INDIVIDUAL RISK MODELS FOR A SHORT TERM 2.1 Introduction In Chapter 1 we examined how a decision maker can use insurance to reduce the adverse financial impact of some types of random events. That examination ‘was quite general. The decision maker could have been an individual seeking pro- tection against the loss of property, savings, or income. The decision maker could have been an organization seeking protection against those same types of losses. In fact, the organization could have been an insurance company seeking protection against the loss of funds due to excess claims either by an individual or by its portfolio of insureds. Such protection is called reinsurance and is introduced in this chapter. The theory in Chapter 1 requires a probabilistic model for the potential losses. Here we examine one of two models commonly used in insurance pricing, reserv- ing, and reinsurance applications. For an insuring organization, let the random loss of a segment of its risks be denoted by S. Then $ is the random variable for which we seek a probability distribution. Historically, there have been two sets of postulates for distributions of S, The individual risk model defines S=X +X +04 X, Q1.1) where X; is the loss on insured unit i and n is the number of risk units insured Usually the X/s are postulated to be independent random variables, because the mathematics is easier and no historical data on the dependence relationship are needed. The other model is the collective risk model described in Chapter 12. The individual risk model in this chapter does not recognize the time value of money. This is for simplicity and is why the title refers to short terms. Chapters 4~ 11 cover models for long terms. Rapier 2 tnahadual Rak Models Yor @ Short Term 7In this chapter we discuss only closed models; that is, the number of insured units 1 in (2.1.1) is known and fixed at the beginning of the period. If we postulate about migration in and out of the insurance system, we have an open model. 2.2 Models for Individual Claim Random Variables First, we review basic concepts with a life insurance product. In a one-year term life insurance the insurer agrees to pay an amount b if the insured dies within a year of policy issue and to pay nothing if the insured survives the year. The prob- ability of a claim during the year is denoted by q. The claim random variable, X, has a distribution that can be described by either its probability function, p.l, or its distribution function, d.f. The pf. is { q x=0 Se) = PX =) = Yq x=b (22.1) 0 elsewhere, and the df. is 0 x<0 Fy(x) = Pr(X Sx) = { -4q Os . (2.2.2) From the p.f. and the definition of moments, ELX] = bg, E[X*] = Fg, (2.2.3) and Var(X) = Bq(1 — q). (2.2.4) These formulas can also be obtained by writing X= Ib (2.2.5) where b is the constant amount payable in the event of death and / is the random variable that is 1 for the event of death and 0 otherwise. Thus, Pr(I = 0) = 1 — 4 and Pr(I = 1) = q, the mean and variance of I are q and q(1 — q), respectively, and the mean and variance of X are bq and b¥q(1 — q) as above. The random variable I with its (0, 1) range is widely applicable in actuarial mod- els. In probability textbooks it is called an indicator, Bernoulli random variable, or binomial random variable for a single trial. We refer to it as an indicator for the sake of brevity and because it indicates the occurrence, I = 1, or nonoccurrence, 1 = 0, of a given event. We now seek more general models in which the amount of claim is also a random variable and several claims can occur in a period. Health, automobile, and other property and liability coverages provide immediate examples. Extending (2.2.5), we postulate that X=. (2.26) 28 Section 2.2 Models for Individual Claim Random Variableswhere X is the claim random variable for the period, B gives the total claim amount incurred during the period, and 1 is the indicator for the event that at least one claim has occurred. As the indicator for this event, I reports the occurrence (I = 1) or nonoccurrence (I = 0) of claims in this period and not the number of claims in the period. Pr(I = 1) is still denoted by q. Let us look at several situations and determine the distributions of I and B for a model. First, consider a 1-year term life insurance paying