advantage that it incorporates the possibility of future accidents in calculations.
*Pierre Lemaire is a doctoral student at the Wharton School, University of Pennsylvania, Philadelphia, Pennsylvania 19104.
46
Comments on Tools Three of the ve actuarial measures promote tough systems, while two of the three regulators measures call for mild BMSs. Lemaire (1995) recommends 50% as an ideal value for the RSAL. This implies that the average policyholder will be situated at the mid-point of the premium levels. Only a very severe system, tougher than all BMSs currently in force, would achieve this ideal value. All 30 systems analyzed by Lemaire and Zi (1994) exhibit RSAL values less than 30%. Lemaire makes no specic recommendation about a satisfactory value for the coefcient of variation. Lemaire suggests an ideal value of 1 for the elasticity. None of the 30 systems has an elasticity exceeding 0.45 for a claim frequency of 0.10. So the elasticity criterion favors BMSs that are more severe than all systems in force. According to Lemaire, reasonable optimal retentions are desirable, because they amount to a small deductible: it may be preferable to handle small claims this way. Large retentions are not welcome, because they induce hit-and-run behaviors. While no specic limit is suggested, this criterion calls for moderate systems. Venter and Ferreira wish the BMS to have the best possible predictive accuracy and the smallest possible mean squared error. This can be achieved only by reducing solidarity by increasing the premiums for the worst drivers in order to award better discounts to the good risks. Tougher systems result, up to a point. Ferreira wishes to reduce the number of uninsured motorists. With the selected criterion, the percentage of policyholders paying more than 300% of the average premium has to be as small as possible. This calls for a mild system. According to Ferreira, BMSs are better accepted if the effect of an accident quickly disappears. This requires a mild BMS. Application All the measures described have been calculated for three BMSs: a sophisticated system (Switzerland), an average system (Denmark), and a very simple system (Thailand). These BMSs are fully described in Lemaire (1995). Lemaire and Zi (1994) have used principal components analysis to dene an Index of Toughness, a ranking of 30 different BMSs according
to their severity toward policyholders. The Swiss system places rst in this classication, the Danish BMS twelfth, and the Thai BMS twenty-second. The Swiss system has 22 classes, with premium levels ranging from 45 to 270. The starting level is 100, in class 10. Each claim-free year results in a one-class discount. Every single claim is penalized by four classes. The Danish BMS has 10 classes, with premium levels from 30 to 150, and a start in class 8, at level 100. Each claim-free year provides a one-class discount. Each claim is penalized by two classes. The Thai system has 7 classes, premiums levels from 60 to 140, and a start in class 4 at level 100. After a claim-free year, drivers in the malus zone return to the basic premium level 100. Other drivers gain a class. After one claim, all policyholders return to class 4. After two or more claims, policies are placed in different malus classes, depending on the preceding class. To calculate the Ferreira measures, it is necessary to model the entire portfolio of a company. We have adopted the same approach as Subramanian in the preceding discussion, discretizing a gamma distribution by subdividing the portfolio into ten. This represents an important step toward a realistic analysis, because Ferreira considered only a three-point distribution for . The results are summarized in Table 1. For tools 1, 2, 3, and 4, a policyholder with claim frequency 0.10 is considered. For the other tools, the entire company portfolio is used.
