-
1/18
Modelling Dependence in Insurance ClaimsProcesses with Lévy
Copulas
Benjamin Avanzi∗ Luke C. Cassar† Bernard Wong∗
Actuarial StudiesAustralian School of BusinessUniversity of New
South Wales
ASTIN & AFIR ColloquiaMadrid, 19-22 June 2011
∗The authors acknowledge support from an Australian Actuarial
Research Grant†
The author acknowledges support from an Ernst & Young
Honours Scholarship
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
2/18
Outline
1 Motivation
2 Literature review
3 Lévy copulas for Compound Poisson processes
4 Presentation of some Lévy copulas
5 Fitting Lévy copulas to real data
6 Conclusion
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Motivation 3/18
Outline
1 Motivation
2 Literature review
3 Lévy copulas for Compound Poisson processes
4 Presentation of some Lévy copulas
5 Fitting Lévy copulas to real data
6 Conclusion
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Motivation 3/18
Dependence between insurance claims processes
A single event may give rise to claims in multiple insurance
riskprocesses due to dependence between types of claims or
classesof business.
Such dependence has potential implications for pricing,
reserving,solvency, and capital allocation of an insurance
company.
Some standard approaches in the literature include (1)
commonshock models, and (2) distributional copulas, applied to
aggregateclaims/claim counts.
Lévy copulas straddles the advantages of these approaches. It
isTime-consistent;Coherent modelling of dependence in frequency
separate todependence in severity;Parsimonious;Make full use of the
available data;Enables a "bottom-up" approach to model
building.
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Literature review 4/18
Outline
1 Motivation
2 Literature review
3 Lévy copulas for Compound Poisson processes
4 Presentation of some Lévy copulas
5 Fitting Lévy copulas to real data
6 Conclusion
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Literature review 4/18
Existing Literature
Theoretical results:original: Tankov (2003); Cont and Tankov
(2004); Kallsen andTankov (2006)also: Barndorff-Nielsen and Lindner
(2007); Bäuerle et al. (2008);Eder and Klüppelberg (2009)
Applications:Bregman and Klüppelberg (2005): ruin probabilities
with twoprocesses (Clayton Lévy copula)Esmaeili and Klüppelberg
(2010a,b): fitting of the Clayton LévycopulaBäuerle and Blatter
(2011): optimal investment and reinsuranceproblemsOperational risk
modelling: Böcker and Klüppelberg (2008);Biagini and Ulmer (2009);
Böcker and Klüppelberg (2010)
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Lévy copulas for Compound Poisson processes 5/18
Outline
1 Motivation
2 Literature review
3 Lévy copulas for Compound Poisson processes
4 Presentation of some Lévy copulas
5 Fitting Lévy copulas to real data
6 Conclusion
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Lévy copulas for Compound Poisson processes 5/18
Bivariate compound Poisson process
!
" # $%
constituted of unique (⊥)and common (‖) jumps:{
S1(t) = S⊥1 (t) + S‖1(t)
S2(t) = S⊥2 (t) + S‖2(t)
S‖1(t) and S‖2(t) with
intensity λ‖
Joint survival function ofcommon jumps F
‖(x1, x2)
(may be dependent)
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Lévy copulas for Compound Poisson processes 6/18
Sklar’s Theorem for Lévy copulas
A Lévy copula C couples the marginal tail integrals and the
joint tailintegral so that
C(U1(x1),U2(x2)) = U(x1, x2)
For the marginal compound Poisson processes Si(t) (i = 1,2)with
Poisson parameter λi and jump size survival function F i(x),the
tail integral Ui(x) is given by
Ui(x) = λiF i(x).
The joint tail integral measures jumps which occur
simultaneouslyand is given by
U(x1, x2) = λ‖F‖(x1, x2).
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Lévy copulas for Compound Poisson processes 7/18
Effect of the Lévy copula
On the common jumps:
λ‖ = C(λ1, λ2) (intensity of common jumps)
F‖(x1, x2) =
1λ‖
C(λ1F 1(x1), λ2F 2(x1)) (joint survival function)
F‖1(x) =
1λ‖
C(λ1F 1(x), λ2) (marginal survival function 1)
F‖2(x) =
1λ‖
C(λ1, λ2F 2(x)) (marginal survival function 2)
On the unique jumps:
λ⊥i = λi − λ‖ (intensity of unique jumps)
F⊥i (x) =
1λ⊥i
(λiF i(x)− λ‖F
‖i (x)
)(marginal survival function i)
(see Böcker and Klüppelberg, 2008)Modelling Dependence in
Insurance Claims Processes with Lévy Copulas Avanzi, Cassar and
Wong (2011)
-
Presentation of some Lévy copulas 8/18
Outline
1 Motivation
2 Literature review
3 Lévy copulas for Compound Poisson processes
4 Presentation of some Lévy copulas
5 Fitting Lévy copulas to real data
6 Conclusion
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Presentation of some Lévy copulas 8/18
Pure Common Shock Lévy copula
The Pure Common Shock (PCS) Lévy copula is given by
C(u1,u2) =δ(u1 ∧ λ1)(u2 ∧ λ2) + [u1 − δλ2(u1 ∧ λ1)]I{u2=∞}+ [u2
− δλ1(u2 ∧ λ2)]I{u1=∞},
for 0 ≤ δ ≤ min(
1λ1,
1λ2
).