an extra benefit in case of accidental death. To be specific, if death is accidental, the benefit amount is, 50,000. For other causes of death, the benefit amount is 25,000. Assume that for the age, health, and occupation of a specific individual, the probability of an accidental death within the year is 0.0005, while the probability of a nonaccidental death is 0.0020. More succinctly, Pr(I = 1 and B = 50,000) = 0.0005 and Pr(I = 1 and B = 25,000) = 0.0020. ‘Summing over the possible values of B, we have Pr(l = 1) = 0.0025, and then Pr(I = 0) = 1 — Pr(f = 1) = 0.9975. The conditional distribution of B, given I = 1, is Pr(B = 25,0001 = 1) = Pr(B = 50,000|I = 1) = Let us now consider an automobile insurance providing collision coverage (this indemnifies the owner for collision damage to his car) above a 250 deductible up toa maximum claim of 2,000. For illustrative purposes, assume that for a particular individual the probability of one claim in a period is 0.15 and the chance of more than one claim is 0: Pr(I = 0) = 0.85, Px(I = 1) = 0.15. This unrealistic assumption of no more than one claim per period is made to sim- plify the distribution of B. We remove that assumption in a later section after we discuss the distribution of the sum of a number of claims. Since B is the claim incurred by the insurer, rather than the amount of damage to the car, we can infer two characteristics of I and B. First, the event I = 0 includes those collisions in which the damage is less than the 250 deductible. The other inference is that B's distribution has a probability mass at the maximum claim size of 2,000. Assume Ghapter2_Indiadual Risk Models for a Short Term 2this probability mass is 0.1. Furthermore, assume that claim amounts between 0 and 2,000 can be modeled by a continuous distribution with a p.d.£. proportional to 1 — x/2,000 for 0 < x < 2,000, (In practice the continuous curve chosen to represent the distribution of claims is the result of a study of claims by size over a recent period.) Summarizing these assumptions about the conditional distribution of B, given I = 1, we have a mixed distribution with positive density from 0 to 2,000 and a mass at 2,000. This is illustrated in Figure 2.2.1. The d.f. of this con- ditional distribution is 0 xs0 oe _(,--yY mwa =at=1)~foo[1~ (1-55) ] o0, fi) = 26 x >0, fix) =3e* x50, Using (2.3.7) twice, we have Chapter 2 Individual Risk Models for a Short Term 7fe) = f fe — wftaddy = [ever 20 dy eae [eray =e -2* x>0, fo) = 7%) = fs = MAU) ay ices ene weer [Lemay em [lerty Ger — 30) — (66° - 6) =3e7—b6e* 430% x > 0. v Another method to determine the distribution of the sum of random variables is based on the uniqueness of the moment generating function (m.g.f.), which, for the random variable X, is defined by My(t) ~ Ele. If this expectation is finite for all t in an open interval about the origin, then M,(t) is the only m.g.f. of the distri- bution of X, and it is not the mg¥. of any other distribution. This uniqueness can be used as follows. For the sum $ = X, + X,+ +++ + Xy M(t) = Efe*] = Elect. eX, 23.8) Blettx0e--80] If X,, X,,..., X, are independent, then the expectation of the product in (2.3.8) is equal to Efe**] Efe™] «++ Efe] so that Mg(t) = My,(t) My,() +++ Mx,(8) (2.3.9) Recognition of the unique distribution corresponding to (2.3.9) would complete the determination of $’s distribution. If inversion by recognition is not possible, then inversion by numerical methods may be used. (See Section 2.6.) rn Consider the random variables of Example 2.3.3. Derive the p.d.f. of $ = X, + Xz + Xj by recognition of the mg-f. of S. Solution: By (2.3.9), M(t) = (Ale eS ‘)(=35). which we write, by the method of partial fractions, as 38 ‘Section 2.3 Sums of Independent Random VariablesThe solution for this is A = 3, B = -3, C = 1. But B/(B ~ #) is the moment generating function of an exponential distribution with parameter B, so the p.d.f for S is fale) = 3(e*) — 3Q2e*) + Be), Vv ‘The inverse Gaussian distribution was developed in the study of stochastic pro- cesses. Here it is used as the distribution of B, the claim amount. It will have a similar role in risk theory in Chapters 12-14. The p.d.f. and m.g.f. associated with the inverse Gaussian distribution are given by Sid) M(t) = exp [« ( Find the distribution