stewart coutts · 2012-01-19 · stewart coutts -~-&1_r71 this peper outlines a verr aiiiple...
TRANSCRIPT
MOTOR PREMIUM RATI~G
Stewart Coutts
-~-
&1_r71 This peper outlines a verr aiIIple structure tor the praotioal
analysis ot .,tor presiUlis trca a tJI( point ot view. It indioates
that it is possible to !IOdel dirterent types ot o1&illS oost. These
results are OOIIbined together with sOlIe st.ple e~nOllio assumptions,
to arrive at proeld.u... Then the peper develops a -points system
"'ioh is st.llar to a nwaber ot presiwa systess operated by UJ(
Insurance Ccapanies. This -points syst.- is used to ocapare
ditterent sets ot assumptions. Finally, an analysis ot surplus is
described vith lID ex_pie.
1. INTRODUCTI<If
The purpose ot this paper is to outline sOlIe ot the statistical and
tec~.ni~al aspects ot a premium basis tor motor insurance. At present,
only broad outlines are availab.le in the tJI( Institute ot Actuaries'
literature and 10 particular, they do not highlight the pragmatic aspects
ot the problem. There have been sOlIe papers written in the ASTIN
bulletin, e.g. Pitkanen (1974), however, they are senerally theoretical
based. The tirst tull analysis tound in the literature was written by
Kahane and Levy (1975). However, the emphasis was statistical modelling
rather than discussing the practical p~oblems involved in premium
rating. Recently, two groups have published stUdies concerning the
problems ot.rating. In The Netherlands, a group ot actuaries were asked
to revise the motor rating structure and this is written up in Uethgrland
Gr~~p Report (1982). At the tJI( General Insurance Statistical Group
meeting in Strattord-upon-Avon (1982), there were tour case studies trom
different countries on premium rating. Hence, in seneral, there is
little written assistance to help either the actuarr new to motor
insurance or the actuarial student tryinS to appreciate the practical
problems ot premium ratins.
The theme ot this paper will be to em",asise a detailed within-porttolio
analysis, taking into account simultaneously, inter al1a, some ot the
major UK motor underwriting tactors and separate st4tistical analyses of
claims frequency and costs, together with expenses, to ar'rive at a
breakeven ploemiUIII. By pertorminS this detailed analysis, it is believed
that \','e pel"son responsible tor the premium decision l-rill be able to
1'(;:; ;;ri ct his at \;ention to the seroiti va are3S uhere jlld<3·-:I~ont h3.:l to be exercilled.
-6-
This paper will describe a procedure which ill applicable in a competitive
.,tor premium urket situation, suohas in the U1C. However, it is
contended that in CO\mtries where fiud preadUIII rating set by a tariff is
in operation, an in-depth analysis is also necessary. This will enable
.an~nt to judge where in their ~tor portfolio the business is
profitable. In addition, the analysis will bring out warnings that where
all CCDpany pooled data is uaed, it oan be harmful, in particular to
.. ller companies.
The structure for this paper is lUI fo11<*s:-
Section 2
Sectio~ 3
Section 4
Section 5
Section 6
Section 7
Section 8
SecUc.n 9
Section 10
Section 11
Section 12
Section 13
Diacusses the reason for a detailed data breakdown.
Gi ves the general background to premium rating by asking,
wbJ project? ~
Introduces the Group which is involved in the decision
process.
Gives the fonnula to be used.
Defines the database.
Outlines the general principles for claim frequency and
cost statistical analyses.
Discusses bodily injury analysis.
Outlines the inclusion of expense allocation.
Discusses economic factors.
Performs the actual premium calculation.
Discusses the presentation of the premium rates.
Discusses a points table analysis.
-7-
Section "
Section 15
Section 16
Compares different seta of assumptions used in the premium
bases.
Marketing aspects of preaiu. rating.
Anslysis of surplus.
Finally, to keep this paper to manageable length, it has been necessary
to restrict some of the discussion o~ statistical aspecta to a minimum
and only give references to papers Which will give details of these
analyses.
2. DATA BREAKDOWN
The first question to consider is Whether the premium rates are to be
reviewed in an overall fashion, e.g. adding 10J over the whole portfolio,
or whether to apply selective increases to different sections of the
portfolio. It is argued that with the use of computers, a selective
breakdown of underwriting factors should be undertaken, which by.
aggregating the results can automatically give the overall level of
premium adjustment needed. If, however, the system of analysis is only
gcared to the overall review, it is very much harder to obtain
information about the selective parts of the portfolio.
If the data are sub-divided by underwriting rating factors, we are left
with figures which are small in exposure and claim numbers, therefore,
simple averages will be.suspect. Hence, it is suggested that simple
statistical modelling be preferred. This will enable:-
1. Extension of actuarial judgement to small databases which implies
that the portfolios of small Insurance Companies can and should be
analysed. This statement is strongly worded, but it is believed that
it can be accomplished since the theory of statistics in the· past 10
years has made 'space age' progress in the analysis. of small
databases, in far more critical and sensitive. areas than motor
insurance, namely, medical statistics. An urgent plea is made to the
actuarial profession to step up research on small databases, in order
lhat small co:npanies ceon employ actuaries with the relevant expertise
to offset uninformed comments such as 'the data is too scanty to
support any meaningful analyses'.
-8-
2. When the data is analysed in sufficient detail, then the effects of
portfolio changes are reduced and judgements on these effects can be
.. de with confidence.
3. The establishment of statistical structures, however simple, provides
'bench marks' that can be used as a basis to monitor actual results
as they emerge. This will be discussed in the final section of the
paper.
_. As will be indicated later, analysis reveals that the process of
premium rating involves many different assumptions. Changes in some
of these assumptions can affect the premium rates significantly, e.g.
different bodily injury assumptions, or in varying mathematical
models. The effects can only be assessed if the data is suffiCiently
detailed and the structure sufficiently defined.
In spite of these pOints, it has been argued by non-actuaries that any
actuarial input which might alter rates within the motor portfolio is
practically irrelevant, when compared with overall marketing
considerations or in countries where the motor rates are fixed. Hence,
the stati~tical process is considered a mere theoretical exercise, the
cost of which, gh'en the personnel involved, is hard to justify. In
addition, actuaries within the general insurance field have tended to
depreciate any detailed analyses, as recently as the UK General Insurance
Statistical Group conference at the one-day seminar on Premiums at
Stratrord-upon-Avon. One argument was that an experienced actuary did
not need to perform any detailed premium analysis, since he should ba
aware of the premium situation and adjust the rates accordingly. It ~~s
also al'gued that past analyses should be a sufficient basis for changes
in premiuM, if selected data were collected in order to monitor the
process.
It is sur.Le~ted tr.at all these remarks are half-truths. True, actuarial
rat.;:'! rc..y differ ct';1'llder3.bly frol!! meorket rates. Howev.cr, it is the
management's decision and they should be aware that the rates eharged may
in fa~t generate potential losses, the size·of which should be
quantified. If decisions are mC'.de without having all facts available,
t.hc'.n t.h1 r' l'::~.>:t. L~ (_I):.~: fl!::!".",!I~ r.oor mon,:l~C:·I~(~nt. A topl~~l example i~ that
tt>e undeJ'\ll'itir\f', r"tjnr factor 'car age' is generally ignored or given
insufficient weir:ht in market structure; in part:icular, from s\".atist~c~
-9-
anilable, newer oars are being undercharged. On the opposite side, a
detailed analysis II&y reYeal unsuspected _rketins opportunities within
the present stNoture. Herein lies the area ot greatest potential tor statistical analysis. In addition, in countries where there is fixed
pr_ium ratins, set by a taritt, this _y eventually break down, e.g. in
the Netherlands. Hence, it is advisable to be able to pertol'll a detailed
analysis to cover this eventuality.
As tor the argument that the experienced actuary haa litUe need ot
analysis, it is aocepted that jud ..... nts could otten be _de without any
in-depth investigation, how is this aohie'ftld? The existins body ot knowledge has been accumulated largely by trial and error. In tact,
however, the actuarial protession ought to stri'ftl to establish detailed
procedures. It is contended that the present international actuarial
lit8~ature does not give sutticient intonaation tor this purpose.
The final argument, appeal1ns to past analyses, tacitly assumes that
premium stNctures are stable over time. It is agreed that frequent
changes to the system may be undesirable, but regular analyses are
necessary to 'ftIrity assumptions _de in the original caloulations. In
particular, if the database is small, this will necessitate regular
checks to judge whether the statistical interences were reasonable. For
the UK, in the late 1970's, _ny companies reviewed their premium leVels
quarterly because of the rising rate ot inflation. Furthermore, the
\.D'1derlying insurance risk pattern may also vary over time.
Finally, cost and time are put torward as reasons for not performins
regular analyses. However, with the advent ot microprocessor technology,
it is believed that the cost has been cut to a minimum and time reduced
to an irrelevant factor.
In sunmary, a breakdown ot the data to take into accoun"t rating factors,
claim frequency and cost is tundamental to establishing premium bases.
To analyse data in this way, modern statistical techniq\.les must be
understood and employed regularly since the systems are not necessarily
stable. To monitor the results, some torm ot analysis of surplus has to be carried out.
-10-
3. WHY PROJECT?
Before discussing what background information is required in practice, to
make judgements about premium levels, it would be useful to visualise the
time span involved in the premium analysis. This is best explained by
means of a simple example.
