pavel v. shevchenko quantitative risk management csiro ... · • basel ii requires that banks hold...
TRANSCRIPT
Quantitative modelling of financial risks
Pavel V. ShevchenkoQuantitative Risk Management
CSIRO Mathematical and Information Sciences
Vienna University of Technology12 January 2010
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
CSIRO Mathematical and Information Sciences
Commonwealth Industrial Research Organisation (CSIRO) AustraliaMathematical and Information Sciences (CMIS)Financial risk teamIndustry: banks, insurance/electricity companies, hedge fundsActivities/modes of engagement: strategic research, consulting, model
development/validation, software developmentBeneficiaries of our work in risk management:• Financial institutions and their shareholders (risk reduction)• Stability of banking/insurance systems• Science and CMISOur mission – To conduct research in financial risk quantification,
measurement and calculation; To develop innovative quantitativemethods and tools in financial engineering
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Financial Risks – quantifying uncertainties
Risks are modelled by random variables mapping unforeseen future states of the world into values representing profits and losses.
Market Risk: modelling of losses incurred on a trading book due market movements (interest rates, exchange rates, etc)
Credit risk: modelling of losses from credit downgrades and defaults (e.g. loan defaults)
Operational Risk: modelling of losses due to failed internal processes, people or external events (e.g. human errors, IT failures, natural disasters, etc)
Quantifying the uncertainties: process uncertainty, parameter uncertainty, numerical error, model error
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Links•Extreme Value models•Expert Elicitation•Combining expert&data•Bayesian methods•Dependence modelling•Numerical methods (PDE, MCMC, MC)•Time series analysis•State-space models•High performance computing•Optimisation
Other risk areas•Air transport•Ecology•Environment•Infrastructure •Security•Weather/Climate•Health
Financial Risk• Market Risk • Credit Risk • Operational Risk• Underwriting risk • Derivative pricing• Interest Rates• Trading strategies• Portfolio Management• Commodity/Energy• Carbon Trading• Liquidity risk
Financial risk and other risk areas
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Regulatory Requirements for banking industry
•Regulatory standards in banking industry:Basel Committee for banking supervision 1974Basel I, 1988: Credit Risk, Basel I Amendment 1996: Market Risk (VaR) Basel II 2001- ongoing: Credit Risk, Operational Risk
•Insurance regulation: International Association of Insurance SupervisorsJoint Forum on Financial Conglomerates 1996Solvency II (similar to Basel II for banks)
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
2010 - Basel III?
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Basel II• Basel II requires that banks hold adequate capital to protect
against Market risk, Credit risk and Operational Risk losses. Three-pillar framework:
Pillar 1: minimal capital requirements (risk measurement)Pillar 2: supervisory review of capital adequacyPillar 3: public disclosure
Advanced Measurement Approaches allow internal models for calculations of capital as a large quantile of loss distribution.Operational Risk is new risk category
• In Australia, the national regulator APRA is now applying the Basel II requirements.
• Solvency 2 for insurance industry
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Regulatory Requirements – Risk MeasuresValue at Risk (VaR)
loss probability density
0
0.01
0.02
0.03
0.04
0$ loss
Expected Loss
Unexpected Loss
10daysover 0.99 :RiskMarket 1year over 999.0 :RiskCredit year 1over 999.0 :Risk lOperationa
}1]Pr[:inf{ :measureRisk
loss expectedloss Unexpected
=•=•=•
−≤>=•
−=•
qqq
qxLossxVaR
VaR
q
q
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Data sufficiency
0.1)get tolosses annual 100,000or (0.1get to000,000,1 )9999.0()999.0( then ,10][ if :Example
])[/)1(1()());(1]([~)(1
,)1(;sfrequencie some and dfs edheavy tailFor
]Pr[),(~;...
