cT6-08 ical Bayes Credibility theory Pege Chapter 6 Emperical Bayes Credibility theory m 2 ‘Swlabus objectives (©) Explain the fundamental concepts of Bayesian statistics and use these concepts 10 calculate Bayesian estimators. | 8 Explain the empirical Bayes approach to credibility theory, in particular {ts similarities with and its differences from the Bayesian approach. 9 State the assumptions underlying the two models in 8. 10. Calculate credibility premiums for the wo models in 8. Q Introduction In this chapter, we will study Empirical Bayes Credibility Theory (EBCT) Models 1 and 2. The algebra for the EBCT models that we will study in this unit is fkirly heavy going Try not to get too bogged down in the mathematics and don't lear the formulae. (Most ‘of them are in the Tables.) Concentrate on appreciating the differences between the models and being able to follow the numerical examples. ‘The Actuarial Education Company @IFE: 2010 Examinations14 Page 2 76-08: Emperical Bayes Credibility theory Empirical Bayes Credibility Theory: Model 1 Introduction of business written under each risk in euch year. Risk parameter Both approaches use an auxiliary risk parameter @. With the Bayesian approach, the chantity we wish to estimate is @, whereas, for the EBCT models, it is m(O), ie a function of @. Unlike the Bayesian approsch, the EBCT models don’t assume any Specific statistical distribution for 0. Conditional claim distribution Just using the assumption that the mean m(B) and variance s2(@) of X,|@ can be expressed as functions of 8. The values of ‘m(@) and s*(@) are then estimated by the models. Credibility formuta With both approaches, the resulting formula for estimating the claim frequency or risk Premium for a given risk can be expressed using a credibility formula, ie a linear combination (~ weighied average) of the average derived ftom ast claims and an overall average. @IFE: 2010 Examinations ‘The Actuarial Education Company42 CTE-06: Emperical Bayes Credibility theory Page? Empirical Bayes Credibility Theory is the name given to a particular approach to the problems in Section 1 of the previous chapter. This approach has lod to the development of a vast number of different models of varying degrees of complexity. In this chapter, two of these models will be studied. in this section the simplest possible model will be studied. Although this model, which will be felerred to as Model 4, is not very useful in practice, It does provide a good intteduction to the principles underlying Empirical Bayes Cradibilty Theory (EBCT). in particular, it shows the similarities and the differences between th, Gmpirical Bayes and the pure Bayesian approaches to credibility theory. In Section 2 an extension of Model 1 will be studied that is far more useful in practice. Model 1: specification In this section the assumptions for Model 1 of EBCT will be set out, EBCT Model 1 can be regarded as a generalisation of the normal/normal model. ‘This Point will be considered in more detail later in this section. The problem of interest is the estimation of the pure premium, or possibly the claim frequency, for a risk. Let X;,X2,... denote the aggregate claims, or the number of claims, in successive periods for this risk. A more precise statement of the problem is that, having observed the values of X4,X2,...Xp, the expected value of Xne needs to be estimated. From now on X4Xz....Xq will be denoted by X. The following assumptions will be made for EBCT Model 1. Assumptions for EBCT Model 1 | The distribution of each X; depends on a parameter, denoted 8, whose value Is. fixed (and the same for all the X; s) but is unknown, (14) Given 0, the X;'s are independent and identically distributed. (12) The parameter @ Is known as the risk parameter. It could, as in Section 2 of the Previous chapter, be a real number or it could be a more general quantity such as a set of real numbers, The risk parameter in the Poisson/gamma model is (the Poisson parameter) 2. The risk Parameter for the normal/normal model is @ The Actuarial Education Company © IFE: 2010 ExaminationsPage 4 CT6-06: Emperical Bayes Grecibilty theory A consequence of these two assumptions is th the random variables { Xj} are identically distributed, (1.3) An important point to note is that: the X;'s are not (necessarily) unconditionally independent. (14) chape,Tetmal model of Bayesian credibility in Section 2.8 of the Previous chapter, See (2.17), (2.18), (2.19), (2.23) and (2.24) In the provisus chapt Note that (1.1) corresponds to (2.17), (1.2) eomesponds to (2.18) and (2.19), (1.3) corresponds to (2.23), and (1.4) corresponds to (2.24), Next some notation is introduced. Define m() and s*(@) as follows: m(8)=E (x; |@) 8? (6)=Vvar(X, 10) Two things should be noticed about m(@) and s?(@). The first is that since, given @, the X,’s are identically distributed, neither m(8) nor s?(8) depends on J, a8 their notation suggests. The second is that since 8 Is regarded as a random variable, both m(@) and s?(@) are random variables, Note also that they are functions of the unknown parameter 0. Ir the value of @ and the distribution of X, given @ were known, the obvious getimate of the expected aggregate claims, or the expected number of claims, in any future year would ba m(@), Since it Is assumed that @ ls not known, the Problem is to estimate: m(8) given X (1.5) @IFE: 2010 Examinations ‘The Actuarial Education CompanyCT6-06: Emperical Bayes Credibility theory Page 5 Example | 2 species insurer concentates on providing titd party moter insurance to drivers of | ceriain type of car who are based in London. Explain what each of the variables and functions in EBCT Model 1 would represent if this model were used to describe the numbers of claims made by different drivers in different years. The only risk factor considered relevant is the safety standard adopted by each individwal driver Solution Here, the risk parameter @ Tepresents the “safety coefficient” of a particular driver. It is assumed that each driver has a constant inherent level of safety, which could (in theory) be measured by some means. The value of @ fora particular driver influences the likely number of claims that the driver will make. For example, a value of @=1 ‘may | correspond to a driver who is “100% safe” (ie never makes any claims), while a value of | =0 may correspond to a driver who is “0% safe” (ie makes a claim ‘every time the car is used). The distribution of the values of @ , which will vary from one driver to the next, is assumed to be determined by a definite (but unknown) probability distribution, ~; is @ random variable, representing the number of claims made by a particular driver in year j. X;|@ represents the number of claims made by a particular driver whose safety coefficient is @ in year J. m(@) = E(X, |@) is the (theoretical) average number of claims made by a driver whose | safety coefficient is @. In this example, m(@) would be a decreasing fimction of 8, since a higher safety coefficient would reduce the average number of claims, $°(8) = vat(X,,|@) is the (theoretical) variance of the number of claims made by a driver whose safety coefficient is @ in different years. s?