of S = X, + X, + X, + +++ + X, where the random variables Xy Xp. ++, X, are independent and have identical inverse Gaussian distributions. Soluti Using (2.3.9), the m.g.f. of S is given by M(t) = IM,(Or" = ovn[ na ( - fi- x) The mg-f. Mg(t) can be recognized and shows that $ has an inverse Gaussian dis- tribution with parameters na and B. v 2.4 Approximations for the Distribution of the Sum The central limit theorem suggests a method to obtain numerical values for the distribution of the sum of independent random variables. The usual statement of the theorem is for a sequence of independent and identically distributed random variables, X,, X;,..., with E[X,] = » and Var(X,) = 0%. For each n, the distribution of Vn (X, — w)/o, where X, = (X, + X, + +++ + X,)/n, has mean 0 and vari- ance 1, The sequence of distributions (n = 1, 2, . . . ) is known to approach the standard normal distribution. When 1 is large the theorem is applied to approxi- mate the distribution of X, by a normal distribution with mean p. and variance o?/n, Equivalently, the distribution of the sum of the # random variables is ap- proximated by a normal distribution with mean np. and variance no?. The effect- iveness of these approximations depends not only on the number of variables but also on the departure of the distribution of the summands from normality. Many elementary statistics textbooks recommend that 1 be at least 30 for the Chapter 2 individual Risk Models for a Short Term 39approximations to be reasonable. One routine used to generate normally distrib- uted random variables for simulation is based on the average of only 12 inde- pendent random variables uniformly distributed over (0, 1) In many individual risk models the random variables in the sum are not iden- tically distributed. This is illustrated by examples in the next section. The central limit theorem extends to sequences of nonidentically distributed random variables, To illustrate some applications of the individual risk model, we use a normal approximation to the distribution of the sum of independent random variables to obtain numerical answers. If S=X FX te +X, then E[S] = = EX, and, further, under the assumption of independence, Var(S) = 2 Var(X,). For an application we need only + Evaluate the means and variances of the individual loss random variables + Sum them to obtain the mean and variance for the loss of the insuring orga- nization as a whole + Apply the normal approximation. Illustrations of this process follow. —— 2.5 Applications to Insurance In this section four examples illustrate the results of Section 2.2 and use of the normal approximation. k= 8 =— A life insurance company issues 1-year term life contracts for benefit amounts of 1 and 2 units to individuals with probabilities of death of 0.02 or 0.10. The following table gives the number of individuals n, in each of the four classes created by a benefit amount b, and a probability of claim 4,. . b, k 1, me 1 002 1 500 2 002 2-500 3 010 300 4010 2 10 40 Section 2.5 Applications to Insurance‘The company wants to collect, from this population of 1,800 individuals, an amount equal to the 95th percentile of the distribution of total claims. Moreover, it wants each individual's share of this amount to be proportional to that individual's ex pected claim, The share for individual j with mean E[X,] would be (1 + 9)E[X|]. The 95th percentile requirement suggests that @ > 0. This extra amount, 0E[X)], is the security loading and @ is the relative security loading. Calculate 8. Solution: The criterion for @ is Pr(S = (1 + @)E[S]) — 0.95 where § — X, + X, + +++ + Xs: This probability statement is equivalent to [8 = F{S] _ _eE(S] VWarS) VVar(S) Following the discussion of the central limit theorem in Section 2.4, we approximate the distribution of (S — E[S])/‘VVar(S) by the standard normal distribution and use its 95th percentile to obtain #5 IS] Waris) It remains to calculate the mean and variance of $ and to calculate 0 by this equation. For the four classes of insured individuals, we have the results given below. Variance k — big ms 1 002 0.0196 +500 2 0.0784 500 3 0.0900 300 4 7 0.3600 500 Then 100 ‘ E[S] = > FIX] = > mbes = 160 a a and Var(S) = 5) Var(x) = 3) mbigi(l ~ 44) = 256. Thus, the relative security loading is 0 = 1645 