1.18get we1000for 0.1;accuracy 10%get to986,140)2,0(:Example
/][stdev2get to))((
)1(4)(
)1(][stdev
~...~~,~quantile)(from,...,
11
110
1
22
211][
1
=====
−−=−−
∞<+
==++=•
====⇒==
=−
=⇒−
=•
≤≤≤=••
−−
−−
∞
=
+
∑
εε
αα
δ
εεσμ
εε
αααααα
αααα
αα
nFFNE
NEFFzFNEzF
p
pnNxFXXXZ
nnLN
qqqqf
nnqf
q
xxxxqxfiidXX
XZ
XZXZ
nn
n
nXiN
nn
n
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Market Risk
}60
,max{VaR)(
Capitalry k RegulatoMarket Ris
by modelled often is of tailThe
)(,:exampleFor
stationaryisandiidare,...,assume
/)()ln(ln
etc FX,s,commoditie stocks, ),...,();(
contracts financial of Portfolio
60
1
10,199.0
10,0.99
21
21110
21
1
111
1
)()1(
∑
∑
=
+−
−−−−
−
−−−
=
=
+−+==
+=−≈−−=
==Π
i
itt
tttttt
ttt
tttt
tttttt
K
k
Mttttkt
VaRktC
GPDZ
XX
XZZZX
SSSSSX
SSQ
K
βσμαασλμ
σμ
SS
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Credit Risk
)(VaR :Capital
economy) theof (statefactor systematic is )1,0(~ ,1 );exp(
)1,0(~ ),1,0(~ ;1
otherwise ,0)(,1
:(2003)Pykhtin model,factor Latent
defaultof prob. is]1Pr[ ;0or1collateral is loss; is ,0]-max[1
:obligors of Portfolio
0.999
22
2
1
1
LC
YNWWZYRRC
NZNYZYX
pNXI
pIICCL
ILwALM
iiiiiiiiiiii
iiiii
iii
iii
iii
M
iiii
=
−−++=+=
−+=
⎩⎨⎧ <
=
====
×=•
−
=∑
γβγβσμ
αα
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Operational Risk: Loss Distribution Approach
• Severity Distribution type F(x):e.g. Lognormal, Gamma, Pareto, g-and-h, EVT, mixtures
• Frequency distributions P(N): e.g. Poisson, Negative Binomial, Binomial
• Annual loss
• Typical assumptions:
• Distribution of Z: Monte Carlo, Fast Fourier Transform, Panjerrecursion, direct inversion of characteristic Functions.
∑=
=N
iiXZ
1
tindependenare)(andtindependenare),...,1(and
jiXXNiXN
ji
i
≠=
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Basel II, Operational Risk
•Basel II defined Operational Risk as: the risk of loss resulting from inadequate or failed internal processes, people and systemsor from external events. This definition includes legal risk but excludes strategic and reputational risk.
•Losses can be extreme, e.g.1995: Barings Bank (loss GBP1.3billion)1996: Sumitomo Corporation (loss US$2.6billion)2001: September 112001: Enron, USD 2.2b (largest US bankruptcy so far)2002: Allied Irish, £450m2004: National Australia bank (A$360m)2008: Société Générale (Euro 4.9billion)
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Basel II Operational Risk
Three-pillar framework:• Pillar 1: minimal capital requirements (risk measurement)• Pillar 2: supervisory review of capital adequacy• Pillar 3: public disclosure
Operation Risk Capital Charge• The Basic Indicator Approach:
GI(j) is Gross Income in year j.• The Standardised Approach: (12-18% per business line)
• The Advanced Measurement Approaches (AMA): Internal model for 56 risk cells (7 event types x 8 business lines)
%.15,]0),(max[11
=×= ∑=
ααn
jjGI
nC
%]18%,12[,]0),(max[31 3
1
8
1∈∑ ×∑=
= =i
ji
ii jGIC ββ
∑=
>=3
1}0)({1
jjGIn
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Operational Risk Basel II Business Lines/Event Types
Basel II Business Lines (BL)• BL(1) - Corporate finance• BL(2) - Trading & Sales• BL(3) - Retail banking• BL(4) - Commercial banking• BL(5) - Payment & Settlement• BL(6) - Agency Services• BL(7) - Asset management• BL(8) - Retail brokerage
Basel II Event Type (ET)•ET(1) - Internal fraud:•ET(2) - External fraud: •ET(3) - Employment practices and
workplace safety: •ET(4) - Clients, products and
business practices•ET(5) - Damage to physical assets:•ET(6) - Business disruption and
system failures: •ET(7) - Execution, delivery and
process management:
