Bonus-malus systems 6.1 INTRODUCTION

Transcription

1 6 Bonus-malus systems 6.1 INTRODUCTION This chapter deals with the theory behind bonus-malus methods for automobile insurance. This is an important branch of non-life insurance, in many countries even the largest in total premium income. A special feature of automobile insurance is that quite often and to everyone s satisfaction, a premium is charged which depends for a great deal on the claims filed on the policy in the past. In experience rating systems such as these, bonuses can be earned by not filing claims, and a malus is incurred when many claims have been filed. Experience rating systems are common practice in reinsurance, but in this case, it affects the consumer directly. Actually, by charging a randomly fluctuating premium, the ultimate goal of insurance, namely being in a completely secure financial position, is not reached. But it can be shown that in this type of insurance, the uncertainty is greatly reduced. This same phenomenon can also be observed in other types of insurance; think for instance of the part of the claims that is not reimbursed by the insurer because there is a deductible. That lucky policyholders pay for the damages caused by less lucky insureds is the essence of insurance (probabilistic solidarity). But in private insurance, solidarity should not lead to inherently good risks paying for bad ones. An insurer 127

2 128 BONUS-MALUS SYSTEMS trying to impose such subsidizing solidarity on his customers will see his good risks take their business elsewhere, leaving him with the bad risks. This may occur in the automobile insurance market when there are regionally operating insurers. Charging the same premiums nationwide will cause the regional risks, which for automobile insurance tend to be good risks because traffic is not so heavy there, to go to the regional insurer, who with mainly good risks in his portfolio can afford to charge lower premiums. There is a psychological reason why experience rating is broadly accepted with car insurance, and not, for instance, with health insurance. Bonuses are seen as rewards for careful driving, premium increases as an additional and well-deserved fine for the accident-prone. Many think that traffic offenses cannot be punished harshly and often enough. But someone who is ill is generally not to blame, and does not deserve to suffer in his pocket as well. Traditionally, car insurance covers third party liability, as well as the damage to one s own vehicle. The latter is more relevant for rather new cars, since for reasons of moral hazard, insurers do not reimburse more than the current value of the car. In Section 6.2, we describe the Dutch bonus-malus system, which we consider to be typical for such systems. Also, we briefly describe the reasons which have led to this system. Bonus-malus systems lend themselves for analysis by Markov chains, see Section 6.3. In this way, we will be able to determine the Loimaranta efficiency of such systems, which is the elasticity of the mean asymptotic premium with respect to the claim frequency. In Chapter 7, we present a bonus-malus system that is a special case of a so-called credibility method. In Chapter 8, we study among other things some venerable non-life actuarial methods for automobile premium rating in the light of generalized linear models. 6.2 AN EXAMPLE OF A BONUS-MALUS SYSTEM Every country has his own bonus-malus system, the wheel having been reinvented quite a few times. The Dutch system is the result of a large-scale investigation of the Dutch market by five of the largest companies in 1982, prompted by the fact that the market was chaotic and in danger of collapsing. Many Dutch insurers still utilize variants of the proposed system. First, a basic premium is determined using rating factors like weight, catalogue price or capacity of the car, type of use of the car (privately or for business), and of course the type of coverage (comprehensive, third party only, or a mixture).

3 AN EXAMPLE OF A BONUS-MALUS SYSTEM 129 This is the premium that drivers without a known claims history have to pay. The bonus and malus for good and bad claims experience are implemented through the use of a so-called bonus-malus scale. One ascends one step, getting a greater bonus, after a claim-free year, and descends several steps after having filed one or more claims. The bonus-malus scale, including the percentages of the basic premium to be paid and the transitions made after 0, 1, 2, and 3 or more claims, is depicted in Table 6.1. In principle, new insureds enter at the step with premium level 100%. Other countries might use different rating factors and a different bonus-malus scale. The group of actuaries that proposed the new rating system in the Netherlands investigated about policies of which 50 particulars were known, and which produced claims. Both claim frequency and average claim size were studied. The factors that were thought relevant about each policy were not all usable as rating factors. Driving capacity, swiftness of reflexes, aggressiveness behind the wheel and knowledge of the highway code are hard to measure, while mileage is prone to deliberate misspecification. For some of these relevant factors, proxy