Consider a company that reviews its premium rates on 1st October 1980 and
let us assume that these rates are expected to ~e in force for one year.
New recommendations would have to be made in practice 3 months in
advance. This time lag would be needed to alter computer output and
prepare documentation for the rate book to be passed to the broker.
The average policy will be effected halfway through the period over which
the'- premium series is expected to be in force, i.e. on 1st April 1981.
This policy will be on risk for one year and, should it have a claim, the
claim date will on average be 6 months after date on risk, i.e. 1st
October 1981.
The material damage costs will be expected to be settled within 3 months
from the date of accident. However, the third party bodily injury costs
will take on average 2 years to settle. Prospectively, the average
future time span is 3~ years from the date which the premium decision has
to be made, 1.e. by the 1st July 1980. Hence, data concerning claj.ms and
expenses have to be projected to the 1st October 1981 and beyond.
Expen~e data will he based on information which is reasonably up to dat.e
at the 1st July 1980 and is then projected to 1981. The claiM data will
be built up from claim numbers, material damage costs and third party
bodily injury costs. All but the latter costs will be based on recent
data, say 1980. However, the most reljable third party bodily injury
data is likely to be 5 years old, i.e. clai~s occurring in 1975. The
reason why the average settlement figure of 2 years is not appropriat.e i~
becalJ~c, :In I'rllctiC'e, the largp.r end proportionately more importa'"1t
claims take in excess of 5 years from date of accident to settlement.
-11-
Hence, the total time span to be projected is on average 8l years. This
large tt8e span necessitates a nuaber of subjective decisions, the main
ones being:-
(i) Are third party claims occurring in 1975 relevant in respect of
liability olaillS expected to occur in 1981 and to be settled on
average in 1983 or later? The .. in probl.. lies in the
settlement figures, as court judg .. ents change with changing
social conditions. This has been commonly referred to as
judgement drift. If the Company has an individual liability
claim eati .. tion process (referred to as a manual basis) then it
is possible, though not necessarily reliable, to use later
liability information based on recent manual estimates as a
substitute for settled data.
(ii) Whatever method is employed as the base for projecting claim
costs and expenses, a view of past and future inflation has to
be taken. In particular, for liability claims, a view of the
rate of inflation to bring 1975 values up to 1980 is needed and
thereafter a future rate to project these costs into 1983 or
later.
(111)
4. THE GROUP
Finally, a view of future levels of claim frequency has to be
decided. Factors including future weather conditions, petrol
prices and speed restrictions have to be considered. It will be
shown later that these aspects may have a relatively large
effect on the profitability of results.
In a Company, it is desirable to make premium decisions within a Group.
The decision-maker is the person who ultimately says what the premlum
rates are going to be; he would usually be a General Manager or the Motor Underwriter.
The rest of the Group would act as advisers to the decision-maker,
supplying information on various aspects of the business. The size of
the Group might ranee from one person for each main function, to, in a
small co~pany, as few as two, the decision-maker and the underwriter.
-12-
The following are the main aspects of the business which have to be
considered:-
1. General Underwriting principles, which vary from company to company
and reflects, i) prior views on occupations which the marketing
should attempt to attract, e.g. teachers, civil servants; ii) wording
in the policy conditions to take into account, the introduption of a
new rating factor such as protected No Claim Discount and so on.
2. The overall market position of the Company compared to its
competitors, with regard to growth and pricing. Within the Company,
production summaries will be available showing lapse and new business
figures. Harketing will also be pushing for sales increases over
selected parts of the portfolio, by trying to keep any rate increases
.' to a minimum
3. Analyses from the claim negotiator, who will report the latest manual
estimates on present third party liability claims.
~. Statistical analyses will produce information for various members of
the Group; concerning in particular, past claims frequency, claims
cost and production results,
5. General economic factors which will be used by the Group to project
into the future, past claim costs and expenses. This will involve,
inter alia, a view as to the effect of the Government's curren~
economic policy on inflation rates, salaries and prices.
5. FORKULAE
It might have seemed ideal to relate the time period basis for
calculation of premium rates to a cohort of policies a~l effected during
a particular calendar year. However, this would require an
overr-;op!list.ic::ted d"tahase, whiC'h is not available. Hene.", a pragmatic
approach has been taken, namely, the averaging process outlined in
Section 3, which uses basically a calendar rate year.
It would be ll.~.?f")] at t.hi.s po:i.nt. to eive the premium fO!'Tohlla to be
appHed by corr.liination of riSK factol's where appropriate.
-13-
The risk prellliUII (RP), i.e. prea1_ excludins expensea.
RP .. to
where t .. clat. trequency .. number ot claims
exposure
o .. projected olaims cost
where AD .. average accident damage cost per claim it settled
1aIediately.
.'
and
,TPPD = average third party property damage cost per claim it settled
immediately •.
TPBI = average third party bodily injury cost per claim it settled
immediately.
M = average llliscellaneous cost per claim it settled immediately.
e,t,gth, are inflation rates tor respective types or claim cost.
n1,n2,n
3,n4 are average settlement periods tor respective
type~ ot claim cost.
++ An alternative suggested by Gunner Benktander is to replace (1+e)n1 by
Where the inflation rate is eCt, i.e. (e~ -1) p.a. and t is the average
time to settlement.
-14-
The calculation of the brealceven praaium (or office praail.ll) P, depends
on how expenses are introduced. There are two main variants:-
(a)
or
(b)
P=~
1-8
where S is the total expenses
total premium income
1.e. a f1xed percentage of premium
(nb + L) P = RP + (cc.f + --t- + r + ed)
1-w
wh.ere cc = claims cost expenses and f is claim frequency
nb = new business expense
L = lapse expense
t = time period till lapse
r = renewal expense
ed = endorsement expense
w = commission plus another expense related to premium
and the costs per unit are inflattd to 'the relevant date of pr~~ium.
This method entails variable plus fixed expenses.
I. short dis:!ussion about these fonnulae appears in Sect!.on 9, but
brie:ly, the total expected expense will be the Saffie for cach method and
only thc division of expense alters.
Two omissions will be noticed:-
(a) No contingency loading (or solvency loading)
It is not certain l:hether co:np!lnies explicitly take this factor into
account. t.lt'1ouCh it has been o:nitted, ::.lr.!!bralcall~' its Inclu:oioi1
wwld be very easy. The p;"obler.JS are of estimating a value and its
effect on the final rate.
-15-
(b) Investment Income
It is acknowledsed that the market does not explicitly take this
factor into account. ImpliCitly, income is taken into consideration
when the whole motor acoount is scrutinised, in that the ultimate
trading results can be caapared with the underwriting results. This
subject deserves a paper on its own and, hence, is left with the
simple comment that research should be carried out to investigate the
effect the factor haa on pralliums.
6. DATA (Sub-titled: Now the Story Begins)
It is now generally accepted that past clalll information must be matched
with the relevant underwriting factors at the t1llle of the cla1ll. In
addition, for IIIOtor insurance, the number of cars exposed to risk will
also be related to the claim period analysed.
There has been some work performed on the relevance of rating factors, in
particular traa the Belgian school, e.g. Hallin and Ingenbleek (1981),
but it is still in its early stages of development.
It will be assumed that all the existing market underwriting factors will
continue in use, since it is unlikely that any Underwriter would consider
altering theil, without other CClllpanies following suit. In order to
illustrate the principles under discussion, the rest of the paper will be
concerned with a worked example. The following rating factors will be
used:-
(1) Type of CO'ter - Caaprehensiw or Non-CClllprehenaive (2) Policyholder's Age - 17-20, 21-24, 25-29, 30-34, 35+ (3) Car Age - 0-3, 4-7, 8+ (II) Vehicle Group - A, B, C, D
Two ooaaents have to be made:-
(a) .sCllle sisnificant rating factors have been omitted from this list
(e.g. district, no claim discount and use of car), but the analysis
can easUy be extended to take th_ into account.
-16-
(b) Following discussion with the Underwriter, it may be possible to
a~gregate some of the detailed data into relevant groupings, in order
to make the analysis more manageable. In our example, as an
illustration, Policyholder Age 35+ was grouped instead of
sub-dividing into 35-50, 50-60 and 60+.
7. THE GENERAL PRINCIPLES OF CLAIMS ANALYSIS
The principal objective of the analysis is to project past claims data
for the relevant period. If the example in Section 3 is taken, then on
1st July 1980 a premium for the 1st October 1980 has to be decided. The
premium rates will be in force for one year and the average date of
claims arising will be 1st October 1981. Hence, all claim information
has to be projected to 1st October 1981 and onwards. There are two
levels of decision to make, namely,
(a) t~e overall levels of frequency and claims cost,
and thereafter
(b) the within-portfolio levels, i.e. the relationship between rating
facto~s.
The first stage is dealt with by the Group with very little statistical
analYSiS, but the ultimate decision requires a great deal of judgement.
The second stage is basically where the statistical modelling takes over
and actuarial judgement comes into play.
The following sections give a brief resume of the process for different
parts of the analysis. Details of statistical modelling were contained
in a talk given to Post Oualification Course for UK Actuaries on General
Linear Models (GLM) by Professor A.P. Dawid and Messrs. Baxter, Coutts
and Ross in 1980. Further details were given in a paper entitled
"General Linear Models in Insurance" at the 1980 Congress of Actuaries in
Switzerland. More recently in the ASTIN 1982 a paper by Albrecht
discussed all the literature on this type of analysis. Even more
accurate models may be achieved by ascertaining the claims amounts
distribution by Ziai's methods (1919) and inputting into the powerful GL~
algorithm.