(8) will take lower values for drivers with @ high safety coetfcien:, since they are likely to make no claims, or possibly one claim, each year, whereas the numbers of claims made each year by drivers with low safety coefficients may range from none to five or more, The Actuarial Education Company @ IFE: 2010 ExaminationsPage 6 16-08: Emperical Bayes Credibility theory who are not golfers, a player's handicap is a measure of hig orher golfing ability. A golfer with a low handicap is a better player than a golfer with a high handicap. A | [ 8204 golfer will achieve a lower score on a round of golf than a bed golfer.) The sir imilarities between EBCT Model 1 and the normal/normal model can be summarised as follows, @ (i (ily EBCT The role of 8 is the same for both modele: it characteris the underlying distributions of the processes being modelled, eg the aggregate claim istribution for each year of business. Soe (2.17) frome previous chapter and (1.1). Assumptions conceming the unconditional distribution of the X;'s are (2g Same: they are identically distributed in each case. See (2.23) and (2:24) from the previous chapter, and (1.3) and (1.4). Assumptions conceming the (conditional) distribution of the X;'s given @ are the same: they are conditionslly independent in (2.19) and (2.22) from the pravious chapter, and (1.2). Mode! 1 can be regarded as a generalisation of the hormalinormal mods The particular points where it differs from, fe generalises, the normal/normal model f) a) £(X; 16) Is some function of @, m(8), for EBCT Model 1 but Is simply 6 for the normai/normal model, Henck var [m(@)| {E8CT} corresponds to 3 (= var(@)) {normal/normaly. var (218) Is @ funetion of @, 5°(@), for EBCT Model 4 but le a constant, 7, for the normal/normal model. Hence: r(x, 18)). *(@) | corresponds to o? © IFE: 2010 Examinations ‘The Actuarial Education CompanyCT6-06; Emperical Bayes Crecibilily theory Page 7 (ii) The normalinormal model makes very precise distributional assumptions about both X; given @, which is N (6,02), and @, which is N (4.03). EBCT Model 1 makes no such distributional assumptions. (iv) The risk parameter, @, is a real number for the normal/normal model but could be a more general quantity for EBCT Medel 1. 1.3 Model 1: the credibility premium In the previous section the assumptions for Model 1 of EBCT were studied and the similarities between this model and the normalinormal model were emphasised. In this section a solution to the problem summarised in (1.5) will be considered, that is, the estimation of m(6) given the data X. The derivation of the credibility premium under this model is beyond the scope of the CT6 syllabus and is not covered here. Q Question 6.2 Without looking back to the previous chapter, write down the general credibility | formula. Result of EBCT Model 1 | The estimate of m(@) given X given by EBCT Model 4 is: (1-Z)E[m(ay]+2% where: KaY xin at and: | - | ae £[s?(@)| /var[m(o)] oe ‘The Actuarial Education Company © IFE: 2010 ExaminationsPage 8 CT6-06: Emperical Bayes Credibilty theory The first, and most important, point to note about the solution is that it is in the form of a credibility estimate. In other words, it Is in the form of formula (1.1) from the previous chapter with: &[m(8)] playing the role of 4 and > Xy/7 playing the role of X. in The second point to note is the similarity between the solution above and the solution in the normal/normal model, and, in particular, the similarity between ‘the formulae for the crecibility factors, ie (2.18) from the previous chapter and (18). Formula (1.6) is a straight generalisation of tonne (2.16) from the Previous chapter, since £[s?(@)] and var[m(@)] can be regarded as Seneralisations of af and of, respectively. We saw this earlier, The final point to note is that the formula for the credibility estimate given by (1.6) involves three parameters, E[m(a)], e[s? (@)] and var{m(0)}, which have been treated so far as though they were known, EBCT makes no distributional assumptions about the risk parameter @, unlike the Bayesian approach to Credibility, but the values of these three parameters need to be known. it can be Noted that these three parameters relate to first and second order moments as fhe deren, higher order moments or some other quantities, Thle Ie due to {he derivation of the credibiity premium and the detail leshind this is beyond the Soe, Of this course. Tha way in which these parameters would be estimated is discussed in the next section, GIFE: 2010 Examinations ‘The Actuarial Education Company1.4 CT6-06: Emperical Bayes Credibility theory Mode! 1: parameter estimation Im this section the solution to the problem of estimating m(8) given X will bo completed by showing how to estimate E[m(é)], var[m(9)] and e[s?(0)]. To do this a further assumption is needed: that there are available data from ‘some other risks similar, but not identical, to the original risk, This requires the Problem to be set up a little more generally, to make some extra assumptions and to change the notation slightly. An important distinction between a pure Bayes approach to credibility and EBCT is that no data are required to estimate Parameters in the former case but data are required to estimate the parameters in the latter. What is of interest is estimating the pure premium, or the expected number of claims, for a particular risk, just as in Section 1.2 and Section 1.3, and that this tisk Is one of WV risks in a collective. By a collective is meant a collection of different risks that are related in a way to be made clearer later in this section, For simplicity, suppose the particular risk that is of interest is Risk Number 4 in this collective. It is assumed that the aggregate claims, or clalm frequencies, for ach of these N risks for each of the past n years have been observed. Let Xy denote the aggregate claims, or number of claims, for Risk Number /, i=4,2,...N,inyear j, J=4,2,..,n. ‘These values are summarised in the following table: Table 1 Year 1 2 n Risk 1 Mii Xe Xin Number 2 Kari. |, Mes Xan N Xue Xna Xn ‘The risk numbers could refer to risks from different companies, insurers, shops ete. It is helpful to remember that small » represents the number of years while big represents the number of risks. ‘The Actuarial Education Company @ IFE; 2010 ExaminationsPage 10 Each row of Table 1 represents observations for a different risk; the first row, relating to Risk Number 1, is the set of observed values denoted X in Section 1.2 and Section 1.3. (Notice that X;,X3,... 