4 130 BONUS-MALUS SYSTEMS measures can be found. One can get a good idea about mileage by looking at factors like weight and age of the car, as well as the type of fuel used, or type of usage (private or professional). Diesel engines, for instance, tend to be used only by drivers with a high mileage. Traffic density can be deduced from region of residence, driving speed from horse power and weight of the car. But it will remain impossible to assess the average future claim behavior completely using data known in advance, hence the need arises to use the actual claims history as a rating factor. Claims history is an ex post factor, which becomes fully known only just before the next policy year. Hence one speaks of ex post premium rating, where generally premiums are fixed ex ante. In the investigation, the following was found. Next to the car weight, cylinder capacity and horse power of the car provided little extra predicting power. It proved that car weight correlated quite well with the total claim size, which is the product of claim frequency and average claim size. Heavier cars tend to be used more often, and also tend to produce more damage when involved in accidents. Car weight is a convenient rating factor, since it can be found on official car papers. In many countries, original catalogue price is used as the main rating factor for third party damage. This method has its drawbacks, however, because it is not reasonable to assume that someone would cause a higher third-party claim total if he has a metallic finish on his car or a more expensive audio system. It proved that when used next to car weight, catalogue price also did not improve predictions about third party claims. Of course for damage to the own vehicle, it remains the dominant rating factor. Note that the premiums proposed were not just any function of car weight and catalogue price, but they were directly proportional to these numbers. The factor past claims experience, implemented as number of claim-free years, proved to be a good predictor for future claims, even when used in connection with other rating factors. After six claim-free years, the risk still diminishes, although slower. This is reflected in the percentages in the bonus-malus scale given in Table 6.1. Furthermore, it proved that drivers with a bad claims history are worse than beginning drivers, justifying the existence of a malus class with a premium percentage of more than 100%. An analysis of the influence of the region on the claims experience proved that in less densely populated regions, fewer claims occurred, although somewhat larger. It appeared that the effect of region did not vanish with an increasing number of claim-free years. Hence the region effect was incorporated by a fixed discount,

5 MARKOV ANALYSIS 131 in fact enabling the large companies to compete with the regionally operating insurers on an equal footing. The age of the policyholder is very important for his claim behavior. The claim frequency at age 18 is about four times the one drivers of age have. Part of this bad claim behavior can be traced back to lack of experience, because after some years, the effect slowly vanishes. That is why it was decided not to let the basic premium vary by age, but merely to let young drivers enter at a more unfavorable step in the bonus-malus scale. For commercial reasons, the profession of the policy holder as well as the make of the car were not incorporated in the rating system, even though these factors did have a noticeable influence. Note that for the transitions in the bonus-malus system, only the number of claims filed counts, not their size. Although it is clear that a bonus-malus system based on claim sizes is possible, such systems are hardly ever used with car insurance. 6.3 MARKOV ANALYSIS Bonus-malus systems are special cases of Markov processes. In such processes, one goes from one state to another in time. The Markov property says that the process is in a sense memory less: the probability of such transitions does not depend on how one arrived in a particular state. Using Markov analysis, one may determine which proportion of the drivers will eventually be on which step of the bonus-malus scale. Also, it gives a means to determine how effective the bonusmalus system is in determining adjusted premiums representing the driver s actual risk. To fix ideas, let us look at a simple example. In a particular bonus-malus system, a driver pays a high premium if he files claims in either of the two preceding years, otherwise he pays with To describe this system by a bonus-malus scale, notice first that there are two groups of drivers paying the high premium, the ones who claimed last year, and the ones that filed a claim only in the year before. So we have three states (steps): Claim in the previous policy year; paid 3. Claim-free in the two latest policy years; paid at the previous policy renewal; No claim in the previous policy year, claim in the year before; paid