-17-
7. 1 CLAIM FREOUEICY
(a) OVerall Levels Data for past years would be available by, say, s~b-groups
Comprehensive and Non-Comprehensive, in the.following format in
Table 11-
Table 1
Frequency per 1,000 Comprehensive
vehicles Quarter of Accident Year
Year of Accident 2 3 II
1977 142 125 115 152 138
1978 152 137 138 154 148
1979 1811 134 141 164 156
1980 159 134-
- includes an IBNR estimate
Discussion on projecting the 1980 results into 1981 would be centred
around such items aSI-
(1)
(i1)
(U1)
(iv)
weather conditions,
petrol prices,
road repairs,
general economic conditions, since these ~isht affect'the
frequency with which policyholders have their cars serviced.
In our example, it was decided to use the 1919 overall level (i.e.
156) as the projection for 1981.
(b) Within-Portfolio Analysis
All the rating factors mentioned above are considered, with separate
analyses for Comprehensive and Non-Comprehensive. The method of
smoothing applied was basically the Johnson and Hey (1971) model,
slightly adjusted by using logit of proportions as the dependent
variable (see Baxter et al (1980) for details) to obtain the smoothed
claim frequency for all combinations of rating factors. The smoothed
values were adjusted to give a prOjected overall frequency equal to
the 1979 values, such as those shown for Comprehensive in Table 1.
By breaking down the data according to all the rating factors, the
effect of portfolio changes is minimised.
-18-
7.2 PROPERTY DAMAGE AND MISCELLANEOUS COSTS
(a) Overall
For Accidental ~amage (or Fire or Theft for Non-Comprehensive), Third
Party Property Damage and Miscellaneous Costs, the object is to
obtain the average cost or a cla1m in October 1981 if settled
immediately. This value is then projected forward after taking into
account the average settlement date. In practice, inflation is
assumed to be the main projectin~ factor for overall levels. Hence,
for each respective claims cost, actual values are projected by
agreed specific inflation rates up to the projected settlement date.
(Inflation rates will be discussed in Section 10).
From detailed analyses, the average settlement periods used were 3 months, 6 months and 3 months respectively.
(b) Within-Portfolio Analysis
To date, statistical modelling using all relevant rating factors has
been seriously attempted only in accidental damage claims (see Baxter
et al (1980». A great deal more work i~ required to get reasonable
models to represent the data. Models similar to the Johnson-Rey
(1971) ~o not give reasonable results, since they do not take into
account the skew dist.ribution inherent in the accidental damage
claims (see Ziai (1979».
For Th.1rd Party Property Damage, separate averages were used based on
car grouping. For Miscellaneous, two overall values were usee,
Comprehensive and Non-Comprehensive. The reason for not modelling
these types of claims was simply lack of time.
8. BODILY INJUPY
This is by far the hardest part of the statistical analysis, since, on
average, only about 5J of all claims per year involve bodily injury
costs, yielding for the example portfolio about 1,000 claims. This
section deserves a paper to itself, so it is only possible to highlight
some of the main problems and present the prag~atic solutions used in
this example. However, before discussing details, it is necessary to
look at the arguments against pooling data from different companies to
arrive at an input to premium calculations.
-19-
Since the n~ber of claims for each company is, in practice, small, it
has been s~gested that all companies should pool their data. That would
not, however, solve the problem, for a reason best illustrated by way of
an example. Table 2 shows some typical average bodily injury costs per
claim for an individual company. They have been adjusted for earnings
inflation plus 'judgement drift' (as previously defined) to bring t~em
all up to the value of 1982, the projected date.
Table 2
Year or Average Bodily Injury Costs per Claim
Accident Inflated uE to 1982
£
1912 119
1913 80
1914 115
1915 58
1916 104
1911 90
1918 19
1979 61
If inflation were the only factor operatinR on these averages, relatively
constant values would be expected. P.owever, it is clear that there is
wide variation. This is due to the extreme skewness of the underlying
distribution of bodily injury claims; there is no reason to suppose that
the distribution shifts from year to year.
Corresponding pooled ~ata for all companies is available in the UK from
the British Insurance Association motor statistics but is not available
for publication. However, if it ,,'ere, then the sample size would be far
greater and the variability smaller than for any single company. Suppose
an average claim of £90 (in 1982 values) resulted from pooled data for
years 12 to 19. If the Company based rates on this, large profits would
be shown if the 1915 experience were repeated, or large losses if the
1972 figures occurred again. The real problem is that the Company
experiences a small sa~ple of the total market experience and therefore
its premium rates should make an allowance for its own variability.
-20-
(a) Overall
The Group has to decide the basis of projecting bodily injury
claims. Assumin~ that they have decided to use their company's data,
they have to ~elect the base period. The possibilities are:-
(i) Using the latest results (1978 and 1979), which, however,
entail predo~inantly manual estimates.
(ii) Groupin~ earlier years, which contain claims fully settled.
In our example, y~ars 72-76 and 72-77 were grouped and then inflated
(see Section 10) to the 1982 projections. The average settlement
period was established to be 2 years.
(b) Within-Portfolio Analysis
There are two basic decision areas:-
(i) The selection of rati~ factors is very difficult. In our
exa~ple, single factors based on Policyholder Age were
considered appropriate. This is a good example where GLM would
be very effective, as it could take the shape of the
distribution into account. There is another area of research
where the profession should lead. Some work by Chan~ and
Fairley (1980) shows that modelling is useful.
(ii) As the data are broken dcwn into even smaller groups, t~e effect
of large claims becomes very important, since a large claim in a
small cell will disproportionately affect the results. Hence
some method of smoothing has to be applied. In the example, all
claims over £10,000 were cut off at that value and the excess
was respread over the whole portfolio. This.method is open to
criticism, but it was felt appropriate at the time.
Another factor to be considered is whether the large claims are
in fact correlated with particular rating factors or randomly
spread.
-21-
This completes the statistical analysis of costs and it would be
useful to summarise the analysis so far.
Table 3 - Summarises Analyses
Whether Modelling
TVEe of Anal vsis "atin~ Factors Technig:ue used Overall Level
Claim Frequency All YES 1979
Accidental Damage All YES Inflation
Rroperty Damage Compo & Non-Comp. NO Inflation Car Group
~iscellaneous Compo & Non-Comp. NO Inflation
Bodily Injury Comp., Non-romp. NO 1972-76 Policyholder age 1972-77 )
+ Inflation)
9. EYPENSES
It is not proposed to go into detail about the collection of data to
obtain a breakdown of expenses. A paper by Ruston (1977) has described
the process very well. As far as the Group is concerned, it will cbtain
this information directly from the accounting department within the .
Ccmpany, Which projects future expenses. The only factor it would
consider is the future rate of inflation related to expenses (see Section
10) •
As far as the premium formula is concerned, two approaches are available.
The first is to consider all expenses as fixed in the short term - see
(a) in Section 5. The total value of expenses is obtained from the
accountants and then expressed as a percentage of premium income.
CommisSion and other premium related expenses are added to this
percentage. In practice, the value should be between 25J to 45J,
depending on type of ccmpany. The rationale behind this is that in the
short term (say 2 years) staff levels (which make up 80S of expenses) are
Virtually fixed. Thus, it is argued, further sophistication is needless.
The second method (b) is to divide up costs into those which are fixed
and expressed as a percentage of premium, and those which are
identifiable as separate totals such as claim cost expense, new business
expense, lapse expense and endorsement expenses.
-22-
For the example, the following were used and "then inflated to 1981
levels:-
Fixed as a percentage of premium (incl. commission)
Claim Cost
New Busi"ess
Lapse
Renewal
Endorsement
17J
£14 per claim
£6 per Pol~cy
£1.50 per Policy
£ 1.50 per Policy
£2.25 per Policy
Both methods will ~ive the same overall expense allocation, if the
portfolio is similar to that of the base period. However, if the
portfolio changes, the second method will give more variable results.
10. ECONO~IC ASS~TIONS
Tre Group will have to form some overall views on inflation. As
mentioned earlier, general economic factors and government policies will
dominate. Usually, several scenarios will be followed. Different rates
of inflation will be applied to different parts of the analysis. In this
respect, in the ur, the British Insurance Association Economic Advisory
Group is very helpful in reporting and projecting inflation separately
for each type of claim. Another problem is to establish which index will
be an appropriate indicator for each type of inflation risk.
The following is a summary of the rates and assumptions used by the Group
in this example.
Property Claims
Bodily Damage
Expenses
Table 4 - Inflation S~~ry
INFLATION INDEX
E8rni~s + ~Aterial C~ds
Farnings + Judgement Drift
Internal, based on Salary Increase
SCENARIOS
!!!2!i ~ 11J 7J
17J 13~
10J 10J
Since inflation has a major effect on the results, the decision maker
will have to dominate this Group decision.
-23-
11. 'DIE CALCDIATrON CF PREMIUM RATES
The stage has now been reaohed when the preJB1U11 oaloulation oan be
pertonaed. This was oarried out on aioroprooessor by applying both
formulae in Seotion 5 to the data generated by Seotions 7, 8, 9 and 10.