'n_ in Section 1.2 and Section 1.3 have now become X44,X42,....X4q in this section.) Section 1.2 two assumptions, (1.1) and (1.2), were made about the connection between the observed values for the single risk then being considered, Section exactly the same assumptions for each of the N deka ny the ‘aremade. These assumptions are given as (1.7) and (1.8) below, For each risk i, /=4,2,...,N; ‘The dlairtbution of each Xy,.J/24, 2... i, dpenca on thesis of a parameter, Glenoted 6), whose value Is fixed (and the same for each value of J) but is unknown, (1.7) Given @;, the Xy's, J=12,...,n, are independent and identically distributed, (1.8) Notice that the risk parameter, which was denoted In Section 1.2 and Section 1.3, is now denoted @;, and that an implication of these two Cranubtions is that the risk parameters for different risks have different values. (However, as in Section 1.2 and Section 1.3, the risk parameter fare given risk does not change value from year to year,) ‘These twa assumptions show Something about the relationships within each row of Table 1 bue they do not show anything about the relationship between different rows, ie between aifferent risks in the collective. The assumption thet shows somett 19 about the relationship between different risks In the collective Is the following, For Jk, the pairs (6X) and (@,Xiq) are independent and identically distributed. (1.9) This assumption shows that the rows of Table 1 are Independent of each other. Two Immediate consequences of this assumption are: For i*k, Xz and Xm are independent and identically distributed, (1.10) ‘The risk parameters 6;,02,...y are independent and identically distributed, (1.41) ‘© IFE: 2010 Examinations ‘The Actuarial Education CompanyCT6-06: Emperical Bayes Credibility theory Pago 11 The connection between the different risks, fe rows of the table, is a result of the Teabtion that the Fisk parameters, 64,6....8y), are identically distributed, intuitively, this means that if, by some means, the Values of 2,95,...,8y were known, then something about the common distribution of the 6;'s would be known and hence something about 6), or, at least, about the distribution it comes from, would be known, The functions m() and s?() were introduced in Section 1.2, Keeping the same Gefinitions for these functions and applying them to all the risks in the collective: (6) —E (Xi 16;) $7 (6) = var(xy 16,) Notice that, as in Section 1.2, neither (Xj 16)) nor var(xy 16) depends on j Since, given 6,, the random variables Xjq,Xj2,...Xp are identically distributed. Notice also that, since 662..0y ara. identically distributed, £[m(6,)), £[5?(6))] and var[m(é))] do not depend on /. These are precisely the Parameters denoted E(m(8)], E[s?(8)| and var[m(@)] in Section 4.2 and Section 1.3 and which the collective will be used to estimate, More notation is needed. Denote: D Xpin by X, im and: x %r denotes the sample mean of the data from Risk Number ¢ and % = Genotes the sample mean of the data from all of the risks, ‘The Actuarial Education Company © IFE: 2010 ExaminationsPage 12 CT6-06: Emperical Bayes Credibility theory This new notation will be used to reformulate the credibility estimate of the pure premium, or the number of claims, for the coming year for Risk Number 1 in the collective as: (1-Z)E[m(0)]+ 2%, wher and: a 1.12 n+E[ s* (6) |/var[m@)] 3h {tls important to realise that formula (1.12) Is exactly the same as formula (1.6) Dut In this case the notation used for the data from the risk itself is X, and x; y rather than X and X,_ Estimators for E[m(@)], &[s*(@)] and var[m(o)] can now be produced. nae These estimators will bo functions of {[xy]”_ 1 + whose values will be known When the crecibility estimate of m(6,) is computed. Each row of Table 1 corresponds to a fixed value of @. Bearing this and the definitions of m(@;) and s?(6;) in mind, obvious estimators for m(@) and s?(6) are: X and (a-1" = (Xp - KP ist respectively. So we estimate m(6,) by the sample mean of the values for Risk Number i and we estimate s? (6,) by the sample variance of the values for Risk Number a. © IFE: 2010 Examinations The Actuarial Education Company76-08: Emperical BaYeS crodtilty Now, Elmcenl average” (over the aistribution of 8) of the values of (9) for different value? ‘of @. The obvious estimator for E(m(o)) 1 the average of the estimates oe matey: ob Cine ev ctee words, the eximator for E7700] © s so we estimate E{m(ay] YY the average of the sample means Kjsendtn This 2S equivalent to the Some snean of the full set oF (both direct and collateral) dat » value of 62(@) and so an obvious average’ x syro. ie itis xy le-d) ail this is a “nested” 5 Note that ums multiplied taper an be shown {m(9))- Unfortunately, this & would be an obvio¥! to be a biased esti ositively bias e produced by sr{mn(ay} a2” destimator for in fact it is P it can subtracting a corre vvar(m(@)) 15° N 0 af ire: 2019 Examinations = CompanyPage 12 (CT6-00: Empencal Bayes Credibilly theory This new notation will be used to reformulate the credibility estimate of the pure Premium, or the number of claims, for the coming year for Risk Number 4 in the collective as: (1-Z)E[m(o)]+ 2X, where: Rad xyin i and: 1.42) n+E s?(8)|/var[mi@)| 78) it is important to realise that formula (1.12) is exactly the same as formula (1.6) but in this case the notation used for the data from the risk itself is X%, and Xj rather than X and X; Estimators for £[m(6)], E|s?(@)) and var[m(@)] can now be produced. int lity estimate of m(é) Is computed. " These estimators will be functions of {ea} i » whose values will be known = when the credi Each row of Table 1 corresponds to a fixed value of @. Bearing this and the definitions of m(8,) and s*(@;) in mind, obvious estimators for m(6;) and 57(6;) are: x and (n-9" Y (xy - KP i respectively. So we estimate m(@,) by the sample mean of the values for Risk Number i and we estimate s? (6,) by the sample variance of the values for Risk Number /. © IFE: 2010 Examinations ‘The Actuarial Education CompanyPage 16 CT6-06: Emperical Bayes Credibility theory estimate var[m(8)) and & 70], whose values are then used to calculate Z The explanation is that in principle the credibility factor, Z, as given by formula (1.12), does not depend on the actual data from the risk being rated, Unfortunately, the formula for 7 involves two parameters, var[m(@)] and £[s?