TABLE 1 SUMMARY OF RESULTS Tool Thailand Denmark Switzerland
RSAL 8.03% 3.78% 6.47% Coefcient of Variation 0.1925 0.3017 0.4595 Elasticity 0.081 0.165 0.449 Average Optimal Retention 6.20% 128.58% 186.03% (in % of average premium) Predictive Accuracy 2330.57 436.30 482.63 (mean square error) Financial Responsibility 0% 1.075% 5.927% Public Acceptance 3 12 42
The Swiss BMS ranks rst on most measures; the Thai system last. So the Swiss system would be recommended by actuaries but not by regulators. Supervisory authorities would very much prefer the Thai system, a BMS that would be unpopular with actuaries.
AND
ASIAN APPROACH
TO
MERIT-RATING
47
Conclusions The revision of a BMS always seems to lead to delicate and lengthy negotiations between insurers and supervisory authorities. The nal selection of a BMS is always a compromise between the objectives of all parties. BMSs in force are considered too mild by insurers, because they involve a high degree of crosssubsidization between the various risk classes. Regulators often consider existing BMSs to be excessive, because they induce hit-and-run behaviors and uninsured driving. The preceding results illustrate the profound differences in the concerns of regulators and actuaries. They help explain why the design of a BMS is such a delicate process.
REFERENCES
FERREIRA, J. 1974. The Long-Term Effect of Merit-Rating Plans on Individual Motorists, Operations Research 22:954 978. LEMAIRE, J., AND ZI, H. 1994. A Comparative Analysis of 30 Bonus-Malus Systems. ASTIN Bulletin 24:287309. LEMAIRE, J. 1995. Bonus-Malus Systems in Automobile Insurance. Boston: Kluwer. VENTER, G. 1991. A Comparative Analysis of Most European and Japanese Bonus-Malus Systems: Extension, Journal of Risk and Insurance 58:542547.
Additional discussions on this paper can be submitted until July 1, 1998. The author reserves the right to reply to any discussion. See the Submission Guidelines for Authors for detailed instructions on the submission of discussions.
DISCUSSIONS
OF
PAPERS ALREADY PUBLISHED
143
to reect diverse utility preferences. The right-tail index is a special case in which the weights of T * are based on integrals of [S (x)]1/2. One nal link between the two methodologies can be drawn from the parameter risk section of Mr. Wangs paper. The normal approximation presented above for multiple policies leads to a risk premium formula E[X] rzE[X], which can be written E[X] rz E[X] E[X] z E[X]. n1/2
Approximating (/) by d [X] D [X]/E[X] yields E[X] z D[X] E[X] E[X] D[X]. n1/2 E[X]
This indicates that the right-tail index may be especially useful in situations in which normal approximations are reasonable, as is often the case with a large number of nite-variance policies. Environments in which the number of policies is small or tail probabilities are excessive might be more appropriately addressed through explicit utility functions or the exible forms r *.
BONUS-MALUS SYSTEMS: THE EUROPEAN AND ASIAN APPROACH MERIT RATING, JEAN LEMAIRE, JANUARY 1998
TO
LIVIANA PICECH,* PATRIZIA GIGANTE, AND LUCIANO SIGALOTTI