An exaaple of the final praailB is giwn in Appendix 1 ColUllll (4) (break
even preJB1U11) for all sets of ratins faotors. The underlyins assumptions
vere low inflation, bodily injury 1912-16 and expenses all r1xed.
In a later aeotion, the erfeots of different sets of assumptions on the
praai. rates are di80usaed.
12. PRESENTATI~ CF RESULTS
The traditional way of presentins rates was via a rate book, where every
oaabination of ratlns faotors was defined with its respeotive rates. In
the early 1970's, the Co-operative Insuranoe Caapany introduoed a points
system and several oaapao1es are nov usins similar systems. In the
follOWing seotions, we look at the problem of derlving a premium for a
points table. Firstly, we must define a points table. Then we will
examine why a Caapany would want to detenaine praaiums in this way. We
will then oonsider the mathematioal background and develop an algorithm
for deriving a points table. Finally, we will seek to interpret the
results. However, it is emphasised that the "points system" is only an
alternative to the rate book and is used as a selling point.
WHAT IS A POIl-1TS SYSTEM?
The explanation of a points systea 1s probably best explained by way of
an example. Consider the followins:-
Table 5 - Points Table
CCItP NCIf-caU' Cover 19 0
Policyholder's 17-20 21-2~ 25-29 30-34 35+ Age 30 12 1) 2 0
O-J II-I 8+ Car Age 8 ij 0
Vehicle A B C D Group 0 5 10 20
-24-
It can be 'seen that we have four rating factors and with each is
associated a scale. For each point on the scale, there is a value,
expressed in pOints. The differences in the points on each scale,
reduced to percentages, are called relativities.
The procedure used is to look up the points score for each rating factor
and then a~regate them. We then use a pOints conversion table to arrive
at the premium to be charged. The oonversion table is simply a list of
points scores with associated monetary values -'see Appendix 2. For
example, oonsider Comprehensive, Polioyholder Age 35+, Car Age 4-7 and
Vehiole Group A -
19 + 0 + 4 + 0 pts. = 23 pts. £12.35
13. WHY USE A POINTS SYSTEM?
The advantages of a pOints system are mainly praotical. They inolude,
not necessarily in order of importance:-
(a) It. is cheaper to produce than the rate book.
(b) It is easy to revise rates.
(c) The calculation is relatively straightforward.
(d) The method is easy to understand.
There are however problems associated with such a system. Examples ~re:-
(a) It may be regarded by outsiders as over-simplified.
(b) Fine tuning of the rate structure is no longer possible for marketjng
purposes.
(c) The pOil'ts valt'~z (a~ opposed to tr.e n;(1nctar:: valued ch&.n('.e only
rarely, and changes over time in the underlying variabl~s tend to be
ncslt'cted.
(rl) B(·c?u~c U'e po.i.pt.~ syster.> 11' eazy t.o jr>tc'rpj-"l, an:! ad.1ul't~·~nts to
the points whjch reflect statistical experience, arc th~ur.ht. by
oulsidere to be errors.
(e) The re~ult.inp: premium slrllcture is not exact, because the underlyjne
llroalysj s is ulUmatcly a simpU fication.
-25-
Returning to equation (1), take lop to the baH 1.0325, whence
log 1.0325 Pijkl = 9
We then fit this relationship by weighted least squares:-
.:;- A2 min,- Wijkl (log Pijkl - e)
"~,
" where denotes the least squares estimator
Wijkl Is some set of weights to be selected
A ",.. A Ci PHj C'k and VGl etc. are the estimated points for the
table.
We now require to know whether we can dispense with any of the terms in
(2), particularly the interaction terms (about which more will be said
later). We must then estimate values for the factors retained.
The method of Johnson and Hey (1971) can be used to arrive at a points
systell' rolloHing equation (2). To ascertain whether any of the terms can
be discard~d use General Linear Models. However, for simple assumptions,
the methods of Orthogonal \Oleighted Least Squares (OWLS) and of Johnson
and Fey and the General Line?r '.!odel (Baxter et al (1980» all give
similar 'ref;ults.
OELS l"eHef; on tl"e fact that. it is possible to factorif;e the ",eie.hts in
equaUon (1), that is
This is demonstrated in Appendix 3.
This ap~",'oxi:1'l'1tj on h~s one hj,p:hly pract.ic!'l advantage. It does not
require the invercion of matrices. Hence, it'is possible to analyse a
large nu~ber of factors without computational capacity becoming a
conct;r'nint, ~nd this method is c:c'opted here.
-26-
13.1 METHOD OF CALCULATION
We wish to convert the break even pre.ium into a points system. Then,
the problem is to compare our fitted points premium (est1lllated from our
~odel) with the break even pre.ium. That is
fOl' each combination of rating factors, where
~ P = fitted premium
P = break even premium
Proceed as follows. Let
• Pijkl = (1.0325)
Pijkl = premium associated with a given rating
where i = level of cover (C)
j = level of Policyholder's age (PH)
k = level of car age (CA)
1 = level of vehicle group (VG)
Then define
where M is a constant
In = all interactions between terms
factor
(2)
The choice of the figure 1.0325 is quite arbitrary. This is an example
of a multiplicative sy~tem.
This complicated analysis is used instead of, say, single tables, in
order to explain the relationships between the factors. The interested
reader i.s rE'ferred to Professor A.P. Dawid (Post Qualification Seminar
for Vi: 1:(Jtu"des in 19[.0).
-27-
13.2 AN ALGORITHM
An algorithm follows for the computation of a points table.
(1) The first step is to choose a set of suitable weights. A suitable
choice would be standing business, since, as will later be
demonstrated, the method is reasonably robust to the choice ot weights selected.
(2) A standard analysis of variance is performed, using
log Wijkl (log of standing business).
A" " A (3) T,h~-factors ci (ph)j (ca)k (vg)l are selected, after checking
for interactions from the Analysis of Variance.
A '" '" (4) The Wijkl is replaced by the product of the estimates of c, ph, ca,
~ from the previous step, i.e. we find 0ijkl from the expression
min 2. ~i (~) j (~)k (~)l (log p-8}2 -J"'l
(5) The Weighted Analysis of Variance of log P is then checked to ensure
that there are no significant interactions.
(6) Assuming that no si~nificant interactions appear, then the estimates
of the main effects serve as the points wi thin the table.
'(1) The fitted values for premiums computed through the model are
co~pared with the break even premiums, in order to assess the
goodness of tit.
(8) Then an overall adjustment is made to make sure that the sum of tl1e
breakeven prE'mium 1s equal to the sum of fitted premiums.
A!l ey'p",p1 n I-Jill now t-p prf'<:'?nted. Appendix 1 ShOHS the bnsic data. The
ratin,!!; fael co,'1! can toe seen 0:1 the left. (Column (1) sl10ws the Standard
BUSiness, which is being used for the weip.hts, i.e. Wijkl •
-28-
Analysis of Variance on log Wijkl 8ives the following results:-
Table 6 - ANOVA of loS W
Degrees of ~ Sum of ~uares Freedom Mean Sguare Ratio
Cover 1.5 1 258 Car Age 5.9 2 487 Vehicle Group 4.6 3 253 Policyholder's Age 33.0 4 1360 Cover x Car Age 12.4 2 1028 Cover x Vehicle Group 0.0 3 1 Cover x PH Age 3.9 4 161 Car Age x Vehicle Group 2.3 6 62 Car Ae;e x PH Age 0.2 8 3 Vehicle Group x PH Age 0.5 12 7 Residual 0.5 74 TOTAL 64.8 119
From this, it can be seen that, in addition to the main effects, there is
a significant association between cover and car age. This point will be
returned to later.
Th~ standing business distribution yields weights for the factors:-
Table 7 - "'eillht.s for Stl'nding Flu~j ne:o.s
COII!P.
11.8
Car Aile
0.3 1:1i
Non-Comp. ~8--
4.7 5.3
Vehicle Group
A B 2.3 6:1
Po1ic~'h(lld(!r lire
17-20 21_;>ij Q.9 --;-:g-
C D 5.5 2.ii
25-20 30-34 35+ 2:8 11:3- 32.3
Tt"., valu"s reprc:"'.ent tl'e cli;et.ri bution of standin!': bu~ine:;s wi thin the portfolio. For f)x~r.>plet there are 32 times as Jn~ny policie~ in Age 35+ a3 in Al~e 17 -20.
-29-
Using these estimates Wijkl , 0 can be estimated from equation (3).
The Analysis of Variance on log PiJkl yields:-
Table 8 - ANOVA log P
Degrees of Sum of Squares Freedom Hean Squares Hean Square Ratio
Cover Car Age Vehicle Gr6up Policyholder's ARe Cover x Car Age Cover x Vehicle Gp. Cover x PH Ap;e Car Age x Vehicle Gp. Car Age x PH Age Vehicle Group x PH Age Residual TOTAL
(1) 4 545 659 3 2511 634 4 0011 061 2 758 813
231 3111 811 683 110 913 5 869 5 576
21 566 "i 191
14 99 306
(2) 1 2 3 4 2 3 II 6 8
12 711
119
(3)=(1).(2) (II)=(3)/Residual(3 4 5115 659 7 788 1 627 317 2 788 1 334 687 2 287
689 703 1 102 115 670 198 28 228 118 10 228 18
978 2 697 1 797 3 584
It is important to note that this analysis does not allow us to make any
statement about the goodness of fit of the model, since we do not have
available an independent estimate of the residual variance. All that the
analysis shows is that, relatively, the main effects are far more
important than the interaction terms, which can therefore be neglected.
