Title: Chain Ladder Bias and Variance Approximations under a Compound Poisson Model
1Chain Ladder Bias and Variance Approximations
under a Compound Poisson Model
2Overview
- Part 1
- An outline of Yogaranpan, J., Clarke, S., Ferris,
S. and Pollard, J. Approximating the Bias and
Variance of Chain Ladder Estimates Under a
Compound Poisson Model. Journal of Actuarial
Practice 11 (2004) 147-167. - Part 2
- Further research
- Consequences for GLMs applied to run-off
triangles - Extending the theory - run-offs with iid entries
3Part 1 Yogaranpan et al. (2004)
Known incremental claims and outstanding claims
estimates
OS1
OS2
OSn-2
OSn-1
Accident years and development years are numbered
from 0 to n-1
4The Chain Ladder Method
Chain ladder development ratio
Relevant product of development ratios
O/S claims estimate for accident year i
5Stochastic Assumptions
Claim sizes are required to be iid (here
exponential variables with a mean of 500 are
assumed)
6Simulation Study
- Results for biases of simulated outstanding
claims
7Why are there biases?
- An analogy an out of the money call option is
not worthless, there is a time value which
depends on stock price volatility. - Similarly with outstanding claims estimation -
volatility has an impact on the mean. - Fundamentally the expectation of a ratio does not
equal the ratio of expectations.
8Development Ratio Biases
- Consider a development ratio m of the form m TA
/ TB
The VBA term is the approximate proportional bias
in m
9Intermediate formulae
- Define Vn-iC as the sum of all V terms
corresponding to the development ratios that
constitute Mi - Define t1 and t2 as the first and second order
moments about the origin of the claims size
distribution - Define g as the ratio of the variance to the mean
for each of the run-off entries. g is the same
for all run-off entries - Define ai as the expected ultimate total claims
cost for accident year i, so that ai li t1 - Define
10Practical formulae final results
11Simulation Study
- Biases of outstanding claims
12Simulation Study
- Simulated vs. approximate covariance matrices
Overall Variance 4.362108
Overall Variance 4.270108
13Simulation Study
- The approximate bias in the overall outstanding
claims estimate is quite accurate (854 vs. 870) - The approximate variance of the overall
outstanding claims estimate is only 2.1 less
than the value from the simulations (4.27108 vs.
4.36108) - The simulations used highly skewed claims sizes
(exponential with a mean of 500). With less
skewed distributions the approximations match the
simulations even more accurately.
14Features of the Approximations
- The approximate biases are linear in g as has
been observed in other papers - An unintuitive result if the portfolio grows
such that all ais increase by a common factor
(eg, doubling in size), then (ceteris paribus)
the approximate biases do not change. - Approximate covariances between accident year
claims estimates grow in proportion to portfolio
size. - The first two terms of the Var(OSi) approximation
behave similarly to the covariances, although the
last term is effectively constant in respect of
changes in portfolio size (similar to biases).
15Shortcomings of the theory
- Restrictive assumptions
- Skewness statistics cannot be found
- Parameter values are required in advance -
estimation creates further biases and variances - Inflation has been ignored
- Biases may be immaterial
- Chain ladder methods are not widely used for
incurred costs - Computer simulations are fast and are probably
more practical than learning the theory
16Part 2 Further Research
- Implications for bias and variance estimates
based on GLMs - Extending the theory iid run-off entries
17GLMs and run-off triangles
- Poisson Count Model
- Well known result of the equivalence of maximum
likelihood estimates (under the Poisson count
model) and Chain Ladder estimates - It follows that the biases and variances are also
equivalent - The Poisson data assumptions are the same as
those used in deriving the approximations, albeit
with claim sizes restricted to a value of 1 - The features of the Poisson count model
outstanding claims estimates are therefore best
described by the approximations
18GLMs and run-off triangles
- GLMs are also based on maximum likelihood
- If the simple Poisson count model cannot possibly
describe its own complex biases, variances and
covariances, what about more intricate GLMs and
other MLE methods? - Serious errors could exist when non trivial GLMs
are applied to run-off triangles - We can expect the complexities of the biases and
variances of the estimates to grow with the
intricacy of the run-off models
19Extending the theory
- The key to the entire theory is the fact the
ratio of the variance to the mean (ie, g) is the
same for every cell of the run-off. - This is true for the compound Poisson model where
the claim size distribution is restricted to be
iid. The same can also be found for compound
binomial and compound negative binomial models
under very severe restrictions. - While the constant g assumption is generally not
true for all other models, it is satisfied if
run-off entries are iid.
20iid run-off entries
- If run-off entries are iid with mean m and
variance s2, then the approximations do not
simplify appreciably - Trial and error has lead to the following
approximations based on the approximations for
the overall chain ladder outstanding claims
estimate
21iid run-off entries
- The iid approximations were tested by simulating
a variety of distributions, and in each case the
bias and variance of the overall outstanding
claims estimate were compared to the
approximations. - In general there is excellent agreement with the
iid approximations, except when very small
denominators are quite likely in the development
ratios (for example, when entries are
exponentially distributed)
22(No Transcript)
23(No Transcript)
24(No Transcript)
25(No Transcript)
26Summary
- Reliable bias and variance approximations have
been found for chain ladder estimates under a
compound Poisson model - The results can be linked to maximum likelihood
methods, raising questions about the reliability
of GLMs and other MLE methods applied to run-off
triangles - The results can be extended, under severe iid
restrictions, to non Poisson models