(@)], whose values are unknown but which can, in practice, be estimated ‘rom data from the risk itself and from the other risks in the collective, In other words, we include the data for the risk we're interested in in the calculation of var[m(@)] ana z[s?(@)] There is one further comment to be made about the assumptions for this model, Assumptions have been made concerning the identical distribution of the X's both from different risks and from the same risk. Ifthe Xy's come from different tisks it has been assumed that they are (unconditionally) identically distributed (see (1.10). If the Xy's come from the same risk, /, It has been assumed they are unconditionally identically distributed (see (1.3) and conditionally, given 6;, identically distributed (see (1.2) and (1.8)). A consequence of these assumptions is that neither E(Xy |@;) nor var(Xj16,) depends on j. This consequence ‘was the only step in the derivation of the credibility estimate (1.12) that required assumptions about the identical distribution of the Xj's. In fact, assumptions (1.8) and (1.9) (and the corresponding assumption in Section 1.2, (1.2)) could have been replaced by the following assumptions and it would still have been Possible to derive the same credibility estimate. Gtven 6}, the X's, J =1,2,...,, are Independent for each risk 4,142, (1.18) For Tek, the pairs (8X9) and (,Xqj) are. independant and the risk Parameters 61,62,....y , are also identically distributed, (1.17) For each risk, /, /=4,2,....N, neither &(Xy18)) nor var(xXy |6)) depends on de (1.18) © IFE: 2010 Examinations The Actuarial Education Company676-08: Empencal Bayes Credioity theory Page 17 None of the results or formulae in Section 1 would be altered in any way by making these weaker assumptions. The reason for making the slightly stronger assumptions, as was done in Section 1.2, is that they make the presentation a little easier, The reason for pointing out now that weaker assumptions could have been made is that this will help to link EBCT Model 1 with EBCT Model 2 in Section 2. Example The table below shows the aggregate claim amounts (in £m) for an international insurer's fire portfolio for a 5-year period, together with some summary statistics. Fill in the ‘missing cuties and calculate E[m(@)] and £[ s°(@)] using EBCT Model 1. Toal claim aout | 1g = | Year quake ZI com | if 2]3)4i)s 1 a [3 [2 [50 | 59 S04 393 ij fi o [fo | [0 | oa 173 | 3 as | 31 | 7 | os | 90 | 70 2153 4 4“ [alels|[a 2 2 Solution An this example, there are 1 =5 years and N =4 risks (= countries), 3, isthe average claims for risk 1 for the S-year period. So the missing entry is: 82 (4452469455471) /5 We can then calculate the other missing entry: (44-58.2)? +--+ (71-382)? ]=132.7 ‘The Actuarial Education Company © IFE: 2010 ExaminationsPage 18 T6-6: Emperical Bayes Credibility neory | To find the estimates we need Y, Which is: | 1d (304+ 684+74.0+582)/4= 62.75 a | | | | | | ‘We can then the evaluate the estimates directly: - E[m(9)) = 3 +132.7)/4= 101.2 Q | auestion 63 | Batimate the value of var{m(@)] eae @IFE: 2010 Examinations The Actuarial Education CompanyCT6-06: Emperical Bayes Credibility theory Page 15 Now consider formula (1.6) for the credibility factor for ERCT Model 4. The different “Ingredients” of this formula have the following defini interpretations: a is the number of data values in respect of the risk, Usually this will correspond tothe number of years of data available £[s?(0)] is the average Variability of data values from year to year for a single risk, ie the average variability within the rows of Table 4 var[m(@)] _ is the variability of the average data values for different risks, ie the Variability of the row means in Table 4 ooking at formula (1.6) for Z, the following observations can be made: () — Z is always between zero and one. (i) Z Js an Increasing function of m. This is to be expected ~ the more data there are from the risk itself, the more it will be relled on when the credibility estimate of the pure premium or number of claims le calculates (ili) Z is a decreasing function of e[s? (8)) This is to be expected ~ the higher the value of &[ s?(@)], relative to var[m(@)], the more variable, a L and hence less reliable, are the data from the risk itself relative to the data from the other risks in the collective, (™) 7 Is an increasing function of var{m(6)]. This is to be expected ~ the higher the value of var[m(@)], relative to e[s? (@)], the more variability There Is between the different risks in the collective and hence the less Likely It is that the othor risks in the collective will resemble the risk thet ig of Interest, and the less rellance should be placed on the data from thes other risks. ‘The Actuarial Education Company © IFE: 2910 Examinationsae 22 Page 20 06: Emperical Bayes Credibilty theory Empirical Bayes Credibility Theory: Model 2 Introduction Model 2 is @ generalisation of Model 1. ‘This section is therefore very similar to Section 1. The important thing is to note the differences in this section the techniques of Empirical Bayes Credibility will be applied to a Second, and slightly more complicated, model, ill be exactly the the sean Section 1. First of all, in Section 2.2, the problem will he stated and the assumptions set out. The problem will bo the same ae In Section 1, fe to estimate the pure premium, or the expected number of claims, in the coming Non an grat. The assumptions will be slightly diferent from thews in Section 1. Next, in Section 2.3, the credibility estimate for the pure Premium or expected umber of claims will be considered. Finally, in Section 2.4, the method of estimating the values of the paremeters that are part of the credibility estimate will be discussed, In Keeping with the terminology in Section 4, the model in this section will be referred to as EBCT Model 2 ‘The examiners will also refer to the model in this way. Model 2: specification The problem is to estimate the expected aggregate claims, or the expected number of claims, In the coming year fora given risk. Let Y,¥z, +Yq be random years from this risk, In Model 1 these quantities were called X,. You will see shortly why a different letter has been used. Wie aasiuined thet the values of ¥4,¥s1..sYp have alreaty been obser vediand the expected value of ¥,,.; needs to be estimated. @IFE: 2010 Examinations The Actuarial Education Companyal Bayes Credisilty meory 4 | Example | Find the credibility factor for the example above and hence calculate the EBCT premium | for each of the countries for the coming year, Solution We can use the estimates we have calculated to find the credibility factor: 5 prs = 08169 ss 90.33 Because we have the same number of time periods for each country, the credibility factor is the same for each. ‘We can then use the basic credibility formula P= ZX, +(1-Z)E[m(@)] to find the EBCT premiums for each country: Country P= 08169504 +(1— 08169) x 62.75 = 52.66 Country 2: 0.8169 x 68.4 +(1-08169) x 62.75 = 6737 Country 3: 0.8169 74.0+ (1 ~ 0.8169) x 62.75 = 7194 Country 4: 08169 x58.2 + (1— 0.8169) x 62.75 = 59.03 Q Question 6.4 @ () © ‘The Actuarial Ecucation Company Are the following statements irue or false for EBCT Model 1? 