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Hierarchical and matrix structures
Business Line (BL) Risk Type (RT)
top levelBL1
BL8
BLi
RT1 RT7RTk
Basel II risk matrix
Zik
Bank hierarchical structure
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Loss Data Collection Exercise 2007 Japan - Annualized data: Number of events(100%=947.7) / Total Loss Amount (100%=JPY 22,650mln)
100% (no.events)100% (loss)
38.6%54.6%
10.9%3.9%
1.9%4.5%
8.8%24.8%
1.5%1.0%
36.5%8.3%
1.8%2.9%
Total
0.4%0.2%
0.02%0.00%
0.09%0.00%
0.13%0.00%
0.00%0.00%
0.02%0.00%
0.11%0.00%
0.09%0.18%
Other
3.5%2.0%
1.06%0.13%
0.18%0.04%
0.00%0.00%
1.40%0.53%
0.01%0.00%
0.00%0.00%
0.84%1.32%
BL(8)
2.1%1.5%
1.50%1.02%
0.05%0.00%
0.00%0.00%
0.48%0.49%
0.1%0.04%
0.00%0.00%
0.00%0.00%
BL(7)
5.3%3.8%
3.14%3.00%
1.19%0.22%
0.01%0.00%
0.92%0.62%
0.01%0.00%
0.00%0.00%
0.00%0.00%
BL(6)
0.6%0.1%
0.23%0.00%
0.38%0.04%
0.00%0.00%
0.00%0.00%
0.00%0.00%
0.00%0.00%
0.00%0.00%
BL(5)
25.7%45.4%
15.05%16.87%
5.87%3.22%
0.69%4.19%
2.02%18.32%
1.16%0.71%
0.64%1.99%
0.29%0.13%
BL(4)
57.2%21.5%
13.21%8.43%
2.83%0.26%
1.11%0.26%
3.62%4.77%
0.14%0.18%
35.73%6.36%
0.54%1.24%
BL(3)
4.7%25.2%
4.12%25.03%
0.26%0.00%
0.00%0.00%
0.22%0.09%
0.04%0.04%
0.00%0.00%
0.01%0.00%
BL(2)
0.5%0.3%
0.28%0.18%
0.07%0.04%
0.00%0.00%
0.11%0.09%
0.01%0.00%
0.01%0.00%
0.00%0.00%
BL(1)
TotalET(7)ET(6)ET(5)ET(4)ET(3)ET(2)ET(1)
Ban
k of
Jap
an, e
t al (
2007
)
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Loss Data Collection Exercise 2004 USA - Annualized data: Number of events(100%=18371) / Total Loss Amount (100%=USD 8,643mln)
Fede
ral R
eser
ve S
yste
m, e
t al (
2005
)
100.0%, 100.0% (US$)
3.2%0.6%
0.8%0.1%
35.3%9.6%
0.7%0.8%
0.7%1.4%
9.2%79.8%
7.6%1.7%
39.0%5.1%
3.40%0.9%
Total
8.0%70.8%
0.08%0.01%
0.07%0.05%
3.45%0.98%
0.02%0.44%
0.12%1.28%
0.40%67.34%
1.75%0.34%
1.66%0.3%
0.42%0.1%
Other
7.3%1.6%
0.20%0.07%
2.20%0.25%
0.01%0.00%
3.30%0.94%
1.38%0.33%
0.10%0.02%
0.06%0.03%
BL(8)
2.4%2.5%
0.09%0.01%
1.82%0.38%
0.04%0.01%
0.00%0.00%
0.13%2.10%
0.10%0.02%
0.26%0.02%
0.00%0.00%
BL(7)
5.1%1.1%
4.52%0.99%
0.14%0.02%
0.01%0.01%
0.31%0.06%
0.04%0.02%
0.03%0.01%
0.01%0.02%
BL(6)
4.5%0.6%
0.23%0.05%
0.01%0.00%
2.99%0.28%
0.05%0.02%
0.01%0.00%
0.04%0.01%
0.18%0.02%
0.44%0.13%
0.52%0.08%
BL(5)
5.1%1.8%
0.44%0.04%
0.02%0.00%
1.38%0.28%
0.03%0.00%
0.01%0.00%
0.36%0.78%
0.17%0.03%
2.64%0.70%
0.05%0.01%
BL(4)
60.1%12.3%
2.10%0.26%
0.69%0.06%
12.28%3.66%
0.21%0.21%
0.56%0.1%
4.41%4.01%
3.76%0.87%
33.85%2.75%
2.29%0.42%
BL(3)
7.3%8.6%
0.05%0.15%
6.55%2.76%
0.24%0.06%
0.03%0.00%
0.19%4.29%
0.17%0.05%
0.01%1.17%
0.02%0.10%
BL(2)
0.3%0.5%
0.01%0.00%
0.03%0.01%
0.12%0.05%
0.00%0.00%
0.08%0.30%
0.06%0.03%
0.01%0.00%
0.01%0.14%
BL(1)
TotalFraudOtherET(7)ET(6)ET(5)ET(4)ET(3)ET(2)ET(1)
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Internal
Loss Data
External
Loss Data
Scenario analysis
Risk Frequency and Severity distributions
Annual loss distribution (per risk and over all risks)
Annual Capital Charge (Capital Allocation, sensitivity, what if scenarios)
Business environment and internal control factors
Insurance Dependence between risks
Operational Risk - Loss Distribution Approach
Monte Carlo simulations
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Operational Risk: Combining internal data, industry data and expert opinions
• BIS (2006), p.152: “Any operational risk measurement system must have certain key features to meet the supervisory soundnessstandard set out in this section. These elements must include the use of internal data, relevant external data, scenario analysis and factors reflecting the business environment and internal controlsystems.”