where λ1 and λ2 are the Poisson parameters for the
bivariatecompound Poisson process (when used in that context), and
where δis the parameter which will determine the intensity of
common jumps.
Leads to a bivariate compound Poisson process with dependencein
frequency onlyUnique and common jump sizes are all independent
andidentically distributed within the different processes
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Presentation of some Lévy copulas 9/18
Archimedean models
Tankov (2003) shows that for some strictly decreasing and
convexfunction φ,
C(u1,u2) = φ−1 (φ(u1) + φ(u2))
defines a bivariate Archimedean Lévy copula.(this can easily be
extended to the multivariate case)
Example:The generator φ(u) = u−δ yields the Clayton Lévy
copula
Cδ(u1,u2) =(
u−δ1 + u−δ2
)− 1δ.
The Clayton Lévy copula has unique properties that do not extend
toother Lévy copulas in general.Bregman and Klüppelberg (2005);
Böcker and Klüppelberg (2008); Esmaeili and Klüppelberg (2010a,b);
Tankov (2003); Cont and Tankov (2004); Kallsen and Tankov (2006);
Böcker and Klüppelberg (2010); Biagini and Ulmer (2009);
Barndorff-Nielsen and
Lindner (2007); Bäuerle et al. (2008); Bäuerle and Blatter
(2011); Eder and Klüppelberg (2009)
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Presentation of some Lévy copulas 10/18
The Lévy copula density
is defined as
c(u1,u2) =∂2C(u1,u2)∂u1∂u2
, where ui = Ui(xi), i = 1,2.
It is useful to compare Lévy copulas:Its shape indicates the
relative prevalence of common jumps atdifferent sizes in each
component.It reflects the dependence in frequency, as the volume
under thedensity on [0, λ1]× [0, λ2] is the expected number of
commonjumps per unit time.
In what follows, consider a bivariate compound Poisson process
withλ1 = 100, λ2 = 100 and a Lévy copula parameter such that λ‖ =
60.
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Presentation of some Lévy copulas 11/18
Pure Common Shock and Clayton densities
Pure Common Shock Clayton
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Presentation of some Lévy copulas 12/18
Archimedean model I & II densities
Archimedean model I Archimedean model II
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Fitting Lévy copulas to real data 13/18
Outline
1 Motivation
2 Literature review
3 Lévy copulas for Compound Poisson processes
4 Presentation of some Lévy copulas
5 Fitting Lévy copulas to real data
6 Conclusion
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Fitting Lévy copulas to real data 13/18
Description of the data
We use data provided by SUVA, a Swiss worker’s
compensationcompany incorporated under public law.
The dataset consists of a random sample of 5% of claims from
theconstruction sector for accidents incurred in 1999 (developed
asat 2003).
It features two classes of claims: 2249 medical claims and
1099daily allowance claims.
1089 claims are common to both classes.
Compound Poisson processes S1(t) and S2(t), with
dependencecharacterised by a Lévy copula, are fit to the SUVA
dataset.
We use the Gumbel and Normal distribution for the log of
claimsizes in the two classes.
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Fitting Lévy copulas to real data 14/18
Maximum likelihood estimation results
MLEs using new Lévy copulas are
Lévycopula
ln L δ̂ λ̂1 λ̂2 λ̂‖
AM1 8631.27 0.0025358 2239.42 1113.32 1093.74(0.0001110) (47.20)
(32.75)
Clayton 8536.43 2.2632 2176.90 1066.27 984.16(0.0688161) (46.26)
(31.68)
PCS 7845.03 0.0004406 2249.00 1099.00 1089.00(0.0000093) (47.42)
(33.15)
Empirical: 2249.00 1099.00 1089.00
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Fitting Lévy copulas to real data 15/18
Comparing the fit of tail integrals
We can compare the theoretical tail integrals of unique and
commonjumps for each class with the observed number of jumps above
size x .
This provides a global idea of the goodness-of-fit, as it
considers bothfrequency (initial value) and severity (shape).