This discussion on the paper by Jean Lemaire aims to point out some comments on merit-rating systems in automobile liability insurance, inspired by recent Italian experience. After the deregulation originating from the application of the Third European Economic Community Directive, Italian insurance companies began to offer a wide supply of different tariff structures
based on increasingly personalized rating systems. In particular, the insurers are now using more a priori classication variables and various bonus-malus systems (BMSs) are emerging. Moreover, some meritrating systems recently appearing in the market combine bonus-malus and deductibles: the policyholders are spread into classes with different premium coefcients and different deductible levels. In this way a further element of personalization is introduced. Some of the systems recently proposed by the Italian companies have been evaluated and compared by Gigante, Picech, and Sigalotti (1996) by analyzing: the distribution of the policyholders among the bonusmalus classes, the average premium level, and the equilibrium premium and their time development, which allows us to study the dynamics of the system along the periods preceding the steady-state. We want to give here a brief illustration of the comparison between a specic BMS with deductible (which we call system A) and the BMS in use in Italy since before the deregulation (system B). To model the number of losses of an individual policyholder, we consider both a Poisson-gamma process and a Markov chain process. For the evaluation of the abovementioned quantities, we use recursive procedures developed in Sigalotti (1994) and Gigante (1997). This approach also allows us to deal with open insurance portfolios; in this way we can study the impact of new policyholders on the nancial stability of the system. System A. The BMS with deductible is organized as follows: Rules: starting class 15 j
max(h 1, 1) r0 min(21, h 3r 1) r 1, 2, 3
*Liviana Picech is a Researcher in the Dipartimento di Matematica Applicata alle Scienze Economiche Statistiche ed Attuariali: Bruno de Finetti, Universita ` di Trieste, Piazzale Europa 10-34127 Trieste, Italy, e-mail, livianap@econ.univ.trieste.it. Patrizia Gigante is a Researcher in the Dipartimento di Matematica Applicata alle Scienze Economiche Statistiche ed Attuariali: Bruno de Finetti, Universita ` di Trieste, Piazzale Europa 10-34127 Trieste, Italy, e-mail, patriziag@econ.univ.trieste.it. Luciano Sigalotti is a Professor in the Dipartimento di Finanza dellImpresa e dei Mercati Finanziari, Universita ` di Udine, Via Tomadini 30 / A-033100 Udine, Italy, e-mail, luciano.sigalotti@dmf. uniud.it.
where h indicates the initial class, j the nal class, and r the number of claims incurred in one year exceeding the deductible Table 1 reports the deductible levels (expressed as a percentage of each class premium) and the premium scale. System B. The BMS in use in Italy since before the deregulation is organized as follows: Rules: starting class 14 j
max(h 1, 1) n0 min(18, h 3n 1) n 1, 2, 3, 4
where h indicates the initial class, j the nal class, and n the number of claims in one year Table 2 reports the premium scale. Tables 3, 4, 5, and 6 show the distributions of the policyholders among the classes of the two systems for
144
Table 1 Deductible Levels and Premium Scale for BMS System A