• An interview with four industry’s top risk executives: Davis, E. (1 September 2006). Theory vs Reality, OpRisk and Compliance: “[A] big challenge for us is how to mix the internal data with external data; this is something that is still a big problem because I don’t think anybody has a solution for that at the moment.” Or: “What can we do when we don’t have enough data [. . .] How do I use a small amount of data when I can have external data with scenario generation? [. . .] I think it is one of the big challenges for operational risk managers at the moment”.
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Ad-hoc procedures for combining internal data, industry data and expert opinions (SA)
)]ˆvar(:min[argˆ;1...,ˆ...ˆˆ.estimators moreor threecombine toextendedeasily be Can
ˆ,ˆ:)ˆvar( minimize to weightsChoose
)1()ˆvar(;1,ˆˆˆ :estimator unbiased Combined
.2,1,]ˆvar[;]ˆ[ i.e. ,parameter for ˆ and ˆ estimatest independen unbiased woconsider t : varianceMinimum
)()1()()(:onsdistributi Mixingseverity estimate data to external&internal
of sample combined andfrequency estimate data to Internal
111
22
21
21
222
21
22
1
22
21
21
21212211
221
21int21
totKKKtot
tot
tottot
mmm
extSA
wwwwww
ww
wwwwww
mE
xFwwxFwxFw
θθθθ
σσσ
σσσθ
σσθθθθ
σθθθθθθ
==++++=
+=
+=
−+==++=
===
•−−++•
•
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Combining two data sources: internal data with expert opinion (or external data) via Bayesian inference
data internal withdata) external (or opinion expert combining
h
h
XXX
K
n
on distributiposterior :)|(nsobservatio internal of likelihood:)|(
data) external(or expert by estimatedon distributiprior :)(
)()|()|(: theoremBayes'
),...,,( parameters Model),...,,( nsObservatio
21
21
XθθX
θ
θθXXθ
θX
π
π
ππ
θθθ
∝•
==
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Example: combining internal data and expert opinion
estimate of the arrival rate vs year
00.10.20.30.40.50.60.70.80.9
0 3 6 9 12 15year
arriv
al ra
te
Bayesian estimatemaximum likelihood estimate
estimator d Likelihoo Maximum- ˆ)15.0,41.3(prior thewithestimator Bayesian- ˆˆˆ
15.0,41.33/2]75.025.0Pr[,5.0][ opinionExpert 0.6)( from 0) 2, 1, 1, 2, 0, 1, 1, 1, 0, 1, 0, 0, 0, (0, counts Annual
11 ∑=
≈≈×=
≈≈⇒=≤≤===
=ki ik
MLEk
kkBk
N
Gamma
EPoissonN
λ
βαβαλ
βαλλλ
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Combining three data sources: internal&externaldata and expert opinion via Bayesian inference
ion)approximat Gaussian methods, MCMCpriors, (conjugatedensity posterior -),|(
opinionsexpert of likelihood- )|(nsobservatio internal of likelihood-)|(
dataindustry by estimated is ondistributiprior -)(
)()|()|(),|(
),...,,( ),...,,( ),...,,(sParameter opinions Expert data Internal
2
1
21
212121
υXθθυθX
θ
θθυθXυXθ
θυX
π
π
ππ
θθθυυυ
hh
hh
XXX KMn
∝
===
Joint work with ETH Zurich: H. Bühlmann, M.Wüthrich, D.Lambrigger
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Combining internal & external data with expertExample: Poisson-Gamma-Gamma
5.0)|(,7.0ˆ:Expert15.0,41.33/2]75.025.0Pr[,5.0][:data External
(0.6) from 0) 2, 1, 1, 2, 0, 1, 1, 1, 0, 1, 0, 0, 0, (0, :counts Annual
==≈≈⇒=≤≤=
=
λυυβαλλ
VcoE
PoissonN
year
Arr
ival
rate
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Modelling dependence