For the Pure Common Shock Lévy copula:Medical (Unique)
0 2 4 6 8 10 12 14
0200
400
600
800
1000
1200
Logarithm of medical claims (unique)
Tail
inte
gra
l
Medical (Common)
0 2 4 6 8 10 12 14
0200
400
600
800
1000
1200
Logarithm of medical claims (common)
Tail
inte
gra
l
Allowance (Common)
0 2 4 6 8 10 12 14
0200
400
600
800
1000
1200
Logarithm of allowance claims (common)
Tail
inte
gra
l
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Fitting Lévy copulas to real data 16/18
For the Clayton Lévy copula:Medical (Unique)
0 2 4 6 8 10 12 14
02
00
40
06
00
80
01
00
01
20
0
Logarithm of medical claims (unique)
Tail
inte
gra
l
Medical (Common)
0 2 4 6 8 10 12 14
02
00
40
06
00
80
01
00
01
20
0
Logarithm of medical claims (common)Ta
il in
teg
ral
Allowance (Common)
0 2 4 6 8 10 12 14
0200
400
600
800
1000
1200
Logarithm of allowance claims (common)
Tail
inte
gra
l
For the Archimedean Model I Lévy copula:Medical (Unique)
0 2 4 6 8 10 12 14
0200
400
600
800
1000
1200
Logarithm of medical claims (unique)
Tail
inte
gra
l
Medical (Common)
0 2 4 6 8 10 12 14
02
00
40
06
00
80
01
00
01
20
0
Logarithm of medical claims (common)
Tail
inte
gra
l
Allowance (Common)
0 2 4 6 8 10 12 14
0200
400
600
800
1000
1200
Logarithm of allowance claims (common)
Tail
inte
gra
l
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Conclusion 17/18
Outline
1 Motivation
2 Literature review
3 Lévy copulas for Compound Poisson processes
4 Presentation of some Lévy copulas
5 Fitting Lévy copulas to real data
6 Conclusion
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Conclusion 17/18
Summary
Introduction to Lévy copulas
Application to the modelling of insurance risk processes
Development of new Lévy copula models
Comparison of different Lévy copula models
Fitting to a new set of real data
Discussion of the goodness-of-fit of the candidate models
Also: Formal decomposition of the trivariate case in the
paper
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
-
Main References 18/18
Main References
Barndorff-Nielsen, O. E. and Lindner, A. M. (2007). Lévy
Copulas: Dynamics and Transforms of Upsilon Type.
ScandinavianJournal of Statistics, 34:298–316.
Bäuerle, N. and Blatter, A. (2011). Optimal control and
dependence modeling of insurance portfolios with Lévy
dynamics.Insurance: Mathematics and Economics, 48(3):398–405.
Bäuerle, N., Blatter, A., and Müller, A. (2008). Dependence
properties and comparison results for Lévy processes.
MathematicalMethods of Operations Research, 67:161–186.
Biagini, F. and Ulmer, S. (2009). Asymptotics for operational
risk quantified with expected shortfall. ASTIN
Bulletin,39(2):735–752.
Böcker, K. and Klüppelberg, C. (2008). Modelling and Measuring
Multivariate Operational Risk with Lévy Copulas. Journal
ofOperational Risk, 3:3–27.
Böcker, K. and Klüppelberg, C. (2010). Multivariate models for
operational risk. Quantitative Finance, pages 1–15.
Bregman, Y. and Klüppelberg, C. (2005). Ruin estimation in
multivariate models with Clayton dependence structure.Scandinavian
Actuarial Journal, 2005(6):462–480.
Cont, R. and Tankov, P. (2004). Financial Modelling With Jump
Processes. Chapman & Hall/CRC, London.
Eder, I. and Klüppelberg, C. (2009). The quintuple law for sums
of dependent Lévy processes. The Annals of Applied
Probability,19(6):2047–2079.
Esmaeili, H. and Klüppelberg, C. (2010a). Parameter estimation
of a bivariate compound Poisson process. Insurance:Mathematics and
Economics, 47(2):224–233.
Esmaeili, H. and Klüppelberg, C. (2010b). Two-step estimation of
a multivariate lévy process. Available
athttp://www-m4.ma.tum.de/Papers/.
Kallsen, J. and Tankov, P. (2006). Characterisation of
dependence of multidimensional Lévy processes using Lévy
copulas.Journal of Multivariate Analysis, 97(7):1551–1572.
Tankov, P. (2003). Dependence structure of spectrally positive
multidimensional Lévy processes. Available
athttp://www.math.jussieu.fr/ tankov/.
Modelling Dependence in Insurance Claims Processes with Lévy
Copulas Avanzi, Cassar and Wong (2011)
MotivationLiterature reviewLévy copulas for Compound Poisson
processesPresentation of some Lévy copulasFitting Lévy copulas to
real dataConclusion