6 represents the “true” risk premium for a given risk. ‘The variance of | doesn’t depend on @. None of the random variables or parameters in the model are assumed to have a normal distribution, eee GIFE: 2010 ExaminationsCTE-06: Emperical Bayes Creaibilty tneory Page 2 So far the problem looks exactly like the problem in Section 1. The important difference between EBCT Model 1 and EBCT Model 2 is that Model 2 involves an extra parameter known as the risk volume, P;. Intuitively, the value of P) measures the “amount of business” in year j. For example, Py might represent the premium income for the risk in year J or the number of separate policies comprising the risk in year j. An important point to note is that the value of Pout atthe start of year n~1 is assumed to be known. Next a new sequence of random variables, X;,X2,..., is defined as follow: Xp =Yy1P; F=4 2,0 These ’,'s, which have been adjusted for the risk volumes, correspond more closely to the 27;"s in Model 1. In Model 1, we effectively assumed that P, was alway's equal to 1, fe the volume of business was the same for each risk group, The random variable X, represents the aggregate claims, or the number of claims, in year j standardised to remove the effect of different levels of business in different years. The assumptions that specify EBCT Model 2 are as follows. Assumptions for EBCT Model 2 The distribution of each X; depends on the value of a parameter, 8, whose value Is the same for each j but is unknown. (2A) Given @, tho X,'s are independent (but not necessarily identically distributed). (2.2) E(X;|@) does not depend on j. (2.3) | P;var(X;|8) does not depend on j. (2.4) i The Actuarial Education Company FE: 2010 ExaminationsPage 22 ‘CTO-06: Emperical Bayes Crediollty theory 2s i previous sections, 6 is known as the risk parameter for the risk, and, as for EBCY Model 1, It could be just a single real valued number or a more general Guantity such as a vector of real valued numbers, Assumption (2.1) is the standard assumption for all crei Idered here. See Accumbtons (2.2) and (2.9) from the previous chapter, and assumption (1.1), qesumation (2.2) corresponds to Assumption (1.2) fer BGT Model 1, but notice ‘hat (2.2) is slightly weaker than (1.2). Assumption (2.2) does not require the X;'s to be conditionally (given @) identically distributed, but only to be Senaitionally independent. There is no assumption in EBCT Model 2 that the %;'s ar9 unconditionally, or conditionally given 0, identically distributed. coc aring the sbove assumptions with assumptions (1.16), (1.47) andl (1.18) for EBCT Model 4, if all the P;'s are equal to 1, then EBCT Model 2 is exactly the same as EBCT Model 1, Having made assumptions (2.3) and (2. 4), m(@) and s?(@) can be defined as follows: m(8)=E(x; |6) 3? (0) =P) var(x, |) The definition of m(6) corresponds exactly to the definition tor EBCT Model 4 in Section 1 but the definition of s? (@) is slightly difforont, Remember to include the F; factor in the definition of s?(6) Zo gain a little more insight into assumptions (2.3) and (2.4), consider the following example. Suppose the risk being considered is mace up of a different umber of independent policies each year and that the number of policies in year 1 's Fj. Suppose also that the aggregate claims in a single year from a single Policy have mean m(@) and variance s2(6), where m() and s?() are fanctions of @ and @ Is the fixed, but unknown, rlsk parameter for all those Policies. Now let Y) denote the aggregate claims from all the policies in force year J @IFE: 2010 Examinations The Actuarial Education Company23 (CT6-08: Emperical Sayes Credibility theory Then: Ely) > m(a) var(¥;)=P;s?(@) &(X;)=m(8) Pj var(X;)=s?(0) This example satisfies assumptions (2.3) and (2.4). Model 2: derivation of the credibility premium The problem was stated rather loosely in the previous section as the estimation of the expected value of ¥;,.4, given the values of Y;,Y2,...,.Y,. We can now be rather more precise about this. The quantity to be estimated is the mean (given ) of Ynot- This is given by P,.4 (8). Since in the model P,.1 is known at the start of year n=1, the problem Is to estimate m(@). The data available are the values of each Y; and its corresponding P; for /=1,2,..., 9. As for Model 2 the full derivation of the credibility premium under this model is beyond the scope of the CTS syllabus and Is not covered here. The solution to the problem, /e the “best” linear estimate of m(@) given X is given by: E[m(a)]E| s*(@)| var[m(oy] Py en Sr £[s?(0)) /var[m@)) in Remember that ¥; is just P,X; The Actuarial Education Company © IFE: 2010 ExaminationsPage 24 78-06: Emperical Bayes Credibility tneory This can be written more attractively as follows: | Result of EBCT Model 2 | | | The estimate of m(2) given x given by EBCT Model 2 is; | 2X ~(1-Z)E[m(o)] (2.5) | i on Ja, n a | B= Sax,/Se-dy,/ Se, | 4 fh Jj=t ane: z ae | DP + £/57(0)] vary] sal - ‘This demonstrates the similarities to and the differences from the Model 4 result. | Question 6.5 Check that the “attractive” version really is the same. Note carefully that the above result tells us how to estimate the value of X for the coming year. If we want to estimate ¥,.1, the ageregate claim amount for the coming year, we have wo multiply our estimate of X by Pj, the risk volume for the corresponding year. © IFE: 2010 Examinations The Actuarial Education Company24 C¥6-06. Empericas Gayes Ci Here are some additional points of note about this solution: () all the P's were equal to 1, the solution given by (2.5) is exactly the same as the solution given by (1.6). This is as it should be since if all the 4's are equal to 1 then EBCT Model 2 is exactly the same as EBCT Model a (ii) As for EBCT Model 1, the solution given by (2.5) involves three parameters, E[rm(6)], var[m(@)] and £| s?