1995 General

Insurance Convention

The Bootstrap Method

andsome

Reserving applications

‘Good
simple are
our
most
ideas,....,precious
intellectual
commodity,
so
there
is
noneed
to
apologize
for
the
easy
mathematilcev
ael.”
BradlEefr
yom

Presented
atGISG
I995
by Christian
Larsen

223
TheBootstrap
methodwasintroduced
by Bradley
Efron,Stanford
University
in 1979.
Efron(1982)
begins
withsomegeneral
statements,
suchas:

“Goodsimple
ideas,....,
areourmostprecious
intellectual
commodity
sothere
isno
needtoapologize
fortheeasymathematical
level.”

and

'‘Animportant
themeof what follows,
isthesubstitutionofcomputationalpowerfor
theoretical
analysis.
Thepayoff offcourse,isthe freedom fromthe constraints
of
traditional
parametric
theory,
withitsoverrelianceon a small
setofstandard models
forwhichtheoretical
solutions
areavailable.
Inthelongrun, ....(bootstrap)
... ,
should
makeclearerthevirtues
ofparametric
theory...."

and

"Froma traditional
point
ofview,....themethods...
areprodigious
computational
spendthrifs.We blithely
askthereader toconsider
techniques
whichrequire the
usualstatistical
calculations
tobemultiplieda thousand
timesover.
None of this
wouldhavebeenfeasibletwenty--five
yearsago,before
theeraofcheapand fast
computation.”

The situation

We consider
a randomsample(Xl,...,XN)ofrandomsize
N.ThedistributionG of
N isassumedtobeknownandtherandomvariablesareassumed
to be independent
andequally
distributed
withunknown distribution
F:

and

Itisfurther
assumed
that
Xr isindependent
ofthefrequency
N.

X=(X1,...,XN)
denotes
therandomsample
andx=(x1,.....,xn)
theobserved
realisation.

The problem

LetR(X)bea function
ofX.ThenR(X)isa stochastic
variable
witha distribution
that
isdependent
on G andoftheunknownF.Theproblem
istoestimatethe
distribution
ofR on thebasis
oftheobservation
x.

224
A solution
-theBootstrap
Method

Ifthe
distribution
ofF wasknownthen
the
distribution
ofR could,
intheory,
be
calculated
exactly
orinpractice
beapproximated
withunlimited
accuracy
bythe
MonteCarlo
method.
Thesimple
idea
inBootstrap
istosubstitute
thedistributiof
n
F with
the
empirical
distribution
based
ontheobservatiox:
n

A. Construct
the
sample
probability
distribution
E (i.e.
the
empirical
distribution
ofX),
putting
mass
l/n
ateach
point
x,...,xn.
B. Consider
the
random
sample
c. Approximate
the
distribution
ofR(X)
bythe
(Bootstrap)
distribution
ofR(Y)
Theessential
facts
are
that
the
distribution
ofR(Y)
isonly
dependent
onE and
G and
ifE is
a good
approximation
toFthen
one
canexpect
that
thedistribution
ofR(Y) willbea
good
approximation
tothe
distribution
ofR(X)
and
thedistribution
of(E,G)
is
known
and
therefore,
the
distribution
ofR(Y)
can
becalculated.

Application
I:Calculation
ofthe
uncertainty
ofreserving
estimates
Let
N denote
the
number
ofobserved
claims
inaspecific
period
and assume thatNis
Poisson
distributed
with
aknownmean.Let denotethe information
linked
tothe
claims
including
information
aboutaccident
period,
development period
and
payments.
Only
information
regarding
the
past
is
available
and forsome claims
Xr,
all
payments
andperiods
ofpayments
are
notnecessarily
included.
Let
RESdenote
the
stochastic
total
ofthe
future
payments,
i.e.
thestochastic
reserve.
Letl(X)denoteareserve
estimator,
e.g.
the
Chain
Ladder
estimator
based onthe
claim information.
isafunction
ofthe
claims
X and
therefore
astochastic
variable.
Theproblem
wewish
tosolve
is
toestimate
the
distribution
oftheChainLadder
estimator
RI(X)
rather
than
just
theChain
Ladder
point
estimate
RI(X).

225
Example 1

Consider the accident periods 198.5-1994and assume that the number of claims with
accident date in period i=l985,..., 1994 and notification delay j=0,...,9 measured by
years from accident period to notification period is Poisson with the parameters
outlined below in figure 1.

Assume further that each claim has only one payment which is Gamma distributed
with mean dependent on the accident period i and constant coefficient
of variation equal to 2. The waiting time k =0,...,9 from year of notification to year of
payment is assumed to be independent of the waiting time to notification and with
distribution outlined in figure 1 below. The distribution of the waiting time r from
accident year to year of payment, i.e. the convolution of p and q, is also calculated.