-30-
From these results, we estimate the main effects as fo11ows:-
~
Caup. Non-Caup. 6.5 -10.9
Car Ase
0.3 11.7 '1 .. 0 li:1
Vehicle Group
A B -9.0 -w
Po1iclho1der ABe
17-20 21-24 27.8 20.6
Table 9 - Estimates of Points
Weighted Grand Mean 132.9
8+ -7.8
C D 0.8 17.2
25-20 30-34 6:9 2.4
35+ -2.3
It ~ou1d be easier to make comparisons if there were no negative va1ups.
This is simply accomplished, by transforming the smallest value in each
1i~e to zero, making the same addition to the other values in the line,
and adjusting the overall mean subsequently.
For e1.:am;>10, with car cg~, we can alter the value for '8+' to zero by
adding 7.8, whence 18.8 for '0-3' and 11.9 for '4-7'. The overall mean
is 1ate~ reduced by 7.8, as in the following table.
-31-
Table 10 - Altering the Base to Facilitate Comparison
Unadjusted Adjusted
Overall 132.9 102.8
~
Caap. 6.5 ·17.11 Non-Comp. -10.9 0
Policlholder A!e
17-20 27.8 30.2 21-211 20.6 23.0 25-29 6.9 9.3 30-311 -2.11 0 35+ -2.3 .1
Car Ae;e
'0-3 11.0 18.8 11-7 11.1 11.9 8+ -7.8 0
Vehicle Group
A "9.0 0 B -11.3 11.7 C 0.8 9.8 D 17 .2 26.2
We now wish to examine the goodness of fit of the new premium structure.
There are three tests:-
(1) The fitted premiums can be compared with the break even premiums for
each rating factor (i.e. E/A).
(2) The premium income we would expect, taking account of our standing
business, from our fitted premiums can be compared' with the break
even premi ums •
(3) The expected premium income from the fit.ted premiums can also be
compared with that of the present premium structure.
-32- I 1
ColUllll (6) ot Appendix 1 shows the relationship between the tit ted
preJlliuJIIS and the break even prellliuJIIS, expressed as a percentage. It can
be seen that, generally, the divergence is only tour percentage points,
though in a tew oases, e.g. CCJllprehensive oover tor an eight-year-old
whicle in Group D driven by a teenager, it is very large. Note also
that tor CCJllprehensive whiole age 0-7, the ratio is less than 100,
141llst tor the oorresponding Non-CCJllprehensive values, it exceeds 100. This suggests that CCJllprehensi ve and Non-CCllprehensi ve require separate.
tables, as discussed below.
Column (7) ot Appendix 1 shows the ditterence between the tit ted premiums
and break even prelllium, weighted tor the standing business. Column 7
reveals that the actual ditterence by cell in required premium income is
in s,OIIie cases considerable. It would seem that Canprehensive is being
mdercharged and Non-Canprehensive overcharged. We will return to this.
However, overall the total prelllium income is reasonable, as shown by the
tollawing:-
(A) Total Standing Business 90,362
(B) Point~ Premiums (Fitted using main effects) £7,898,372
(C) B~eak Even Pre=i~s £8,268,308
(D) Difterence (B)-(e) -£369,936
(E) Ratio (B)/(e) (~) 96
Hence t:1e fitted fJre:nium would have to be increased by IIJ to match the
breakeven premium.
A s1t:!:i.lat' co",parison r..ay be made with our present premium structure, for
example:-
(A) Total Standing Business 90,362
(B) Pcir~tn Pl'''eiirl.ur~ (Fitted usine main effects) £g,012~1I51l
(C) Actual Premiums (charged in the existing £8,812,028 rate book)
(D) Differcnce (B)-(C) £200,1126
(E) r:~~' .L'7'J (C)I (i.;) (,. , ,to' 91
This jndicates that a prcmil~ increase of about 3J 1s requir~d to break
even.
-33-
13.3 PROBLEMS
There are a number of problems which such a method might encounter in
practice, such as:-
(1) What are the appropriate weights?
(2) \ihat happens if we later change. the assumptions, e.g. as to expenses
or the impact of inflation on claim costs?
(3) What if we discover interactions between the main effects at the
first stage?
To investigate the effect of changing the weights, an experiment was
und'ertaken. The weights used in the previous example - Column (1) of
Appendix 1 - form a series, starting 32, ~2, 103 ••• and ending ••• 151,
167, 858. This series was reversed, i.e. the weights used on the second
occasion began 858, 167, 151 ••• and ended 103, 62, 32.
The calculations were repeated with the following results:-
Table 11 - Comparison of Different Portfolios
Ori,.inal Revised
Overall 103 103
Cover
Compo 17 19 Non-Comp. 0 0
Policyho] der Aee
17-20 30 32 21-24 23 23 25-29 9 9 30-34 0 0 35+ 0 2
Car Ase
0-3 19 13 4-7 12 10 8+ 0 0
Vehicle GI'~
A 0 0 B 5 5 .r. 10 9 D 26 21
-34-
Although some differences are apparent, these are perhaps not so great as
mjght have been expected, given the drastic nature of the change. We
conclude, therefore, that the method is fairly robust with respect to
changes in weights.
111. COMPARISON OF DIFFERENT SETS OF ASSIJtofPTIONS
One of the great advantages of this method is the relative ease with
which calculations may be performed. Hence the.points system can help
examine the effect various input assumptions have on the premium
structure. We will investigate the following changes:-
(a) Inflation: We will have two scenarios - High, with inflation at 11%
for property damage and 17% for liability settlements, and Low, with
"inflation at 7% for property damage and 13% for liability settlements.
(b) \"Po will compare data for Third Party Bodily Injury (TPBI) claims
based on the period 1972-1976 against data based on the period
1972-1977.
(c) Expenses: We will compare the effect of treating all expenses as
fixed, against the effect of distinguishing bettleen fixed and
variable expenses.
-35-
The results are shown in the following table:-
Table 12 - SUBDarl of Results
AssumE!tions (1) (2) (3) (4) (5) (6) Inflation Low High Low Low Low Present TPBI 72-76 72-76 72-77 72-76 72-76 Points Expenses Fixed Fixed Fixed Fixed & Var. Fixed System Weil5!!ts Present Present Present Present Reversed
Overall 103 106 100 109 103 104
Cover
Coa!p. 17 16 19 15 19 18 Non-Comp. 0 0 0 0 0 0
Pol1c~holder Afie
17-20 30 30 33 27 32 29 21-24 23 23 23 20 23 23 25-29 9 9 10 8 9 8 30-34 0 0 0 0 0 0 35+ 0 0 1 0 2 0
Car A!e
0-3 19 19 19 16 13 7 4-7 12 12 12 10 10 4 8+ 0 0 0 0 0 0
Vehicle GrouE!
A O· 0 0 0 0 0 B 5 5 5 4 5 5 C 10 10 10 8 9 12 D 26 26 26 23 21 23
The effect of varying the inflation assumption can he seen by comparing
Columns (1) and (2). As we might expect, the overall level has risen,
but the relativities are virtually unchanged.
Changes in the base year assumptions for TPBI are reflected in Columns
(1) and (3). The lower overall level suggests that the 72-77 basis
yields a lower overall average. This time there has been a change in the
Poiicyholder Age relativities. The factors affected are Cover and
Policyholder Age, which again agrees with our intuitive expectations.
Altering the expense assulI'ptions from Fixed (Column (1» to Fixed plus
Variable (Column (4», gives results as expected, na~ely, an incrense in
overall level combined with a narrowing of the relativities.
-36-
Column (5) repeats the results obtained when the weights were reversed.
It can be seen that t~e overall level does not change but the
relativities do.
We now turn to the question of what is to be done if associations between
main effects are observed at the first stage. It will be recalled that
in our worked example, an association was noticed between cover and age
of car. One possible solution is to have separate rating structures for
the two different types of cover. The results are summarised below:-
Table 13 - Separate Tables for Cover
Overall
Policyholder AFe
17-20 21-24 25-29 30-34 35+
0-3 4-7 8+
Vehicle GrouD
A B C D
These results suggest that
Comprehensive
118
28 23 10 2 0
23 13 0
0 5
11 28
there may be valid
Non-Comprehensive
105
36 25 11 0 3
11 9 0
0 4 7
21
theoretical grounds fo:-
having separate premium structures. However, we must never forget that
these decisions are taken on pr"ctical as well as theoretical r;rounds.
It is highly probl'ble that despite these results, ~:e will opt fOI' a
single premium structure, becau~e of marketing considerations.
A further advantage flows fr~~ th~ ease of computation. We can take
another c0mpany's rates and perform an analysis upon them. In this way,
their rating structure can be compared with our own.
-37-
15. fo4ARKETING ASPECTS
The Group will now be in a position to discuss the premium recommendation.
Factors they will have to consider will include inter alia:-
(i) What other companies have done since the last meeting.
(ii) Whether the competitive position will allow all or some ot the
recommendatIons to be implemented. For example, in.Table 12,
compare the actuarial premium recommendations Column (1) with
the existing structure Column (6). The outstanding differences
occur when comparing the points tor Car Age. Hence the
actuarial recommendation might indicate an increase in rates for
new cars. However, this is unlikely to be acceptable because of
the market conditions, which give little weight to this rating
tactor.