(8)]. The way in which these parameters are estimated is explained in the next section. Mode! 2: parameter estimation The procedure for estimating the parameters £[m(6)], var[m(@)] and | s? (@)] for ESCT Model 2 follows exactly the same steps as the procedure for EBCT Model 4, which was studied In Section 1.4, As before, we are now moving to a two-dimensional data set with rows representing Gifferent risks. It is now assumed that the risk that is of interest is one of a collective of N risks and that there exist data in the form given in Section 2.2 for each of these N tisks for each of the past m years. These data consist of values for the aggregate claims, or the number of claims, and the corresponding risk volumes. Let Yy be a random variable denoting the aggregate claims, or the number of 11, 124, 2,..N, and let Py bo the claims, for risk number / in year j, /=4,2, corresponding risk volume, For each j and j define: Xy=Yy! Py The Actuarial Education Company © IFE: 2010 Examinations76-06: Emperical Bayes Credibility he data are summarised in the following table, which corresponds to Table 1 for EBCT Model 1: Table 2 Year 1 2 oo n 1 Yas Pay Yaa Pro se Yoo Pan Z Yeas Pas Yeas Pea “ Yen Pan Risk Number N YoPan Yoas Pua : Yon Fin For simpllety itis assumed, as was done in Section 1.4, that the Fisk that Is of particular interest Is Risk Number 1 in this collective, Thie meane ther what were denoted ¥,, Py and X; in Section 2.2 and Section 2.3 are now denoted Yay, Phy and %1j respectively in this section. The problem Is to estimate the expected value of X;,,.4 and the solution to this Problem has already been given by formulae (2.5), remembering that x) and %; In (2.5) are now denoted X,) and P- The purpose of the data from the other risks in the collective Is purely to help to estimate the parameters E[m(8)], var[m(9)] and &[ s? (6) that appear in (2.5), These other risks in the collective satisfy assumptions exact the same as assumptions (2.1), (2.2), (2.3) and (2.4) for Risk Number 1. These assumptions are as follows. For each risk /, i the chetiullon of each Xy Jom 2;). ch tegandslon ai parse: 4, whose value is the same for each j but is unknown (2.6) Sivan 4, the Xy's are independent (but not necessarily lontically distributed) (2.7) there exists a function m() such that m(6,)= &(xy 14)) (2.8) there oxists a function s() such that s?(9, )- Py var(Xy 16). (2.9) © IFE: 2010 Examinations The Actuarial Education Company676-06: Enipentcal Bayes Cresiality theory These four assumptions show that each risk in the collective satisfies the same assumptions as the particular risk that is of interest. The following two assumptions, which correspond to (1.9), (1.10) and (1.11) in Section 1.4, show the connection between different risks in the collective. The risk parameters 6,,62,....0y, regarded as random variables, are independent and identically distributed. (2.40) For iz k, the pairs (@,Xy) and (0,,Xjm) are independent. (2.41) Notice that, since the 's are identically distributed, the values of E[m(6))], var[m(9;)] and E[s?(6;)] do not depend on i so that they can be denoted by £[m(9)], var[m(@)] and E[s?(@)|, respectively, as in Section 22 and Section 2.3. As in Section 1.4, more notation is needed. Denote: Di by F (2.12) i ae i yA byP (2.43) is fh . = (nn) > (1-5 1P) by P* (2.14) in S rpg Bi by & (25) ia Ws itn = D SAK IP-S SL ryxylP wy X (216) im A fs This notation is all given in the Tables (Page 30). Notice that what was denoted X in formula (2.5) is now denoted X, and that X now has a different definition. (This was exactly what happened in Section 1.4.) Notice also that X; and X are weighted averages of the Xy's, the weights being the risk volumes Py. ‘The Actuarial Education Company © FE; 2010 ExaminationsPage 28 CT6-06: Emperical Bayes Credibility tneory With this now notation the credibility estimate of the pure premium, or number of lalms, per unit of risk volume for the coming year for Rick Number 4 in the collective originally given by formula (2.5) can be reformulated aet 2X1 +(1-Z)E(m(a)] (2.17) where: Re Zany /S0y = and: Py i Qy= XA + [8 (0)|/var(mia)] in Note that this credibility factor is speci risks there will be different values of Z jc to Risk 1, so is denoted by Z;. For the other {t ts important to realise that this is exactly the same as formula (2.5) but is written In the notation of this section rather than that of the previous section. Unbiased estimators for E[m(9)], var[m(a)] and £ 5? (@)| can be proposed " based on the observed values {Mery Hy . © IFE: 2010 Examinations The Actuarial Education CompanyCT8-06: Emperical Bayes Credibility theory Page 29 These estimators are: Parameter Estimator E[m(a)] x (2.18) u < 1 sat? &|s*(o) NAY ("Py (Xp - (2.19) a ia ae at var[m(@)] ee |(on-4 YY Py (xy - XY (2.20) is Ja N 2 HX nD Ay (xy) The points of note about these estimators are: “a (il) (iii) They are in exactly the same form as the estimators for EBCT Model 1. Look back at formulae (1.13), (1.14) and (1.15). In particular, if all the Py’s were equal to 1, then the two sets of estimators would be identical. It can happen in practice that formula (2.20) gives a negative value even though var|m(6)] must be non-negative. In such a situation the estimate of var|m(8)| is taken to be zero. A similar point about the variance estimates of EBCT Model 1 was discussed in Section 1.4. Formutlae (2.18), (2.19) and (2.20) are all given in the Tables, And you'll be pleased to know that () ‘The proofs that (2.18), (2.19) and (2.20) are unbiased are beyond the scope of the syllabus. ‘The Actuarial Education Company (© IFE: 2010 ExaminationsPage 30 Example The table belows show the volumes of example on Page Error! Reference CT6-06; Emperical Bayes Credibily theory business for each country for the insurer in the ‘source not foun Caleulate the EBCT premium for Country 1 using Model 2 Solution ‘The original data forthe total claims is Yj and the able above gives values of Fy. ‘The claims per unit volume Xp / By are shown in the table below. ‘Tot! claims per unit volume (Xj) Country (i) Year (/) 1 2 3 4 5 1 | 4.000 3.533 3.231 3.125 5.900 2 3.200 5.071 2.909 4.867 2.333 = 17.000, 6.750 12.667, 5417 22.500 4 2.000 1.486 2.300 3438 7.100 ‘We can then calculate B, P and P* © IFE: 2010 Examinations ‘The Actuarial Education CompanyCT6-06. Emperical Bayes Credibliy theory Page 3 i | The figures are given in the table below oe ] | Ak 1 | 6 | 2 101 3 35 4 113 Pa3is P* = 1181 ” Xo Furthermore, X; and Y can be calculated as: ¥,=)°¥,/% and ¥= > y,/P. 5 iat j=t ci ze = | epi a | Siacaen | ome 1 3818 37.13 38.94 2 3.386 1159 14771 3 los7t | __-1237.82 2,756.56 4 2575 2673 «|~~«a92.04 | | ¥=3984 So this gives: E[m(@)] =. = 3.984 From the other columns in the table, we get: lal Sle p(y, 2? z[°@] ada v-¥) = (57.13 +111,59 + 1,237.82 +267.73)/16 = 10464 ‘The Actuarial Education Company © IFE: 2010 Examinations70-06: Emperical Bayes Credibility ineory | Also: var[mi9)] = (58.94 + 147.71 +2, 756.56 + 492.04)/19 -104.64] “Tisi = 6.539 ‘The credibility factor for Country 1 is: RB; E at Oe 5 ALTO] ORE abe Bir raat 6.539 The risk premium per unit volume is: 44%, +(1-Z)z[nx0)] = 0. 