Period Delay j~p k~q j+k~r

1985 500 1.000 0 0.30 0.35 0.1050
1986 500 1.050 1 0.25 0.20 0.1475
1987 500 1.103 2 0.20 0.20 0.1800
1988 500 1.158 3 0.15 0.15 0.1875
1989 500 1.216 4 0.07 0.10 0.1620
1990 500 1.276 5 0.03 0 0.1095
1991 500 1.340 6 0 0 0.0625
1992 500 1.407 7 0 0 0.0315
1993 500 1.477 8 0 0 0.0115
1994 500 1.551 9 0 0 0.0030
All 5000 12.578 Mean E(j)=l.53 E(k)=1.45 E(j+k)=2.98
figure 1

On the basis of these distribution assumptions we get E(RES) = 2142 and the
expected total claim amount is 500* 12.578 = 6289. As a consequence, the expected
amount paid already is therefore 4147.

Two data sets, PAST and FUTURE generated by simulation in The SAS System
on the basis of the distribution assumptions above. The PAST file consists of 4699
claims where i+j< 1995 and of 3486 payments where i+j+k< 1995. The Future file
consists of the remaining data, i.e. claims with date of notification or date of payment
in 1995 or later. The sum of future payments, i.e. the observed value of RES, is 2004,
compared to the expected value 2142.

226
As an example, the claim information regarding three claims from the PAST file is
shown below:

Observation Claim Accident Transaction Delay jor Payment

identification periodi j+k
108 108 1985 Notification 0
108 1985 Payment 3 2251.89
3086 3086 1991 Notification 1
3086 1991 Payment 2 1466.06
4649 4649 1994 Notification 0
figure2

The PAST data is now triangulated and the Chain Ladder estimate for the
reserve is calculated. The results are outlined in figure 3 below.

DevelopmentPeriod Observed Total Total

ACC. I 2 3 4 5 6 7 8 9 10 Payments Reserve Estimated
PE D Payment
1985 50 81 113 97 87 55 36 4 2 1 525 0 525
1986 59 64 103 85 95 5 43 11 5 1 516 I 516
1987 52 85 132 82 78 54 38 I2 4 1 534 4 539
1988 77 66 94 79 127 44 27 9 3 1 515 14 528
1989 81 111 123 115 87 84 45 12 4 1 602 62 665
1990 79 116 84 96 118 64 42 11 11 4 494 122 616
1991 79 67 112 132 109 65 42 11 4 1 389 231 621
1992 74 121 138 125 128 76 50 13 5 1 333 397 730
1993 78 I23 143 129 132 78 51 13 5 I 200 551 752
1994 112 148 185 67 171 101 67 17 6 2 112 864 976
ALL 740 983 1226 1107 1132 672 442 113 42 IO 4220 2247 6467
figure3

The estimated run off and the model run off are shown in figure 4. For example it is
seen that on average 11.4% of the claim amount is estimated to be paid in the
accident year, and that the model proportion r is 10.5%.

Run-offpattern
Est. 11.4% 15.2% 190% 17.1% 17.5% 10.4% 6.8% 1.7% 06% 0.2%1 65.3% 347%1 100%
Model 10.5% 14.8% 180% 18.8% 16.2% 110% 6.3% 3.2% 12% 0.3%1 65.9% 34.1% 100%
figure4

227
Figure
5 shows
theobserved
andtheestimatedproportion
oftheestimated
ultimate
claim
costspaid
atdifferent
times.
Itisseenthatattheendof31December 1994the
amount4220(65.3%)
waspaid
andthat theamount expected
tobepaid
inthefuture,
i.e.
theoutstanding
claims
reserve
is 2247 (34.7%)
Development
ofAccumulated
Payments/ Reserves

figure 5

Inorder
tocalculate
the
estimation
uncertainty
oftheChain
Ladderestimate
2247,
the
Bootstrap
methodhasbeenapplied
using
Larsen
& Partners’
Actuarial
Claims
Reserving
System.
Theestimated
distribution
ofR(Y)(figure
6)isbased
on200
MonteCarlo
simulations
fromthe
empirical
distribution.
Theprocedure
is:

1. Drawarandom number from

m G = Poisson(4699).
2. Drawarandom sample, Y1,....,YM froE
mE
3. Calculate the Chain Ladder reserve based on Y1,....,YM
4. Repeat
l-3 200 times.

226
 Out.standing
ClaimsReserve
Distribution

figure 6

Normally
we donotknowthedistribution
ofF andatest ofthequality ofthemethod
isdifficult
todefine.
However,inthesituation
above, where thedistribution
ofF is
known,
we caneasily
calculate
thedistribution
of R(X) by simulation:

1. Drawarandom
number m fromG = Poisson(4699).
2. Drawarandom
sample, X1,....,XM fromFF
3. Calculate the Chain Ladder reserve based on X1,....,XM
4. Repeat
1-3
200times.Theresults areshowninfigure7.
Outstanding
ClaimsReservc
Distribution

figure 7

229
Thereserveestimate based ontheobservation
x isexceeding
theexpected
value E(RES)byapproximately
105andthis
error
isinherited
intheentire
distribution
of. However,theBootstrap
error
distribution
has no systematic
erroranditapproximates
totheactual
error
distribution
reasonably
well, seefigure8
below.