•Basel Committee statement“Risk measures for different operational risk estimates must be added for purposes of calculating the regulatory minimum capitalrequirement. However, the bank may be permitted to use internally determined correlations in operational risk losses across individual operational risk estimates, provided it can demonstrate to the satisfaction of the national supervisor that its systems for determining correlations are sound, implemented with integrity, and take into account the uncertainty surrounding any such correlation estimates (particularly in periods of stress). The bank must validate its correlation assumptions using appropriate quantitative and qualitative techniques.” BIS (2006), p.152.Remark: Adding capitals implies perfect positive dependence between risks. This can be conservative but can also underestimate the capital (e.g. the case of heavy tailed distributions).
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Dependence between risks
)parameters c(stochasti profilesrisk ebetween th Dependence•
, and losses annual ebetween th Dependence• timesame at the occured severities ebetween th Dependence•
cellsrisk many affecting eventscommon the viaDependence•
, and sfrequencie ebetween th Dependence•)](|[)(:(ES) Shortfall Expected•
failmay ation diversificformally and measure coherent anot is },)(,min{)()( :VaR•
)C()C()C( :capital ain ation diversific Possible•
...:
11
1
)(2
1
)2(1
1
)1(
1
jiZZ
jiNNZVaRZZEZES
zFzFZVaR
ZZZ
XXXZZlossannualtotal
ji
ji
ZZ
K
KN
i
Ki
N
ii
N
ii
K
kk
≠
≠>=
≥==
+…+≤
+++==
−
====∑∑∑∑
αα
α αα
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Negative diversification problem
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.6 0.8 1 1.2 1.4 1.6
gamma
dive
rsifi
catio
n at
0.9
99 q
uant
ile
);2006,2002( et.al. Embrechts ; andmean finitefor possible 0 large and 1for 0 ; large and 10for 0
][][][1 ,ation Diversific
1,/11)(~,t independen :Example
2999.01999.0
21999.0999.0
21
ααγαγ
γ
α
αα
γ
<>><<<
++
−=
≥−=
DDD
XVaRXVaRXXVaRD
xxParetoXX
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Linear correlation pitfalls
onsdistributi marginalgiven nt with inconsiste becan ),(11],,[),(
0 ),(on with distributi- tbivariate from ifdependent are and 0),(but dependent perfectly are (0,1), :exampleFor
ceindependenimply not does 0 ),(
functionnonlinear (.)or (.) if ),())(),((risks) somefor case (not the required are )var(),var( finite
gsShortcomin0,),,(),( :nsformlinear traunder invariant
dependencelinear of measure a is )var()var(
][][][),(
tcoefficienn correlatiolinear then variablesrandom are and If
maxminmaxmin
2
YXYX
YXYXYXXYX~N
YX
hgYXYhXgYX
YXYcXYX
YEXEYXEYX
YX
ρρρρρρ
ρρ
ρ
ρρ
βαργβαρ
ρ
≤≤≤−∈===
=
≠
•>=++
−×=
•
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
)(behaviour marginal ii xF structure dependence ),,...,(copula 1 duuC
),...,( ondistributijoint 1 dxxF
margins. of (.) ations transformincreasingunder invariant is Copula(1959)Sklar .continuous are margins if dfjoint any for unique is Copula
))(),...,((),...,(
))(),...,((),...,(
variablesrandom (0,1) uniform ofon distributijoint temultivaria is copula)(),...,( rvs ofon distributijoint a is )( Copula
11
111
111
111
i
ddd
ddd
ddd
T
uFuFFuuC
xFxFCxxF
XFUXFUlink
−−=
=
==•
Copula – dependence structure
ondistributijoint is ],...,[Prob),...,( and]Prob[)(~ with marginally ,..., rvsConsider
111
1
ddd
iiiiid
xXxXxxFxXxFXXX
≤≤=≤=•
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Copulas
));(...)((),...,(n Archimedia e.g. :copulasExplicit
(2009) Shevchenko and Luo )),(),...,(),(( ),...,,( copula- grouped dgeneralize
))(),...,(),...,(),(),...,((
,...,1 ),(
),...,,...,,,...,( copula- grouped(.). from sample a is ))(),...,((
.(.)on distributi temultivaria from sample a is ),...,(
square-Chi inv. theis )( , (0,1)~ ),(/)( where
);,...,( ~ ),...,( ),,...,(
df. temultivaria theis (.) df; standard theis (.)