8048 x 3.818 + (1 - 0.8048) x 3.984 = 3.851 Since the volume for Country | forthe coming year is 20 units, the EBCT premium is: 20x3851=7701 Question 6.6 =a Calculate the EBCT premiums for Countries 2, 3 and 4, @IFE: 2010 Examinations ‘The Actuarial Education CompanyCT6-06: Emperical Bayes Credibility theory Page 33, = Question 6.7 Explain the effect cach of the following changes (acting in isolation) would have on the Value of the credibility factor for EBCT Model 2, (@) | s*(0)] is inereased. (©) var[m(6)] is reduced. (©) The unit of currency is changed from £ to $. (@)—Alllthe P,’s are increased by the same factor. Does the eredibility factor “behave” as you would expect? We now finish this chapter with two exam-style questions. Past Exam Question (Subject 106, September 2000, Question 3) ‘The table below shows annual aggregate claim statistics for 3 risks over 4 years. ‘Annual aggregate claims for risk #, in year j, are denoted by Xi, Risk, f 1 2,517 4,121,280 2 7.814 7,299,175 5 2,920 3,814,001 (@ —Caloulate the vatue of the credibility factor for Empirical Bayes Model 1. [3] Gi) Using the numbers calculated in (i) to illustrate your answer, describe the way in which the data affect the value of the credibility factor. [4] [Total 7] The Actuarial Education Company © IFE: 2010 ExaminationsPage 34 76-06: Emperical Bayes Credibiliy theory Exam-style question An actuary wishes to analyse the amounts paid by a group of insures on their respective Portfolios of commercial property insurance policies using the models ot Empirical Bayes Credibility Theory. ‘The actuary obtains the following information about the amounts of claim payments made and the number of Policies sold for each of three different insurers. The data obtained are as follows. 1 Year 1 Year 2 Year3 Year 4 InsurerA | £14.2m £15.8m £22.7m £19.0m 163 189 252 199 Insurer B | £58.6m £63.1m £81.0m £64.2m 4,435 4,761 5,576 4,581 InsurerC | £123m £132m £161m £133m 16,184 114s | 20,102 18,000 © Analyse the data using EBCT Model 1, and calculate the expected total claim Payment to be made by Insurer B in the coming year. (61 (i Analyse the dat using EBCT Model 2, and again calculate the expected payout amount for Insurer B in the coming year, assuming that the expected number of Policies sold for the coming year for Insurer B is 4,800. You may use the summary statistics given below, which have been calculated using the formulae and ‘notation given in the Tables, again working in millions of pounds, Subscripts 1,2 and3 refer to Insurers A, B and C respectively, {8} DAY, -X)? = 0014667 DA %,)- FP =5106461 DAU, - 2) = 0.006103 x) DA Gy 0003979 Gil) Comment on your results, 22) [Total 16] @IFE: 2010 Examinations ‘The Actuarial Education CompanyCTo-0s. Emperical Sayes Creaioity inaory Page 35 Z| Chapter 6 Summary Empirical Bayes credibility ‘This approach to credibility theory assumes that the claims for each risk are dependent on an underlying risk parumeter @. However, no assumptions are made about the form of the distribution of the claim amounts. ‘The credibility premium can be expressed in terms of a credibility factor, which depends on the mean and variance of the conditional claim distribution. These ‘quantities can be estimated based on data derived from a number of different risks. ‘The table below summarises the models. Model Quantity [Assumptions about the | Assumptions about the | estimated underlying distribution conditional claim ofthe isk parameter | distribution X19 EBCT Model 1 | risk premium | a definite but unknown a distribution with a distribution mean m(@) and variance s'(8) EBCT Mode! 2 | claim Model? adj fk just for ar | volume of business Empirical Bayes Credibility Theory model 1 Definitions .X) represents the amount or number) of claims m(8) = E(X; 18) Credibility factor: s°(6) = var(X, 16) { eso] Credibility premium: E[m(6)\ X]=Z X, +(1-Z)E[m(@)] ‘The Actuarial Education Company © IFE: 2010 ExaminationsPage 36 \To-U8: Emperical Bayes Crecioity tneory | eSeepicioan Bayes Credibility Theory | Definitions: m(8)= E(x, 8) Credibility factor: /\ xX a Credibility premium; ‘The formule for estimating Elom@)], Ef ave given in the Tables (Pages 29 & 30), © IFE: 2010 Examinations mode! 2 J; Tepresents the amount (or number) of claims HP) (P; constant) $°(0) =P, var(X, |6) i. eL call var[m(@)] ZX, +(1-Z)E[m(y] $°(@)] ad var[m(6)] for the EBCT models ‘The Actuarial Education CompanyCT6-U8: Emperical Bayes Credibility theory Page 7 Chapter 6 Solutions Solution 6.1 Here, the risk parameter @ represents the handicap for a particular player. The distribution of the values of 8. which will vary from one player to the next, are assumed to be determined by a definite (but unknown) probability distribution. ‘The shape of this distribution could be estimated by finding out the handicaps of a random sample of players. (In order for the EBCT model to apply here, we'll have to assume that players are secretive about their handicaps, so that the actual handicap for a given player isn’t known with certainty.) X; is a random variable, representing the score obtained by a particular golfer on the sth day. X;|@ represents the score obiained by a particular golfer whose handicap is 6 on the jth day. m(8) = E(X,|0) is the (theoretical) average score for a golfer whose handicap is @. Note that, although the average score and the handicap are related, m(@) will not be equal to 8. In this example, m(@) will be an increasing function, since golfers with high handicaps (fe not very good players) will have high average scores (de they will take lots of shots to get round). $96) = var(X,,|@) is the (theoretical) variance in the scores obtained by a golfer whose handicap is @. s(@) will take lower values for players with a high standard of play (a Jow @), since their scores will be fairly consistent, whereas the scores for novices will tend to fluctuate over a much wider range. So again, s*(@) will be an increasing function. You may find it helpful to keep this example in the back of your mind to help you understand how the EBCT models work. Solution 6.2 ‘The general credibility formula is: ZX +(1-Z)u ‘The Actuarial Education Company © IFE: 2010 ExaminationsPage 36 (C76-06: Emperical