6 132 BONUS-MALUS SYSTEMS First we determine the transition probabilities for a driver with probability of having one or more claims in a policy year. In the event of a claim, he falls to state 1, otherwise he goes one step up, if possible. We get the following matrix P of transition probabilities to go from state to state The matrix P is a stochastic matrix: every row represents a probability distribution over states to be entered, so all elements of it are non-negative. Also, all row sums are equal to 1, since from any state one has to go to some state Apparently we have Hence the matrix P has as a right-hand eigenvector for eigenvalue 1. Assume that initially at time the probability for each driver to be in state is given by the row-vector with and Often, the initial state is known to be and then will be equal to one. The probability to start in state and to enter state after one year is equal to so the total probability of being in state after one year, starting from an initial class with probability equals In matrix notation, the following vector gives the probability distribution of drivers over the states after one year: Drivers that produce a claim go to state 1. The probability of entering that state equals Non-claimers go to a higher state, if possible. The distribution over the states after two years is independent of since

7 MARKOV ANALYSIS 133 The state two years from now does not depend on the current state, but only on the claims filed in the coming two years. Proceeding like this, one sees that So we also have The vector is called the steady state distribution. Convergence will not always happen this quickly and thoroughly. Taking the square of a matrix, however, can be done very quickly, and doing it ten times starting from P already gives Each element of this matrix can be interpreted as the probability of going from initial state to state in 1024 years. For regular bonus-malus systems, this probability will not depend heavily on the initial state nor will it differ much from the probability of reaching from in an infinite number of years. Hence all rows of will be virtually equal to the steady state distribution. But there is also a more formal way to determine it. This goes as follows. First, notice that hence But this means that the steady state distribution is a left-hand eigenvector of P with eigenvalue 1. To determine we only have to find a non-trivial solution for the linear system of equations (6.5), which is equivalent to the homogeneous system and to divide it by the sum of its components to make a probability distribution. Note that all components of are necessarily non-negative, because of the fact that Remark (Initial distribution over the states) It is not necessary to take to be a probability distribution. It also makes sense to take for instance In this way, one considers a thousand drivers with initial state 1. Contrary to the vectors as well as do not represent the exact number of drivers in a particular state, but just the expected values of these numbers. The actual numbers are binomial random variables with as probability of success in a trial, the probability of being in that particular state at the given time. Efficiency The ultimate goal of a bonus-malus system is to make everyone pay a premium which is as near as possible the expected value of his yearly claims. If we want to investigate how efficient a bonus-malus system performs this task, we have to look at how the premium depends on the claim frequency To this end, assume that the random variation about this theoretical claim frequency can be described as a Poisson process, see Chapter 4. Hence, the number of claims in each year is

8 134 BONUS-MALUS SYSTEMS a Poisson variate, and the probability of a year with one or more claims equals The expected value of the asymptotic premium to be paid is called the steady state premium. It of course depends on and in our example where and the premiums are it equals This is the premium one pays on the average after the effects of in which state one initially started have vanished. In principle, this premium should be proportional to since the average of the total annual claims for a driver with claim frequency intensity parameter is equal to times the average size of a single claim, which in all our considerations we have taken to be independent of the claim frequency. Define the following function for a bonus-malus system: This is the so-called Loimaranta efficiency; the final equality is justified by the chain rule. It represents the elasticity of the steady state premium with respect to For small it can be shown that if increases by a factor increases by a factor which is approximately so we have Ideally, the efficiency should satisfy In view of the explicit expression (6.6) for for our particular three-state example the efficiency amounts to As the steady state premium doesn t depend on the initial state, the same holds for the efficiency, though both of course depend on the claim frequency Remark (Efficiency less than one means subsidizing bad drivers) The premium percentages in all classes are positive and finite, hence and hold. In many practical bonus-malus systems, we have over the whole range of This is for instance the case for formula (6.9) and all see Exercise Then we get