))(),...,((),...,( :copula-
))(),...,((),...,( :copulaGaussian
11
1
2211
2211
2121121111
12121121111
1
)(1
1111
)(11
)(
11
1)(1
11
1)(1
dd
nn
nn
nmnnnn
kk
dmnnnn
td
td
k
dNdd
t
dtdt
dNNNdGa
uuuuC
XtXtXtZWZWZWt
XtXtXtXtXt
mkVGW
ZWZWZWZWZWtCYtYt
tXX
VuniformVVVGW
zzZZZWZW
ttt
ututuuCt
uFuFuuC
φφφ
χχν
ννν
ννννν
ν
ννν
ν
ννν
νν
νννν
++=•
=
′=•
=
==
=•=
Φ
==
Φ=
Φ
Φ=•
Φ=•
−
++
−
++
Σ
Σ
−−−
Σ
Σ
−−ΣΣ
−−ΣΣ
UX
U
XU
X
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Dependence between annual losses implied by dependence between frequencies
-1
-0.5
0
0.5
1
-1 -0.5 0 0.5 1
.parameter dependencecopula thevs),( 21 ρρ ZZS
).,(~),();( ),(
;10,5 );((.)~ ,(1,2)~
and
212121
2211
11
21)(
2
1
)2(2
1
1
)1(1
uuCUUUFNUFN
PoissonFNLNX
XZXZ
Gaiii
i
N
ii
N
ii
ρ
λλλ−−
==
==
===
∑=∑=
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Dependence via common Poisson process
1 yeartodayt
1st risk events
2nd risk events
kiCkiii
iCii
CCC
CC
CC
CC
ppNNptN
ptNtNtN
NN
PoissontNPoissontNPoissontNPoissontNPoissontN
tNtNtNtNtNtN
λ
λλλλλρ
λλλλλλλ
=⇒⎭⎬⎫
⎩⎨⎧
−+
=•
++=
++
++•
),cov(.1probwith)(ˆ
;probwith)()(ˆ)(
))((/),(
);(~)();(~)()(~)();(~)(~);(~)(~
)()(~~)();()(~~)(
2121
2211
2211
2211
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Dependence between risk profiles
(.)~(.);~
onsdistributi marginal e with th))(),...,(),(),...,((),(
(.))function copula (via profilesrisk for on distributijoint
) variableseconomiccommon e.g. (viadependent and stochastic are ),...,();,...,( profilesrisk
)|(.~)();|(.~
1,2,...,,...,2,1cell,risk th-,)(
)()(
)()1()()1(
)()1()()1(
)()()()(
)(
1
)()(
HG
HHGGCF
C
ftXPN
tJ jjtXZ
jt
jt
Jtt
Jtttt
Jttt
Jttt
jt
js
jt
jt
jtN
s
js
jt
ψλ
ψψλλ
ψψλλ
ψλ
=
•
==•
==∑==
ψλ
ψλ
Model fit via MCMC, Peters, Shevchenko and Wüthrich (2009)
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Example: dependence between frequency profiles;dependence between severity profiles
.,between copula)( ;,between copula)(
.parameter copulaGaussian vs),(n correlatiorank sSpearman'(2)(1)(2)(1)
)2()1(S
μμλλ
ρρ
−∗−o
ZZ
(2009) alet Peterscopula; via ,, between dependence
)1,1(~ and )1.0,2(~),10,4(~
),(~)(),(~
)( and )(
)()()(
)()()(
)()()()()(
)(
1
)()()(
1
)()(
jt
jt
jt
jt
jt
jt
jt
jt
js
jt
jt
jtN
s
js
jt
itN
s
is
it
GammaNGamma
LNtXPoissonN
tXZtXZ
σμλ
σμλ
σμλ
•
•
•
== ∑∑==
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Parameter Risk (uncertainty of parameters)
∞=>≥>
∝•−=
−•
−•
−×=•
=−=•
−=
+
+
+
++
=
∫
∑
]|[ then,0]1 Pr[i.e. ,0 with)|( and ),(~ if e.g.