(11i) Whether there are to be rate changes, their timing, and the
likely reaction of the market
(iv) Any special marketing campaign proposed, e.g. introducing new
factors, offering discounts for older drivers, or intro~ucing no
claim discount protected.
As a final thought, the market place for selling motor insurance is
changing, owing to the advent of direct telecommunication on televiSi.on.
The time will come (when the next generation of televisions are made)
whe" quotations will be available to the public via the television in
their own homes. Insurance companies and brokers will have to consider
the implications of this new dimension. This just highlights the dynamic
world of marketing.
The issues raised by marketing considerations are very important, and in
fact dominate any decision process. However, the subject is too complex
to discuss formally in this paper.
16. Al:ALYSIS OF SURPLUS
Premium rates are simply an attempt to forecast claim and expense
experience. For proper control an analysiS of surplus should be
rei.uliJl'ly pel'formed to Ifleasure ,,-here .the forecasts arc failing and the
effects on the prof! tability of business.
-38-
The following analysis of surplus for motor is based on a note published
by Andrew Grant in the UK General Insurance research bulletin (1981).·
Additional papers by Brennan (1968) and Taylor (1974) are also relevant.
In 1977, a projection based on data then available was made for 1978; the
main categories of data are listed below. The results actually
experienced differ from those projected because the assumption$ made in
the projection do not wholly agree with the actual experience. We need a
method of analysing these differences into their component contributions,
which can later be addp.d back to yield the differences between actual and
projected results. This problem is similar to an analysis of surplus and
the information gained will be useful in fine-tuning the assumptions in
the model, as the monetary effect of the differences between the
assumptions made in the model and the actual experience is discovered.
It Would also provide a useful format for analysing these differences for
management.
Companies use models of which the simplest use only premiums and loss
ratio~, while the more complex ones involve assumptions about exposure,
average premium rate, average claim cost, claim frequency and expenses.
In the model consi~ered, assumptions on the following factors are made
for each period and risk group:-
SB Standing Business (number of vehicles at the end of the year)
A\:F Average Written Premium
ACF Average Claim Frequency
Ace Average Claim Cost
VER Variable Expen~e Rates (this would include the co~is~ion rate)
FE Fixed Expensee
The Earned Premiull' EP and the Exposure EX can be easily calculated usinp:
the 1/8 method for ouarterly projections or the 1/24 method for
monthly projections. The Average Earned Premium AEP can be calcu16t:ed hy
dividilY th~ EP 1:Iy tt,c FY. (At p:-esent., tax and ir-vcstr.:E'nt inccmE' m'e
ignored) •
The result from the year's Undel'w!"iting Profit UP (i.e. excludlne-: any
adjtl:>tr.'(';1!; to Out :'.:' ,,:.,~j.ne Clairl.:) for ,prim" years) eN' be E'Xp)':::,:~e,l in the forlll
UP = Earned Premhim - Claims - Expenses (inclusive of commission)
-39."
The Underwriting Profit UP projected by the model is
UP = EX x AEP - EX x ACF x Ace - SB x VER x AWP - FE
We will denote by , the figures derived from the actual results. It is
possible to derive the AEP' fro. the Earned Premium EP' and the Exposure
EX'. The claillll can easily be expressed in the form EX' x ACFI X ACC' as
the Exposure EX' and the Claim Frequency ACF' are known. The
Underwriting Profit oan, therefore, be expressed as
UP' = EX' X AEP' - EX' X ACF' X Ace' - SB' X VER' X AWP' - FE'
The differences between the actual experience and that projected are the
result of differences in the exposure, claim frequency, level of
expenses, inflation rate, 'etc. We need a method of assessing the
numerical contributions from each of the factors. The formulae for a
possible analysis are given below.
Effects of
Exposure
(EX' - EX) x (AEP - ACF x Ace) - (58' - 58) x VER x AWP
Average Prell'ium
EX' x (AEP' - AEP) - SB' x VER X (A~~' - AWP)
Claims Frequency
- (EX' x (ACF' - ACF) x ACC)
Claims Inflation
- (EX' x ACF' x (ACC. - ACC»
Claims Cost
- (F.X' X ACF' x (ACC' - ACC·»
Expenses
- (FE' - FE + 58' X VER' X AWP' - 58' x VER x AWP')
(where ACC. is the forecast of the average claims cost using either the
known or the latest estimates of claims cost inflation rates).
The fonnulae above can easily be checked from the expression for
IJndcrwr~t.jnp: Profit. -40-
The effect of Claim Frequency and Claim Cost and Claims Inflation can be
expressed in several different ways. Two of these are given belowl-
Orisinal Alternative Claim Frequency - El' x (ACF' - ACF) xACC - EX' x (ACF' - ACF) x ACC'
Claim Cost - EX' x ACF' x (ACC' - ACC·) - EX' x ACF x (ACC' - ACC·) Claims Inflation - El' x ACF' x (ACC· - ACC) - EX' x ACF x (ACC· - ACC)
The Original is preferred and is used in in the analysiS because the effect
of the claim frequency is independent of the actual claim cost. The actual
claim cost wil be partly based on the Outstanding Claims estimated and will
change over time until the ultimate settlement. Any adjustment to these
estimates in the analysis will affect only the result attributed to Claim·
Cost ••.
Various other alternative breakdowns of tbe analysis exist and the choice
largely depends on the projection procedure, the data and the factors to be
hl@hll~ted. As an example, the effect of average claim cost can be divided
into its component parts, namely, Accidental Damage, Third Party.Bodily
Injury, Third Party Property Damage and Miscellaneous. Agaj~ the effect of
inflation can be highlighted and the model would be similar to that used for
making the all()\olance ror inflation in the original projection. If it is
required to highlight both factors - that is inflation and types of claim -
and adequa.te data a~e available, the year's claim experience using the year
of account bases can be expressed as
- EX' Y. (ACF1' x AD' + ACF2' x TPBI' + ACF3' x TPPD' + ACF4' x M')
- (ACF1 x AD + ACF2 x TPBI + ACF3 x TPPD + ACFII x H)
The analy~is would pive the following results:-
Claim Frequency
- EJ' x (AD x (ACF1' - ACF1) + TPBI x (ACF2' - ACF2) + TPPD
x (ACF3' - ACF3) + H x (ACFII' - ACFII»
Claims Inflation
- EJ' x (ACF1' x (AD. - AD) + ArF2' x (TPBI. - TPBI) + ACF3'
x (TPPDI - TPPj) + ACFlp X (W' - M»
-41-
Accidental Damage
- £1' x ACF1' x (AD' - AD-)
Third Party Bodily Injury
- EX' x ACF2' x (TPBI' - TPBI-)
Third Party Property Damage
- EX' x ACF3' x (TPPD' - TPPo-)
Miscellaneous
- EX' x ACF_' x (M' - ~)
In the analysis above, AD, TPBI, TPPD and M represent the average
estimated cost per claim, for the Accidental Damage, Third Party Bodily
InjUry, Third Party Propepty Damage and Miscellaneous types of claim and
their respective claim frequencies are denoted by ACF1, ACF2, ACF3 and
ACF_. The - denotes estimates using the updated estimates of the claims
inflation rates.
Further, since ACC' will vary from year to year until utllmate
settlement, it may be useful to extend the analysis to measure what
effect changes in ACC' have on the overall surplus. The extension is
straightforwad but is not carried out.
Anal vais' of Surplus - Example
The following example is taken from the forecasts of part of the Private
Car portfolio of a medium sized UK insurance company. The forecasts are
made on a year of accident basis, this example being 1978. We give four
sets of forecasts for 1918, corresponding to a prOjection at the end of
each quarter in 1977.
-42-
The actual results for claim costs included the latest case estimates and
therefore this analysis would be subject to some adjustments as the claim
p8)'111ents run off.
1st 2nd 3rd 4th
(Figures in 'OOOs) Forecast Forecast Forecast Forecast Results
Standing Business 101.4 88.'4 87.4 85.4 78.3
Exposure 97.5 87.8 87.8 85.8 80.8
Written Premium 6244. 5702. 6114. 6339. 5794.
Earned Premium 6061. 5684. 5902. 6035. 5600.
No. of Claims 13.1 12.6 12.4 12.2 11.7
Claims Costs 3993. 3938. 3821. 3712. 3348.
Expen~es + 2077. 1987. 2184. 2295. 2342.
Profit (Loss) • (9) (241) (103) 28 (90)
• Profit : Earned Premium - Claims Cost - Expenses + including Variable Expenses (15J of Written Premium)
To calculate the effects for the analysis of surplus we need, in
addition, to reculculate the forecast~ of claims cost using the latest
estimates of claims inflation. These averages are given below:-
1st 2nd 3rd 4th
Forecast Forecast Forecast Forecast Results
Fo~ecasted Avera~e
Adjusted Forecast
304.81
289.85
The effect of each factor is then:
312.54
300.13
308.15
304.13
304.26
302.84
286.15
(Figures in 'OOOs) 1st Forecast 2nd Forecast 3rd Forecast 4th Forecast
Exposure -141 -41 -71 -56 Premiums 431 258 121 -81 Claim Frequency -257 -33 -89 -64 r.lai~s Jnrlatjon 175 145 47 17 ('laj",:". ro . .,t s 43 1611 210 195 Expen5Cs -333 -341 -206 -129
-43-
The initial underestimation of the average premium is partly due to the
premium increases during 1977 and 1978 which were not allowed for in the
earlier forecasts. It would appear to be relatively straightforward to
extend the ideas presented in this paper to separate.off the effects of
premium increases and therefore examine more closely the actual method of
projecting premiums.