Bayes Credibilty ineory Solution 6.3 We need to find: var[moy)= 23H, Fp edhe, sia mi 4ie ‘We've already worked out the second term (more or less) when we found £[s°(@)]. So, putting in the numbers gives: to (58.2-62.75)?] Solution 6.4 (@) False, 0 4s just a risk paraimeter that reflects the likelihood of claims, ‘The tue risk premium for a given risk is £[m(o)|x]. (©) False, ‘The variance of 4°,|@ is s°(8), which is a function of @ © True. In fact, none of the quantities in the model are assumed 10 have any specific type of distribution, © IFE: 2010 Examinations ‘The Actuarial Education CompanyCT6-06: Emperical Bayes Credibility theory Page 39 Solution 6.5 We start with X +(1-Z)£[m(8)]. This is equal to: ‘ [2] a ___ vari] Spf, AFO xe Se Var[ni@)] x E[m(@)] a ‘We can see that the second term here is equal to the first part of the original expression for the best linear estimator, If we cancel a factor of )° P, in the first term here, and ia remember that )’ PX, = )°Y;, we see that the first term here is the same as the ial second part of the expression for the best linear estimator. So the altemative form is equivalent to what we got before Solution 6.6 ‘The table below shows the figures forall four countries: Country Credibility factor | Risk premium per | EBCT premium unit volume 1 0.8048 3.851 77.0 2 0.8632 3.468 86.7 3 0.6862 8.504 85.0 4 | 0.8759 2.750 33.0 ‘The Acivarial Education Company © IFE: 2010 ExaminationsPage 40 76-08: Emperical Bayes Credibility theory Solution 6.7 ‘The formula forthe eredibility factor for EBCT Model 2 is ope, #{@]) 2a 38 2 “var[n]| Since increasing £{ s°(@)] increases the denominator, Z decreases, as expected. (0) var[m(8)] represents the variation in the average claim amount for different risks, Since reducing var[m(@)] increases the denominator, Z decreases, as expected, (©) Changing the unit ofcurency should not affect the credibility factor Since the quantities £[.(6)] and var{m(6)] are measured in units of £7, their ratio is dimensionless. So changing the unit of curency to § would not affect Z. (@) The ,’s specify the relative weightings to be put on the claims for each year, We would not expecta uniform increase applied to all the weightings to affect the ‘credibility factor. Since the definition £[s?(6)]= P, var(x,|6) includes P, factor, but var[m(@)] doesn’t, the ratio £[s*(@)]/ var[m@)] varies in proportion to the 77's. So any extra factor incorporated in the P, 's would cancel out, leaving Z unchanged. © IFE: 2010 Examinations The Actuarial Education CompanyCT6-08; Emperica: Bayes Credibiity tneory Page si Solution to Past Exam Question (Subject 106, September 2000, Question 3) (i) Credibitity factor E[s*(0)] is estimated by the average of the sample variances, ie: 4.121,280+ 7,299,175 +3,814,001 ,078, 152 The sample mean of the ¥,’s is: 2,517+7,814 2,920 41? and the sample variance of the T,’s AIT) +(7,814~4,417)° +(2,920-4,417)? (2,517— 3,695,309 So var[m(@)] is estimated by: $695,209 —1x5,078,152 = 7,425,771 The Actuarial Education Company © IFE: 2019 ExaminationsPage 42 76-06: Emperical Bayes Credity theory (3) How the data affect the value of the credibility factor Z 's an increasing function of 7, the number of years of Past data. If we have more than 4 years of past data, the eredibility factor will increase, 7 is a decreasing function of £[s?(9)]. if £[s?(6)] increases, eg if the variance of the claim amounts from one or more of the risks were to increase, then the value of the credibility factor would fall Z is an increasing function of var[ m(0)]. IF var[m(0)] increases, eg if there was greater variation between the individual ‘sample means, then Z would increase. Solution to Exam-style Question () Analysis using EBCT Model 1 Using EBCT Model 1, we obtain the following numerical values 350) -% | T Year1 | Year2 il ] | Insurer A 14.2 15.8 22.7 19.0 17.925 14.1158 TewerB | 386 | 621 | so | a2 | o6ms| sass Insurer C 123 132 161 133 | 137.25 270.9167 | ‘Using the standard formulae for the EBCT estimates, we get: ¥ = 417.925 + 66725413725) = 73967 £[s*@)] =4(14.1158-+96.4358+ 270.9167) = 127.1561 var[m(@)] = 4[ 07.925 73.967)? + (66.725 ~ 73.9677 +#037.25~73.967)? |-4%127.1561 =3,567.1562 @ IFE: 2010 Examinations ‘The Actuarial Education CompanyCT6-06. Emperical Bayes Creuibilily theory Page 43, So the credibility factor for the model is: n 4 els @)| 44 L2TASEL ” 3,567.1562 Z = 0.9912 ~var[m(@)] So the EBCT premium for the coming year for Insurer B is: 66.725 0.9912 + 73.967 x 0.0088 = £66,79m (i) Analysis using EBCT Model 2 We first need to use the data to calculate the statistics we need. Using the notation given in the Tables, we have: F = 163+189+252+199 = 803 ‘The same approach for the other insurers gives the values 19,353 B= 71,729 For the whole portfolio we have; P =803+19,353+71,729= 91,885 803 19,353 71,729 - 129353), 71,799(,-2L729" Af rams 19.353: a) Me ( i) = 2,891.5793 Calculating ¥, for the first insurer: 142+ 1584227419 803 z, ‘The Actuarial Education Company © IFE: 2010 ExaminationsPage 4+ © T0-U6: Emperical Bayes Creaiolity theory Similar calculations for the other insurers give: 0.013791 0.007654 For the whole portfolio we have: 142 +158 46041614133 91,885 = 0,009660 Now using the summary statistics given in the question, we can find our estimates for Elm(6)], var[mn(@)] and El 52(6)]: Elm@)] 0.009660 L L var [mi Sear epenl 70-1064 0.336408 + 0.292641) at [m(8)] sme 161+ 0.336408 + 0.292641) ~Fpg3 01014667 = 0.006103 + 0.003979] = 0.001794 z[*@)] (0.014667 + 0.006103+ 0003979) = 0.002750 3x3 So for Insurer B, the credibility factor is: 19,353 ee acts tpapaiceeeer se 0.0001794 So the credibility premium per unit of risk volume for Insurer B is: 0.999208 x 0.013791 + 0.000792 x 0.009660 = 0.013788 So assuming a risk volume in the coming year of 4,800, the risk premium for Insurer B is £66.18m, © IFE: 2010 Examinations The Actuarial Education CompanyCT6-06: Emperical Bayes Crecibitty theory Page 45 (iii) Comment ‘The two models give fairly similar results. ‘The estimate in Model 2 will depend on the prediction of risk volume for the coming year. In both cases we have used a very high value for the credibility factor. So we are effectively ignoring the data from the other insurers, and are basing our estimate almost ‘entirely on the data from Insurer B. This seems pretty sensible, given that both the Volume figures and the average claim amounts appear to be pretty variable between the three different insurers. So we are tempted to ignore the data from Insurers A and C, and focus on the information that we have for Insurer B. ‘The Actuarial Education Company © IFE: 2010 Examinations