R1(Y) R1(X) R1(Y)-E(R1(Y)) R1(X)-E(R1(X))

Mean 2253 2130 0 0

STD 214 179 214 179
Median 2247 2118 -6 -12
75% fraction
2373 2229 120 99
95% fraction.
2622 2456 369 326
98% fraction
2713 2508 460 378
figure8

Please
note,
that
thedistribution
of andofthee r r o r is
calculated
without
usingtheknowledge
oftheunderlying
distribution
F.

Application
2:Estimation
ofthetotal
uncertainty

We consider
areserving
methodR.Thetotal uncertainty
consists
oftheestimation
uncertainty
(related
tothepast)
plustheuncertainty
related
tothefuture
payments
RES-E(RES).
We assumethat
theselected
modelis‘correct’
sothatthereserve
estimate
isunbiased,
i.e.
E(R(X))
= E(RES).

Inorder
toestimate
reserving
marginswe would
estimate
thedistribution
of the
stochastic
variable
R + ( RES-E(RES)).

Example
2

Again we consider
thedistribution
outlined
inexample
1andthereserving
method
RI defined above.
Sincetheclaims
are
independent
itfollows
that
R(X) and (RES-
E(RES))areindependent andwe therefore
onlyhavetocalculate
the
convolution of the distributions
of and(RES-E(RES)),
forexample by
simulation.

of x)isapproximated
Thedistribution bytheBootstrap
distribution
. The
distribution
ofRES-E(RES)could havebeenestimated
bysimulation
on the
basis
oftheestimated
parametersandthemodelassumptions.
However,the
distribution
hasbeensimulated
onthebasisoftheoriginal
parameters.
The
results
areoutlined
infigure
9 below:

230
 RES
R1(Y)- -E(RES)Total
estimated
Bootstrap
Mean 2253 0 2253
STD 214 77 230
Median 2247 -4 2248
75%fraction
2373 39 2380
95%fraction
2622 120 2632
99%fraction
2713 166 2758
figure9
It
isseen,
that
the
maincontribution
tothe uncertainty
of‘Total’
rises
from the
estimation
uncertainty.
Forexample,the95%fraction
isonly
increased
from
2622
to2632when
thefuture
randomness isincluded.
Itis
often
seen,
asinthis
example,
that
focus
should
beontherandomnessinthepast
rather
thaninthe
future
whenreserving
margins
areestimated,
Application
3:Selection
ofreserving
method
Weconsider
twodifferentreserving
methods
and .Assume
that
both
metho.ds
are
unbiased
estimatori.
se.

Wewould
then
prefer
to
use rather
than if

The
problem
is
that
the
distributions and of
are
unknown.
However,
using
Bootstrap
approximation,
wecan easily
estimate
the
variances
of and
and
base
the
selection
onthese.
Example3
Weconsider
the
distribution
outlined
in
example
1and
two
different
reserving
methods:
R1: Deterministic
Chain
Ladder
method
based
onthe
payments
(as
above)
R2: A stochastic
model
with
unknown
but
constant
claim
inflation.
Both methodsare
reasonably
central/unbiased,E(R1(X))=2130,E(R2(X))=2145and
E(RES)=2143.

231
Itisseen
fromfigure
10that
thestochastic
model gives
amoreprecise
reserve
estimate
butfurther
more,
thestandard
deviation
of isless
than
that
oftheChain
Laddermethod

R1(Y) R1(X)- R2(Y)- R2(X)-

Real dist
Bootstrap’ Real dist
Bootstarp
Mean 2253 2130 2134 2145
STD 214 179 177 153
Median 2247 2118 2137 2144
75% fraction 2373 222 225 2244
95% fraction 2622 2456 2414 2418
98% fraction 2713 2508 2505 2487
figure 10

Conclusion
Itisconcluded
thatiftheindividual
claim
data areavailable
theBootstrapisan
effective
methodtocalculate
approximations
totheuncertainty
distributions
of
claimsreserves.
Onlydistribution
assumptionsregarding
theclaimfrequency
arerequired.
The application
isnotdependent on thereserving
method and
evenforcomplicated reserving
methodstheuncertaintycan be estimated.
It
canbe usedtoselectbetweendifferent
reservingmethodstoobtain the most
robustmethodandtocalculate
reserving
margins.

References
Efron,B.(1979),
Bootstrap
methods:
Another
look
atthe
jacknife.
Ann.Stat,
7.p.
1-
26.
Efron B.(1982)TheJacknife,
TheBootstrap
andOther
Resampling
Plans. Society for
IndustrialandApplied
Mathematics.,
Regional
Conference
series
in Applied
Mathematics.

232