The effect of claims cost unexpectedly increases between the first and
last forecasts, highlighting either a need to examine the methods used to
project claims cost, or possibly an unexpected change in the claims
experience.
The relatively large effect of expenses is due to two problems. Firstly,'
there~is a slight difference between forecasts and results in the basis
for·' alloca ting fixed expenses between classes; also, the inflation ra tes
used to project expenses during 1978 coul~ be out of line with those
actually experienced. An obvious extension of these methods would be to
separate the effect of inflation from effect of expenses in a similar
manner to that employed for claims costs.
CONCLUSION
This paper has argued for a detailed breakdown of motor insurance data. This
allows for an indepth analysiS of claim experience which in turn require~
statistical modelling, which assists in understanding the underlying
structure. In addition, it is suggested that premium rates can be structured
to take into account explicitly the underlying assumptions involved in
rating. This will help actuaries concerned with motor rating to appreCiate
the problems and advise the Underwriter from a position of knowledge.
-44-
A(X~WLEIXlEHmTS :
I would like to giYe credit to an Insurance CClllpany. tor allolling the data to
be produced. Thank Michael Brockman tor his etforts in getting the cClllputer
to perform the various oalculations and Russell Devitt and Miohael Weitzman
tor their help in preparing the paper.
In addition, I would like to thank nUJDerous colleagues tor tbeir instructive
ori ticiSlll, which have helped generate IIOst of the ideas in this paper, in
partioular Grahaa Ross 3nd Graham Masters. Finally, a word ot congratulations
to Jan Stow tor reading my writing and typing the paper aoourately and quickly.
-45-
Albrecht, P. (1982)
Baxter, L.A./Coutts, S.M.
and Ross, G.A.P. (1980)
REFERENCE
Parametric Multiple Regression Risk Models
I Theory and Statistical Analysis 16th
ASTIN Colloquium Liege
Applications ot Linear Models in Motor
Insurance 21st ICA Vol.2
'Test Hypotheses by Statistical
Investigations'
Brennan, A.R. (1968) A note on the Theory ot Analysis ot Surplus
TlAANZ p. 1311
Chang, L. and Fairley, W. (1980) An estimation Model for Multivariate
Insurance Rate Classifications : Journal of
Risk and Insurance Vol.1I7
David, A.P. (1980)
Grant, A.(1981)
Hallin, M. and In~enbleek, J.F.
( 1981)
Statistical Models tor Smoothing and
Predictions. (Talk given at a Post
Qualification Seminar on the Application of
Linear Models to the Analysis ot Motor"
Insurance Data).
Available from Statistical Department
University College London, UK.
Analysis of Surplus in Motor Insurance GIRO
Bulletin December
Etude Statistique des facteurs influencant
Ie risque automobile
15th ASTIN Colloquium Loen
Johnson", P.D. and Hey, G.B. (1971) Statistical' Studies in Motor Insurance JIA
Vol.97 p.199
Kat'ane, Y. and Levy, H. (1775) Regulation in the Insurance Industry
Determination of Premiums in Automobile
Insurance. Journal of Risk and Insurance
Vol. XLII No.1
-46-
Netherlands Group Report (1982)
(Chairman Wit, G.W.)
Pitkanen, P. (197~)
Ruston, I.L. (1977)
Taylor, G.C. (197~)
Ziai, Y. (1979)
New Motor Rating Structure in the
Netherlands
(Publication of the ASTIN-GROEP Nederland)
Tariff Theory - ASTIN. Colloquium in Turku
Expenses in General Insurance. OK Actuarial
Student Society Paper
Symmetry Between Components of Analysis of
Surplus TIAANZ p.37~
Statistical Models of Claim Amount
Distributions in General Insurance
Phd Thesis at The City University, London
-47-
-----------~..;..:..:.-.:.-.:..:..:.
PRIV;'TE CAR ?REM ANALYSIS TPBI 72-76. MAIN EFFECTS Fitted C;OYk::R V5H. ADE vtH. ORP P/H AGi stAND1N() NHNT9 I PR[MIUM
BREAK [VEil om. RAllO SB • BUSINE~S PR[~IUM OIFF
11~ (2) (3) (4) (5) (6) '(7)
(3) - (4) (3)-(4) (2) • (5)
COI1r 0-3 01 17-20 :32. 16" 224.~ 2:$7.77 ··1~.44 94 04'14.13 21-24 62. 162 173.34 101:S.06 ,017.:S2 '11 -10"J(..25 25-29 103. 140 114.06 1~7.(.O 0'12.74 90 -1312.26 30-34 152. 1;59 0:;.34 90.41 -'1;.1; 07 \ 137 -1906.77 3~+ 1446. 1:$9 E;~.C:3 ?3.US o-O.~ 91 -11~0.70
02 17-20 52. 174 260.64 274.10 "13.46 9$ 06".SO 21-24 146. IG., 207.21 20!6.13 -10."2 92 -2"1(.2.21 25-29 .207. 153 t.~3.4~ 140.51 -15.06 90 -4321.96 30-34 472. 144 9~.16 11:>.17 -16.02 06 -7:;39.46 35+ ~63. 144 9?73 lOS. 17 ·'S.44 , 92 ·42741.0S
03+04 17-20 60. 179 306.72 31S.06 -11.34 96 '680.34 21-24 180. 1"12 243.S4 2<.3.'10 "20.14 92 -37~.S6 25-29 .44. 150 157.04 176.64 -19.60 S9 -El702.S2 30-34 820. 149 116.69 139.32 -22.6.3 04 -113:55'1.'12 35+ S142. 149 117.36 130.40 -13.12 90 -106860.34
~+ 17-20 21. 195 ~17.16 478.~7 38.59 0108 S10.32 21-24 15.3. las 411.13 409.69 1.44 100 220.23
I 25-29 291. 174 264.79 2El7.65 -22.El7 '12 -6654.55 ~ • 30-34 630. 165 196.74 234.lt -31 • .36 84 -23~3a.3~
CIII I 35+ 4498. 165 197.S7 21S.70 -20.El3 90 -93672.76 ~
4-7 01 17-20 35. 162 179.6~ 100.96 -"1.31 99 -45.9S ~ .... 21-24 93. 155 142.02 147.:.so '-4.413 97 "416.96 >< 2~-29 126. I'll 91.90 92.91 -.9:5 99 ·117.00 ... 30-34 i' 101. 132 G3.34 70.2'1 ·'1.?4 97 '-351.34 35+ : 2OS1. 132 . 60.7. 67.01 1.72 103 358S.71
02 17-20 94. 167 200.73 ::Zll.~ "2.60' 99 -244.19 21-24 229. 160 1(.:;."4 171.14 "5.20 97 -11?l.07 25-29 30S. 146 106.07 100.55 "1.60 90 0652.15 30-34 744. 0 137 79.41 02.42 -3.02 96 -2244.03 35+ 7911. 137 79.86 70.2.1 1.59 102 12571.22
03+04 17-20 65. 172 245.63 244.76 .07 100 154.69 21-24 199. 165 101:>.27 201.16 "5.09 97 -1171.37 25-29 437. 1~1 125.76 130.96 -5.20 96 -2270.47 30-34 001. 142 93.4:> 101.41 -7.9<' 92 -6376.06 35+ 6077. 14~ 93.90 94.?O -1.00 99 -68S1.97
05. o 17-20 21. lOS 414/15 375.40 30.60 llO S12.18 21-24 113. 1131 329.2:> '320.20 9.05 103 1022.22 25-29 215. 167 212.05 219.23 -7.18 97 -1544.60 30-34 427. 1:59 157.56 176.52 -18.96 89 -009'5.25 35+ 2S18. l~EI l~a.46 165.21 -6.74 96 -19005.80 ,
\
S+ 01 17-20 17. 150 122.89 114.91 7.90 107 135.67 21-24 39. 143 9".70 95.90 1.79 102 69.92
---- --------
F
PRIVATE C~~ ~EM ANA~Y~IS rrOl 72,:76
FITTED COli:.;1 VEH. f,OE VCH. en? P/H AOE: ZTAtll'l;.o) rolt"T~ rm:MlIJM lInrAK tvt:N Dtrr nAno SD •
nu3INE:;~ r-m:l1lL1M Dl1'"F (1) (21 C~) (41 (:)1 (6) (7)
(3)-(4) (;$11(4) (21·(S)