Deductible Level (% of Class Premium) 0% 0 25 25 25 25 25 25 25 50 50 50 100 100 100 100 100 100 100 100 100
Table 2 Premium Scale for BMS System B

Class Premium Scale 0.50 0.53 0.56 0.59 0.62 0.66 0.70 0.74 0.78 0.82 0.88 0.94 1.00 1.15 1.30 1.50 1.75 2.00
Class 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Premium Scale 0.60 0.64 0.64 0.67 0.70 0.74 0.77 0.80 0.84 0.84 0.88 0.92 0.92 0.96 1.00 1.12 1.20 1.44 1.64 2.00 2.40
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
specic years, for both closed and open portfolios and under the above-mentioned hypotheses for the claim number process. In the case of open portfolios, the new policyholders are placed in the starting class of the system.
First, note that the effect of the new policyholders is the same in both BMSs, resulting in a reduced concentration in the rst class. With a rate of 3% of new policies, this reduced concentration is very remarkable. If we compare the two BMSs, we see that System A produces a higher concentration of policies in the rst class. Since the transition rules are essentially the same, this effect is due to the low-value losses, which,
Table 3 Distributions of the Policyholders in BMS System A: Markov Claim Number Process
No Entries Class 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 Distribution in Year 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 Distribution in Year 10 0 0 0 0 0 0.384295 0.000000 0.000000 0.337478 0.000000 0.000000 0.177487 0.000327 0.001932 0.068298 0.001449 0.002865 0.019581 0.001454 0.002073 0.002760 Distribution in Year 30 0.628123 0.066612 0.077772 0.085260 0.024687 0.016376 0.046988 0.007931 0.005148 0.022395 0.002609 0.001827 0.008776 0.000842 0.000678 0.002670 0.000275 0.000250 0.000591 0.000092 0.000098 3% New Entries Distribution in Year 10 0 0 0 0 0 0.294349 0.009895 0.011422 0.271709 0.022800 0.025023 0.163448 0.033495 0.036604 0.089673 0.008742 0.007470 0.017184 0.002832 0.002599 0.002754 Distribution in Year 30 0.363934 0.052779 0.063 0.061301 0.038094 0.035611 0.049681 0.034901 0.035119 0.043965 0.036956 0.037953 0.042129 0.039701 0.040747 0.010248 0.006494 0.003055 0.002315 0.001256 0.000761
DISCUSSIONS
OF
145
Table 4 Distributions of the Policyholders in BMS System B: Markov Claim Number Process
No Entries Class 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Distribution In Year 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 Distribution in Year 10 0 0 0 0 0.344178 0 0 0.321631 0 0.004318 0.186502 0.005539 0.010944 0.078784 0.007947 0.010909 0.022028 0.007221 Distribution in Year 30 0.597287 0.066392 0.131202 0.030461 0.028662 0.055188 0.013575 0.011846 0.02687 0.006254 0.005509 0.012128 0.003044 0.002745 0.004621 0.001509 0.001355 0.001350 3% New Entries Distribution in Year 10 0 0 0 0 0.263784 0.008897 0.010303 0.258434 0.021226 0.026843 0.169089 0.037036 0.043929 0.099311 0.015738 0.015697 0.020619 0.009094 Distribution in Year 30 0.357927 0.053229 0.086488 0.039025 0.040333 0.052727 0.036525 0.037517 0.045469 0.038728 0.040287 0.045030 0.043312 0.045657 0.014992 0.010768 0.006399 0.005588
Table 5 Distributions of the Policyholders in BMS System A: Poisson-Gamma Claim Number Process
No Entries Class 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 Distribution in Year 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 Distribution in Year 10 0 0 0 0 0 0.458321 0 0 0.274161 0 0 0.145213 0.000110 0.000733 0.068634 0.001116 0.002996 0.029809 0.003521 0.006239 0.009148 Distribution in Year 30 0.649221 0.048472 0.052755 0.061513 0.011574 0.007101 0.046630 0.003331 0.002942 0.036289 0.002263 0.003071 0.026160 0.002725 0.003765 0.016212 0.003566 0.004584 0.008997 0.004183 0.004647 3% New Entries Distribution in Year 10 0 0 0 0 0 0.351189 0.011319 0.012582 0.224005 0.022018 0.024104 0.137867 0.032017 0.034859 0.089548 0.008519 0.007716 0.025060 0.004799 0.006219 0.008180 Distribution in Year 30 0.391578 0.042791 0.048389 0.048934 0.028572 0.027869 0.046884 0.029643 0.031290 0.047285 0.034978 0.037296 0.048141 0.040944 0.043159 0.016828 0.009883 0.006967 0.007862 0.005437 0.005271
because of the deductibles, do not result in claims. Both claim number process hypotheses yield this result. If we compare the two stochastic models for the claim number process, we observe a lower rate of policies in the best class when a Markov chain is used. This can be explained by the different effect of past experience in the two models. By assuming a Markov
model, we consider as relevant only very recent experience. Conversely, by using a Poisson-gamma model all past experience is relevant and in a particular way: old and recent claims affect the distribution of future claims to the same extent. Table 7 exhibits the index numbers of the equilibrium premiums relative to the base period 10, under the Poisson-gamma hypothesis for the claim number
146
Table 6 Distributions of the Policyholders in BMS System B: Poisson-Gamma Claim Number Process
No Entries Class 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Distribution In Year 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 Distribution in Year 10 0 0 0 0 0.426592 0 0 0.262172 0 0.001919 0.145168 0.002558 0.007240 0.074379 0.008670 0.015566 0.035079 0.020657 Distribution in Year 30 0.629217 0.046897 0.090593 0.015937 0.013865 0.042671 0.007109 0.007261 0.030292 0.005942 0.007278 0.022351 0.007494 0.009469 0.017687 0.011680 0.015009 0.019249 3% New Entries Distribution in Year 10 0 0 0 0 0.326948 0.010522 0.011676 0.213830 0.020665 0.024181 0.136573 0.033065 0.039873 0.095255 0.016040 0.019374 0.031228 0.020768 Distribution in Year 30 0.391605 0.042745 0.066123 0.029389 0.030038 0.044238 0.030064 0.031903 0.043751 0.035587 0.038476 0.047201 0.044438 0.048793 0.021618 0.017827 0.016517 0.019690
Table 7 Index Numbers of the Equilibrium Premiums