SYS) — 9 645 15 - 0 1645, : ES] 160 v Chapter 2 Individual Risk Models for a Short Term a—_k == =—“‘“‘C:;‘CS;C;~O~™~™~™~C~COC The policyholders of an automobile insurance company fall into two classes. Distibution of Claim Amount, By Parameters Number of Truncated Class in Class Exponential emg 1 500 ox0 125 22000 ____o0s_ 280 A truncated exponential distribution is defined by the df. 0 x<0 F@)=f1-e* Osx rbiqu = 8,000 (1)(0.02) + 8,000 (2)(0.02) = 480 and Var(S) = > mbiq(l — 4.) = 8,000 (1)(0.02)(0.98) + 8,000 (4)(0.02)(0.98) = 784. In addition to the retained claims, S, there is the cost of reinsurance premiums. The total coverage in the plan is 8,000 (1) + 3,500 (2) + 2,500 (3) + 1,500 (5) + 500 (10) = 35,000. The retained coverage for the plan is 8,000 (1) + 8,000 (2) = 24,000. Therefore, the total amount reinsured is 35,000 — 24,000 = 11,000 and the reinsur- ance cost is 11,000(0.025) = 275. Thus, at retention limit 2, the retained claims plus reinsurance cost is $ + 275. The decision criterion is based on the probability that this total cost will exceed 825, a ‘Section 2.5 Applications to InsurancePr(S + 275 > 825) = Pr(S > 550) os [s = BIS] , 550 HSI] WVar(S) VVar(S) pleted in Exercises 2.13 and 2.14, v In Section 1.5 stop-loss insurance, which is available as a reinsurance coverage, was discussed. The expected value of the claims paid under the stop-loss reinsur ance coverage can be approximated by using the normal distribution as the distri bution of total claims. Let total claims, X, have a normal distribution with mean p and variance 6? and let d be the deductible of the stop-loss insurance. Then, by (1.5.2A), the expected reinsurance claims equal BIL) = Ae fF do [- Changing the variable of integration to z = (x — 4)/o and defining B by d= » + Bo, we obtain the following general expression for the expected value of stop-loss claims under a normal distribution assumption: ° [oe ~ Bll - oor} 252) where (x) is the distribution function for the standard normal distribution. (25.1) EUL(X)] ‘Consider the portfolio of insurance contracts in Example 2.5.3. Calculate the ex- pected value of the claims provided by a stop-loss reinsurance coverage where a. There is no individual reinsurance and the deductible amount is 7,500,000 b. There is a retention amount of 20,000 on individual policies and the deductible amount on the business retained is 5,300,000. Solutio a. With no individual reinsurance and the use of 10,000 as the unit, E{S] = 0.02[8,000(1) + 3,500(2) + 2,500(3) + 1,500(5) + 500(10)] = 700 and Var(S) = (0.02)(0.98){8,000(1) + 3,500(4) + 2,500(9) + 1,500(25) + 500(100)] 587.2 Tapter 2 Individual Risk Models for a Short Term’(5) = 50.86. Then, with = = w) _ (750 — 700 © 50.86 the application of (2.5.2) gives us P = 50.86(0.24608 — (0.983)(0.16280)] = 4.377. This is equivalent to 43,770 in the example as posed. b. In Example 2.5.3 we determined the mean and the variance of the aggregate claims, after imposing a 20,000 retention limit per individual, to be 480 and 784, respectively, in units of 10,000. Thus o(S) = 28. B 0.983 Then, with d—p _ 530 ~ 480 _ B= 2B 1.786 the application of (2.5.2) gives us P = 28{0.08100 ~ (1.786)(0.03707)] = 0.414. This is equivalent to 4,140 in the example as posed. v 2.6 Notes and References The basis of the material in Sections 2.2, 23, and 2.4 can be found in a number of post-calculus probability and statistics texts. Mood et al. (1974) prove the theo- rems given in (2.2.10) and (2.2.11). They also provide an extensive discussion of properties of the moment generating function. For a discussion of the advanced mathematical methods for deriving the distribution function that corresponds to a given moment generating function, see Bellman et al. (1966). Methods are also available to obtain the p.f. of a discrete distribution from its probability generating function; see Kornya (1983). DeGroot (1986) provides a discussion of several conditions under which the cen- tral limit theorem holds. Kendall and Stuart (1977) give material on normal power expansions that may be viewed as modifications of the normal approximation to improve numerical results. Bowers (1967) also describes the use of normal power expansions and gives an application to approximate the distribution of present values for an annuity portfolio. ra Section 2.6 Notes and ReferencesExercises Section 2.2 24. 