9 MARKOV ANALYSIS 135 As log decreases with so does from as to 0 as So there is a claim frequency such that the steady state premium for exactly equals the net premium. Drivers with pay less than they should, drivers with pay more. This means that there is a capital transfer from the good risks to the bad risks. The rules of the bonus-malus system punish the claimers insufficiently. See again Exercise Remark (Hunger for bonus) Suppose a driver with claim probability who is in state 3 in the above system, causes a damage of size in an accident. If he is not obliged to file this claim with his insurance company, when exactly is it profitable for him to do so? Assume that, as some policies allow, he only has to decide on December 31st whether to file this claim, so it is certain that he has no claims after this one concerning the same policy year. Since after two years the effect of this particular claim on his position on the bonus-malus scale will have vanished, we use a planning horizon of two years. His costs in the coming two years (premiums plus claim), depending on whether or not he files the claim and whether he is claim-free next year, are as follows: Of course he should only file the claim if it makes his expected loss lower, which is the case if From (6.11) we see that it is unwise to file very small claims, because of the loss of bonus in the near future. This phenomenon, which is not unimportant in practice, is called hunger for bonus. On the one hand, the insurer misses premiums that are his due, because the insured in fact conceals that he is a bad driver. But this is compensated by the fact that small claims also involve handling costs. Many articles have appeared in the literature, both on actuarial science and on stochastic operational research, about this phenomenon. The model used can be much refined, involving for instance a longer or infinite time-horizon, with discounting. Also the time in the year that a claim occurs is important.

10 136 BONUS-MALUS SYSTEMS Remark (Steady state premiums and Loimaranta efficiency) To determine the steady state premium as well as the Loimaranta efficiency for a certain bonus-malus system, one may proceed as follows. Let denote the number of states. For notational convenience, introduce the functions with to describe the transition rules, as follows: if by otherwise. claims in a year, one goes from state to The probability of a transition from state to state when the parameter equals is Next consider the initial distribution where is the probability of finding a contract initially, at time in state for Then the vector of probabilities to find a driver in class at time can be expressed in the state vector as follows: The sum of the is unity for each In the steady state we find, taking limits for with As noted before, the steady state vector is a left-hand eigenvector of the matrix P corresponding to the eigenvalue 1. In the steady state, we get for the asymptotic average premium (steady state premium) with claim frequency with the premium for state Note that depends on but not on the initial distribution over the states.

11 EXERCISES 137 Having an algorithm to compute as in (6.16), we can easily approximate the Loimaranta efficiency All it takes is to apply (6.8). But it is also possible to compute the efficiency exactly. Write then where These derivatives can be determined by taking derivatives in the system (6.15). One finds the following equations: where the derivatives of can be found as Using the fact that the efficiency can be computed for every by solving the resulting system of linear equations. In this way, one can compare various bonus-malus systems as regards efficiency, for instance by comparing the graphs of for the plausible values of ranging from 0.05 to 0.2, or by looking at some weighted average of values. 6.4 EXERCISES Section Determine the percentage of the basic premium to be paid by a Dutch driver, who originally entered the bonus-malus scale at level 100%, drove without claim for 7 years, then filed one claim during the eighth policy year, and has been driving claim-free for the three years since then. Would the total of the premiums he paid have been different if his one claim occurred in the second policy year? Section Prove (6.8). Determine from this that with P as in (6.1). What is the meaning of its elements? Can you see directly must hold?

12 138 BONUS-MALUS SYSTEMS Determine in the example with three steps in this section if in state 2, instead of the premium is Argue that the system can now be described by only two states, and determine P and Show that in (6.9) for every and with When is close to 1? Recalculate (6.11) for a claim at the end of the policy year when the interest is Calculate the Loimaranta efficiency (6.9) by method (6.17) (6.19). Determine the value of such that the transition probability matrix P has vector as its steady state vector, if P is given by If for the steady state premium we have if and for estimate the Loimaranta efficiency at For the following transition probability matrix: determine the relation between and that holds if the steady state vector equals