BMS System A No Entries 3% New Entries 100 103.22 105.43 108.48 110.67 113.84 115.78 117.30 119.19 120.61 121.68 122.83 123.91 124.69 125.41 126.27 126.84 127.30 128.00 128.42 128.72 100 102.31 103.82 105.84 107.21 109.11 110.22 111.07 112.07 112.79 113.34 113.89 114.39 114.75 115.08 115.43 115.69 115.89 116.15 116.34 116.47 BMS System B No entries 3% New Entries 100 103.59 106.44 109.82 113.60 115.29 117.43 119.90 120.96 122.39 124.08 124.79 125.79 126.99 127.49 128.22 129.11 129.48 130.02 130.70 130.97 100 102.53 104.45 106.62 108.91 109.90 111.10 112.42 112.98 113.70 114.51 114.86 115.32 115.85 116.08 116.40 116.76 116.93 117.16 117.42 117.54
Class 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
process. The equilibrium premium, that is, the reference premium that makes the expected global earned premium equal to the expected global payment, is a sound tool for analyzing BMSs because it takes into account the distribution of the policyholders among the classes of the system, the expected premiums, and the expected losses. Moreover, its time development points out the increments of the reference premiums
needed to maintain the nancial stability of the system. Both systems require raising the reference premiums, and the increments are very close. The effect of the new policies is once again very remarkablein this case the needed increments are quite reduced. Analogous results are obtained under the Markov process hypothesis.
DISCUSSIONS
OF
147
These evaluations allow us to sketch the importance of studying the changes in the portfolios composition (entries and/or exits) so far as the nancial equilibrium of the system is concerned.
REFERENCES
GIGANTE, P. 1997. Un modello per una tariffa RCA Bonus-Malus con franchigia, Giornale dellIstituto Italiano degli Attuari LX, no. 12:2949. GIGANTE, P., PICECH, L., AND SIGALOTTI, L. 1996. Valutazioni attuariali per sistemi Bonus-Malus, Atti del V Congresso Nazionale di Scienza delle Assicurazioni, Torino 1-3/12/ 1996, in press. SIGALOTTI, L. 1994. Equilibrium Premiums in a Bonus-Malus System, Technical Report, Univ. Trieste, presented at Speakers Corner, XXV Astin Colloquium, Cannes.
when, for each given bonus-malus class, the BM premium is equal to the expected claim payments. The second index is designed to investigate the level of personalization implied by the system and helps to value the risk heterogeneity present in some bonusvalue classes. In this paper we compare the three BMSs considered by P. Lemaire using these indexes. Because the evaluations are performed from the insurance company point of view, the Poisson-gamma hypothesis for the claims number process is assumed. In order to calculate the quantities that appear in the denitions of the two indexes, we implement the recursive procedure developed in Sigalotti (1994).
The Model
With reference to one policyholder in year t, let: Nt the number of claims the claim amounts the total loss the bonus-malus class.
BONUS-MALUS SYSTEMS: THE EUROPEAN AND ASIAN APPROACH MERIT-RATING, JEAN LEMAIRE, JANUARY 1998 LIVIANA PICECH* SIGALOTTI
AND
TO
Z 1t, Z 2t, . . . Xt Yt
h0
Z
Nt
ht
LUCIANO
In his paper Jean Lemaire has provided a wide range of actuarial tools for evaluating and comparing different bonus-malus systems (BMSs). In his discussion, Pierre Lemaire has concentrated on other important aspects that have to be carefully considered when a new BMS is to be adopted: a measure of the rate of solidarity present in the system (the predictive accuracy) and same regulatory tools (namely, the nancial responsibility and the public acceptance). This discussion of Dr. Lemaires paper also contributes to the development of some measures of the rate of efciency of a BMS. More precisely, we are considering two indexes [see Picech and Sigalotti (1995)], which enlighten some complementary aspects connected to the efciency of a BMS. The rst index is designed to value the fairness in the BMS and therefore has its maximum value
We assume that The number of claims, Nt, is Poisson-gamma distributed and is the random parameter; thus Nt has a negative binomial distribution. The claims amounts Z 1t, Z 2t, . . . are independent and identically distributed [let E (Z ) E (Z 1t ) E (Z 2t) . . .] and are independent from the number of accidents Nt and from the bonus malus class Yt . We calculate recursively The distribution of Yt The distributions of Nt Yt for any given value of Yt The expectations E (Nt Yt ) for any given value of Yt The equilibrium premium P e t , dened as the solution of the equation in Pt E (Xt) Pt
h1
Pr (Y h)
H h t
*Liviana Picech is a Researcher in the Dipartimento di Matematica Applicata alle Scienze Economiche Statistiche ed Attuariali: Bruno de Finetti, Universita ` di Trieste, Piazzale Europa 1-034127 Trieste, Italy, e-mail, livianap@econ.univ.trieste.it. Luciano Sigalotti is a Professor in the Dipartimento di Finanza dellImpresa e dei Mercati Finanziari, Universita ` di Udine, Via Tomadini 30 / A-033100 Udine, Italy, e-mail, luciano.sigalotti@dmf. uniud.it.
where ( 1, . . . , H ) is the premium scale of the BMS. The quantity
h1
Pr (Y h)
H h t
is the so-called average premium level and generally decreases until the system reaches its stationary distribution [see J. Lemaire (1998)]. For this reason, it
148
is important to note that, as Dufresne also points out (1995), it is necessary to rescale the premiums in each period, so that the expected average premium paid by the insured is equal to the expected claim payments. In year t, the rescaled premium for the h-th bonus-malus class is then P e t h and the equilibrium premium P e plays the role of the basic premium. t With reference to one risk in year n 1, the bonusmalus class is determined by the starting class (in year 1) and by the accidents incurred in the n years. Let H (N1 n1, N2 n2, . . . , Nn nn); then y (H ) is the bonus-malus class of the risk in year n 1 and Y(H ) is the premium level. In order to maintain equilibrium in the system, the updated pure premium is P e n1 Y(H ).
e B E[E (Xt Yt ) P t Yt]2,
then the structure of the index can be represented as follows:

t) i( 1

A AB
1/2
Efciency Indexes for BMSs

The notion of efciency in a BMS refers to its aptitude to discriminate among good and bad risks. For this purpose the transition rules should be well designed to spread the risks over the bonus-malus classes, whereas the premium scale should allow the premiums of each bonus-malus class to reect its expected claims. To evaluate the discriminating power of a BMS, the relative stationary average level (see P. Lemaires discussion) can be calculated. On the other hand, to investigate how the premiums P e t h reect the expected claims E (Xt Yt h), we can consider for each t the following quantity, which can be seen as a measure of the efciency in that period:
e E[E (Xt Yt ) P t Yt ]2
The level of personalization implied by a BMS also affects its efciency. It could happen that, in spite of the spreading of policies among the bonus-malus classes and the assessment of premiums reecting their actual expected claims, policies belonging to the same class could have different attitudes toward risk. Since these risks pay the same premium, a certain level of solidarity is present in the system. A measure of the capability of the bonus-malus premiums to reect the risk premiums E(Xt ) can be dened as follows:
e E[E (Xt ) P t Yt]2
Pr(Y h) (E (Z) P
t h
e t
h)2 f (Yt h)d .
This concept relates to the predictive accuracy introduced by Venter (1991). We can dene the following t) index analogously to the index i ( : 1
(t) i2
2 E[E (Xt) P e t Yt] 2 e 2 E[E (Xt) P e t Yt] E[E (Xt) P t Yt ]
1/2
Picech and Sigalotti (1995) show that

e E[E (Xt ) P t Yt]2 B C
Pr (Y h)[E(X Y h) P
t t t h
e t
h] .
2
where C E 2(Z )
To compare BMSs with different transition rules and premium scales, it is helpful to adopt some indexes. According to the suggestions in Cacciafesta (1983), the following index has been introduced:
t) i( 1
Pr (Y h)
t h
[var(Nt Yt h) E (Nt Yt h)]. We see that the more remarkable the heterogeneity of the risks within the bonus-malus classes, the higher the C. t) Hence, the structure of the index i ( is 2
t) i( 2
e E[E (Xt) P t Yt]2 2 e 2 E[E (Xt) P e t Yt] E[E (Xt Yt ) P t Yt]
1/2
t) 2 The term I ( values 0 when E[E (Xt) Pe t Yt] 0, 1 that is, all policyholders pay the same premium. The t) 2 term i ( values 1 when E[E (Xt Yt) P e t Yt] 0, 1 which means that the policyholders pay a premium exactly equal to the expected claim payments in the bonus-malus class to which they belong. If we let e A E[E (Xt) P t Yt]2
A ABC
1/2
The quantities A, B, and C that appear in the denition of the two indexes can be easily calculated through the recursive procedure.
Application
t) t) The indexes i( and i ( have been calculated for the 1 2
and
DISCUSSIONS
OF
149
Thai, Danish, and Swiss BMSs. According to the assumptions in P. Lemaires discussion, we assume that: Nt has a negative binomial distribution with E(Nt) 0.1 and var (Nt) 0.107 The expected claim cost is E(Z ) 10,000$. By applying the cited recursive procedure, it is possible to calculate the indexes regardless of the initial distribution of the risks in the bonus-malus classes (that is, the probability distribution Pr (Y1 h), h 1, . . . , H ) and for any length of time. To compare the three BMSs, the following table shows the values i1 and i2 of the indexes obtained when the BMSs have reached their steady states:
lowest bonus-malus class of the Danish system (77.25% of the policies, as compared to 59% in the Swiss system) and the resulting solidarity among them.
Conclusions
In his discussion P. Lemaire points out how delicate a matter the revision of a BMS is, in particular when the companys demands have to be matched to supervisory requirements. In any case, it is important that the effects of any choice be carefully investigated, and for this purpose there are many actuarial tools to help enlighten many different aspects.
Indexes i1 i2
Thailand 76.72% 23.78%
Denmark 98.59% 56.39%
Switzerland 96.80% 77.97%
REFERENCES
CACCIAFESTA, F. 1983. Sulla misura del grado di personalizzazione nelle forme di assicurazione RCA Bonus-Malus, Giornale dellIstituto Italiano degli Attuari XLVI, No. 12: 12537. DUFRESNE, F. 1995. The Efciency of the Swiss Bonus-Malus System, Mitteilungen der Schweizerische Vereinigung der Versicherungsmathematiker, 1:2942. PICECH, L., AND SIGALOTTI, L. 1995. Equilibrium Premiums and some Efciency Measures for Bonus-Malus Systems, presented at Speakers Corner, XXVI ASTIN Colloquium, Leuven. SIGALOTTI, L. 1994. Equilibrium Premiums in a Bonus-Malus System, Technical Report, Univ. Trieste, presented at Speakers Corner, XXV ASTIN Colloquium, Cannes. VENTER, G. 1991. A Comparative of Analysis of Most European and Japanese Bonus-Malus Systems: Extension, Journal of Risk and Insurance 58:542547.
For the rst index, both the Danish and the Swiss BMSs get a very high score (the Danish system, in particular, seems to work even better than the Swiss), whereas the low value obtained by the Thai system shows that its bonus-malus premiums do not well reect the expected claims. The low value of the second index for the Thai system also shows that the personalization produced by that system is very poor. Conversely, from the personalization perspective, the Swiss system is more effective than the Danish one. At a glance, this result seems to contradict the rst index valueshowever, it is explained by a high concentration of risks in the