2.2. 23. 24. 25. 26. Use (2.2.3) and (2.2.4) to obtain the mean and variance of the claim random variable X where q = 0.05 and the claim amount is fixed at 10. Obtain the mean and variance of the claim random variable X where q = 0.05 and the claim amount random variable B is uniformly distributed between 0 and 20. Let X be the number of heads observed in five tosses of a true coin. Then, X true dice are thrown. Let Y be the sum of the numbers showing on the dice. Determine the mean and variance of Y. (Hint: Apply (2.2.10) and (2.2.11),] Let X be the number showing when one true die is thrown. Let Y be the number of heads obtained when X true coins are then tossed. Calculate E[Y] and Var(¥). Let X be the number oblained when one uue die is tossed. Let Y be the sunt of the numbers obtained when X true dice are then thrown. Calculate E[Y] and Var(Y). The probability of a fire in a certain structure in a given time period is 0.02. Ifa fire occurs, the damage to the structure is uniformly distributed over the interval (0, a) where @ is its total value. Calculate the mean and variance of fire damage to the structure within the time period. Section 2.3 27. 28. Independent random variables X; for four lives have the discrete probability functions given below. Prix, PAX, = 2) PAX, Pe(X, =») 06 07 06 9 00 02 00 00 03 on 00 00 00 00 o4 00 01 0.0 00. ol Use a convolution process on the non-negative integer values of x to obtain F(x) for x = 0, 1,2,..., 13 where $ = X, + X, + X + Xy Let X, for i = 1, 2, 3 be independent and identically distributed with the d.f. 0 x<0 Fo)=4x 0Sx<1 1 x21 Chapter 2 Individual Risk Models for a Short Term 7Let § = X, + X, + Xy a, Show that F.(x) is given by 0 x<0 : fa O 4). Compare this with the exact answer. 2.11. a. Use the central limit theorem to calculate b, ¢, and d, for given a, in the statement ml where the X/s are independent and identically distributed with mean and variance o? and (2) is the d.f. of the standard normal distribution. b. Evaluate the probabilities in Exercise 2.8(c) by use of the normal approximation developed in part (a). coe avi) + bea) 2.12, A random variable U has mg.f M,(t)= (1-2) ot } a, Use the mg.f. to calculate the mean and variance of U. b. Use a normal approximation to calculate points yogs and yoo, Such that Pr(U > y,) = € 3... >seeNote the random variable U has a gamma distribution with parameters a = and B = 1/2, Gamma distributions with a = n/2 and B = 1/2 are chi-square distributions with 1 degrees of freedom. Thus U has a chi-square distribution with 18 degrees of freedom. From tables of dif’s of chi-square distributions, we obtain yoqs = 28.869 and yao, = 34.805. Section 2.5 2.13. 214. 215. 2.16. Calculate the probability that the total cost in Example 2.5.3 will exceed 8,250,000 if the retention limit is a. 30,000 b. 50,000. Calculate the retention limit that minimizes the probability of the total cost in Example 2.5.3 exceeding 8,250,000. Assume that the limit is between 30,000 and 50,000. A fire insurance company covers 160 structures against fire damage up to an amount stated in the contract. The numbers of contracts at the different con- tract amounts are given below. Contract Amount ‘Number of Contracts 10000 80 20000 35 30000 2B 50 000 6 100 000, 5 Assume that for each of the structures, the probability of one claim within a year is 0.04, and the probability of more than one claim is 0. Assume that fires in the struchires are mutually independent events. Furthermore, assume that the conditional distribution of the claim size, given that a claim has oc- curred, is uniformly distributed over the interval from 0 to the contract amount. Let N be the number of claims and let S be the amount of claims in a 1-year period. a. Calculate the mean and variance of N. b. Calculate the mean and variance of S. c. What relative security loading, 8, should be used so the company can collect an amount equal to the 99th percentile of the distribution of total claims? (Use a normal approximation.) Consider a portfolio of 32 policies. For each policy, the probability q of a claim is 1/6 and B, the benefit amount given that there is a claim, has p.d.f. ad-y) O 4). Chapter 2_ Individual Risk Models for @ Short Term 8