severities edheavy tailfor ondistributi predictive mean infinte
densityposterior the),()|()|(]ˆˆ[
MLEestimator,point a is ˆ);ˆ|(ofquantile999.0ˆ)|(ofquantile999.0ˆ
ondistributipredictive full)|()|()|(
,...,1 nsobservatiopast ),(
; year in loss)(
1
999.0999.0
1999.0
1999.0
11
1
YY
θθYYθ
θθ
Y
θYθθY
NXY
M
i
BM
MB
MM
tN
iit
ZEGPDX
hQQEbias
ZgQ
ZfQ
dZgZf
Mt
ttXZ
ξξξπβξ
ππ
π
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Parameter Risk (uncertainty of parameters)
0%
50%
100%
150%
200%
250%
5 10 15 20 25 30 35 40Year
% b
ias
LognormalPareto
Relative bias in the 0.999 quantile estimator induced by the parameter uncertainty vs number of observation years. (Lognormal) - losses were simulated from Poisson(10) and LN(1,2). (Pareto) – losses were simulated from Poisson(10) and Pareto(2) with L=1.
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Research topics for collaboration
• Combining different data sources: internal & external data with expert opinions (credibility theory, Bayesian techniques)
• Dependence between risks: copula methods, structural models• Expert elicitation (Bayesian networks) • Extreme Value Models (Modelling distribution tail)• Efficient Markov Chain Monte Carlo, ABC MCMC, Monte Carlo, finite
element/finite difference methods • State-space models (Kalman/particle filters/SMC)
• Pattern recognition – trading strategies • Stochastic/local volatility models• Modelling operational/market/credit risks/insurance risks• Pricing exotic options• Modelling commodities, interest rates, • Portfolio asset allocation
Thank youCSIRO Mathematical & Information SciencesPavel ShevchenkoEmail: [email protected]/Pavel.Shevchenkowww.csiro.au/people/Pavel.Shevchenko.html
Remember that all models are wrong; the practical question is how wrong do they have to be to not be useful.
George Box and Norman Draper (1987)
CSIRO Mathematical & Information Sciences www.cmis.csiro.au
Recent journal publications
• X. Luo and P.V. Shevchenko (2009). The t copula with Multiple Parameters of Degrees of Freedom: Bivariate Characteristics and Application to Risk Management. Quantitative Finance.
• G. W. Peters, P. V. Shevchenko and M. V. Wüthrich (2009). Model uncertainty in claims reserving within Tweedie's compound Poisson models. ASTIN Bulletin 39(1), 1-33.
• P. V. Shevchenko (2009). Implementing loss distribution approach for operational risk. Applied Stochastic Models in Business and Industry.
• Peters, G., P. Shevchenko and M.Wuthrich (2009). Dynamic operational risk: modellingdependence and combining different data sources of information. The Journal of Operational Risk 4(2), 69-104.
• Shevchenko, P. (2008). Estimation of Operational Risk Capital Charge under Parameter Uncertainty. The Journal of Operational Risk 3(1), 51-63.
• Luo, X., P. V. Shevchenko and J. Donnelly (2007). Addressing Impact of Truncation and Parameter Uncertainty on Operational Risk Estimates. The Journal of Operational Risk 2(4), 3-26.
• D. D. Lambrigger, P.V. Shevchenko and M. V. Wüthrich (2007). The Quantification of Operational Risk using Internal Data, Relevant External Data and Expert Opinions. The Journal of Operational Risk 2(3), 3-27.
• Bühlmann,H., P. Shevchenko and M. Wüthrich. A “Toy” Model for Operational Risk Quantification using Credibility Theory. The Journal of Operational Risk 2(1), 3-19, 2007.
• Shevchenko, P. and M. Wüthrich (2006). Structural Modelling of Operational Risk using Bayesian Inference: combining loss data with expert opinions. The Journal of Operational Risk1(3), 3-26.