Naaj9801 2

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

Naaj9801 2

Uploaded by

Copyright:

Available Formats

BONUS-MALUS SYSTEMS: THE EUROPEAN AND ASIAN APPROACH TO MERIT-RATING

BONUS-MALUS SYSTEMS: THE EUROPEAN

Insurance Plan that is essentially a BMS was introduced in 1990.

The matrix M () [pi j()]

is the transition matrix of this Markov chain.

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 1

TABLE 1 BRAZILIAN SYSTEM (STARTING

TABLE 2 BELGIAN SYSTEM

BONUS-MALUS SYSTEMS: THE EUROPEAN

TABLE 3 BELGIAN BMS: MARKOVIAN PRESENTATION

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 1

3. TOOLS FOR THE DESIGN AND EVALUATION OF BONUS-MALUS SYSTEMS

3.1. The Relative Stationary Average Level

BONUS-MALUS SYSTEMS: THE EUROPEAN

FIGURE 1 MEAN PREMIUM LEVEL

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 1

3.2. The Coefcient of Variation of the Insureds Premium

BONUS-MALUS SYSTEMS: THE EUROPEAN

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 1

3.4 The Average Optimal Retention

BONUS-MALUS SYSTEMS: THE EUROPEAN

The expectation of the number of reported claims is i

The expected cost of a nonreported claim is equal to E i(X) 1 pi

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 1

3.5 The Rate of Convergence of BonusMalus Systems

BONUS-MALUS SYSTEMS: THE EUROPEAN

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 1

DISCUSSIONS KRUPA SUBRAMANIAN* BONUS-MALUS SYSTEMS IN A COMPETITIVE ENVIRONMENT

BONUS-MALUS SYSTEMS: THE EUROPEAN

TABLE 2 SELECTED VALUES

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 1

Time 0 1 2 3 4 5 6 7 8 9 10 Cumulated Prot

BONUS-MALUS SYSTEMS: THE EUROPEAN

FIGURE 1 MARKET SHARES

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 1

FIGURE 2 CLAIM FREQUENCIES

BONUS-MALUS SYSTEMS: THE EUROPEAN

FIGURE 3 MARKET SHARES

FIGURE 4 CLAIM FREQUENCIES

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 1

TIME 15 Premium Income (000s) $10,000 10,000 Annual Prot (000s)

Time 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Cumulated Prot

Next Premium Increase

System Initial B-revised A-revised

Elasticity 0.2736 0.3346 0.3948

Premium Coefcient of Variation 0.3627 0.4222 0.4889

BONUS-MALUS SYSTEMS: THE EUROPEAN

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 1

BONUS-MALUS SYSTEMS: THE EUROPEAN

PAPERS ALREADY PUBLISHED

LIVIANA PICECH,* PATRIZIA GIGANTE, AND LUCIANO SIGALOTTI

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 4

Table 1 Deductible Levels and Premium Scale for BMS System A

Table 2 Premium Scale for BMS System B

PAPERS ALREADY PUBLISHED

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 4

Table 7 Index Numbers of the Equilibrium Premiums

PAPERS ALREADY PUBLISHED

where ( 1, . . . , H ) is the premium scale of the BMS. The quantity

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 2, NUMBER 4

e B E[E (Xt Yt ) P t Yt]2,

then the structure of the index can be represented as follows:

Efciency Indexes for BMSs