COl\? s+ f. 25-~9 76. 130 62.92 :;0..92 4 •. 00 i07 303.37 30-34 170. 120 46.7:5 43.27 ~.4:3 100 591.62 3:;+ 2<)26. 120 0.02 42.20 4.32 111 13613.49
B . 17-20 43. l~S 142.79 142.39 .40 100 17.16 . 21-2'- 103. 14~ Uo3.:a I1S.2;$ -1.71 99 -lC~.21
25-29 219. 134 73.11 71.05 2.()6 . 103 4~0.16 30-3~ 497. 1:25 54.32 51.0.s 2.49 1015 1212.97 35+ 6792. 125 54.63 50.12 4.52 109 31573.06
C .~ 17-20 27. 160 163.03 164.31 3.72 102 100.31 21-~4 60. 1:;3 . 133.~3 1;:4.76 -1.17 . 99 -'O.4~ . ;:S-~7 1:!·3. 139 ~'S.03 96.151 -.47 99 -74.95 30-34 310. 1:S0 "_~.92 65.29 '1.37 9a -4203.41 35+ 3346. 1:::0 ~4.2? 62.34 1." 103 6529.111
I D 17-20 3. 177 ~3.51 144.26 139.05 . 196 '. tl7.1' .. 21-24 19. 16';> 22''';.23,' 207.159 1'1.64 100 33:;.17 '" , ::'5-29 61. 1:;6 145.06 1:;1.06 "6.00 • 96 -366.20 30-34 126. 146 107. '10 120.'13 -13.15 09 -165'.3~ 35" 966. 147 1 ()c·.40 114.90 -6.S0 94 -6357.6'
1"0)1 CO!1? 0-3 A- 17-20 6. 152 120.52 110.39 10.14 116 10S.Sa 21-24 6. 145 10;:.17 7J.6:5 ". 23.53 13'1 1;'1.1S :!~-2'9 6. 131 ~·5.(;0 20.63 37.17 230 223.0' 3')-34 12. 1'''~ 4;).C9 23.46 20.43 172 24S.1~ ,,-35+ 52. 122 4'}.17 44.9" 4.10 l09 217.~'
B ;' 17-20 . IS. 1~7 .. 149.33 132.62 16.71 113 250.6' 21-24 29. 14'';> 110. '11 100.67 10.04 110 523.2, 25-2~ 30. 136 76.46 62.24 14.21 123 426.41 30-34 23: 126 :;6.01 41.613 15.13 136 347.9~ 35+ lSI. 126 57.14 52.22 4.92 109 899.69
C· ." 17-20 17. 162 17:;.73 1:;1.57 24.16 116 410.67 21-24 34. 154 ' 137.70 111.07 27.03 125 946.14 '~-27 'Z7. 141 89.97 71.52 10.46 126 632.92 30-~'l '12. 1.51 66.C:; 50.27 16.53 133 696.::;' 35+ 291. 132 67.24 57.95 9.29 116 2702.0$
D',··, 17-20 12. 17g. 2-:>6.2? 214.S6 81.43 139 977.12 21-24 :25. 171 \ 23'5.~5 170.06 65.49 139 163-'.:!~ 2~'-29 23. 1:;7 1~1. 70 lOS. Sf.> '42.(;4 139 9S'5.37 30-34 2::!. 140 112.72 CO.36 32.36 140 711.9, :)5+ 121. 14~ 113.37 91.7? 21.50 . 124 2611.0~
':-7 I~ •• ' n-~'" $3. .If,:: 10::'.92 11?51 ~·16. 5? Sf.> -14'10.0, :' t-:~··. , -" .. •..... c:. t~:! CJ .. 7 .. t \, ... ~.11 9'1 -164.5Q , .. " ,~ ... "'1 ." -71,.:'4
"-PRIVAT;: CAR fRIZM ANALY~l$ TrCl r:t.-76
FITTED COV:.R VEH. AOE: VEH. GRP P/H AGE >;TANDINO POINTS F'REI1IUM CH<:AK EVEN DIFF RATIO SO •
DUSI NE:::S rR011UI1 DIFF (1) ( '~l .(31 (41 (SI H.I n'l
(31-(41 (;$'''41 (21.(~1
NON COI'J' 4-7 .1. 3~+ 482. 11~ '39.39' 42..2!; -2.87 93 -13e~.02 . \
'B ' 17-20 147. 1~0 119.59 1~:I.28 -1~.6? 83 -2307.08 21-~4 194. 142 9:5.07 94.90 .17 I 100 33.fJ8 25-29 211. 129 61.23 61.:12 -.29 100 -60.ea 30-34- 293. 119 4:1.49 'IJ.44 ~.O:l 105 601.12 35+ 1723. 120 45.76 47.20 ·'1.~2 97 -2621.39
C ... iO. 17-20 140. 1:1~ 140.73 147.00 -6.20 96. -878.93 21-24 220. 147', , 111.00 '103.?6 7.91 100 17"0.~1 2~-29 2:13. . 134 72.05 67.9~ 4.11 106 1038.70 30-34 300. 124 ' 5:1.54 ' 40.20 5. a4 ' 111 1(,01.84 , . 3~+ 163&. 12~ ~~.o .. ~2.10 1.6& 103 271S.n
",
D .... 17-20 6.'. 171 237.20 21~.99 23.29 111 . 1560.2S I 21-24 122. 164 100.(.a 157.69 30.'14 120 :577/).7' ....
2~-29 130. 1:10 121.4'1 106.12 15.36 114 • 1997.20 0 I 30-34 174. 141 90.2/,' 73.23 12.04 115 2014.84
35+ 69:1. 141 90.79 02.41 8.37 ,10 ~820.SP .'
s+ A. 17-20 346. 1:s3 70'.41 ", 93.87 -23.46 ' ri -8117.91 21-24 303. 126 5'5.'17 t..4.69 "0.72 87 -2642.16 25-29 392. '112 36.0~ 41.81 -~.76 86 -22'6.7~ 30-34 .": l504. .. 103 " 26.7'1 29.02 -2.2S· 92' -112S.0~ 3~+ 3217. 10~ 26.94 31.23 -4.29 86 -13815.31
B. 17-20 ~28. 133 81.80 104.49 -22.69 78 -11979.27 21-24 609. 131 65.03 72.03 .. '1.0~ 90 -4211.87' 25-29 774. ' 117 , u.sa 4S.9? -4.11 91 -3178. Ie) • 3"-34 90'1. 107 . 31.12 31.74 -.62 9G -613.66 35+ 6036. 108 ' 31.30 3:S.0~ -3.75 ' 87 -22608.4,
,C' ,_ 17-20 209. 143 ' .... 27 113.~ -17.~8 a -3674.', 21-24 204. U6 76.~;S 79.14 -2.61 97 -742.11 2~-29 346. 122 4?29 lSO.75 -1.46 97 -lSOS.7' 30-34 558. 113 36.62 35.~7 1.06 103 :5G7.0? 35+ 2035. 113 36.&3 30.43 -1.:'9 96 -4~17.9' '
11-20 ~6. 1~9' 1~2.31 .
166.29 -3.97 98 -222.4l1 '0' , 21-24 136. l!i2 129.04 120.32 0.72 ' 107 UO:S.91 2~-2.? 1~1. p~ 03.11' 79.74 3.37 . 104 l508.1' 30-3" 167. 129 61.7.5 58.30 3.37 106 1562.27 3~+ ~8. 129 62.10 62.24 -.13 100 . -'U3.96
.' TOTALS
90362. I iA" STA~DING 9US:NCSS (:)) POH.:iS F'R!ZMIUMS 7890372.00 • , (el DFl!;,;AK EVE~: PIlEI11UMS 826~307.50 " (01 DIFFER:::NC::: iBI-IC) -36990$5.5<>
Appendix 2
-51-
APPENDIX 3
Orthogonal Weighted Least Squares (OWLS)
It has been suggested that weights, which are the recriprocals of the
estimated binomial variance ViJk, where
Hence, with an interaction term included the expression to be minimised would
be:-
Choice of Weights
The minimisation can be simplied by assuming V1Jk factorises such thatl-
(2A)
where a1 is the' age effect di is the district effect and ck is the car effect. By substituting equation (2A) in expression (lA) it can be shown that
for a given set of constraints the model is 'orthogonal'. Hence, the
estimates of the parameters and an ANOVA on accident proportions are
relatively straightforward to compute.
The necessary constraints and estimators are as follows:-
[4\.' '(. : f. :olj\(j~ 2clC~'" ':. ~
. .. .. ~
\
L4;'~~~ L .... (~\()~ = 0 .a.h-;
. .l J
-52-
)
Estimators:
and similar
and similar
where
(I 2- 2 2 ~ . .l~ ~ ( ~~) f'""
~
• • ~ Q.. ;.. C.
"-L~
,J..' el( (~) 'ls"c:.
J ~. v. ~. 111':''''
.... equations for \(. and ""CI(
;)
"-
(Y\{l.j .. equations for
~q., • A. ...
" 2. £!l <. C!L><) - t-e. 1'\:,,, '" ~ ~)~ ... and l\(~) ,v.
}
/\~) ... ... t"-
... y. ..
z Co", :. c.. ...
" \(. ~
When we apply an orthogonal weighted least squares analysis (OWLS)
To justify the assumption of factorisation in equation (2A) the following
model is considered:-
Where H is the overall mean, A is the age effect,'Dj is the district, Ck is the car erfect, etc. Then I:. standal'd ANOVA on -10g10 Vijk is carried
out to see which effects are significant, by examining the mean squares. If
the mean squares of the interaction effects are small compared to the mean
squarcs of the main effects, then the interaction effect~ can be omitted
we can write " " " Ai " to" c.-" " ,J
A.' d.. ID ..l. t:< 10 ~.\ c..' I- lO ~ ~ J
/' ,.. :stimates of ai' OJ and ck can then be calculated, where ai' dj and
ck are only defined wil-ilin an arl>itrary multiplicative constant which for convenience \1111 be taker. as 10M!3
-53-
and