quantitative modeling of operational risk: between g-and-h
Post on 04-Jun-2022
3 Views
Preview:
TRANSCRIPT
Basel II LDA g-and-h Aggregation Conclusion and References
Quantitative Modeling of Operational Risk:Between g-and-h and EVT
Paul Embrechts Matthias Degen Dominik Lambrigger
ETH Zurich(www.math.ethz.ch/∼embrechts)
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Outline
Basel II
LDA
g-and-h
Aggregation
Conclusion and References
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
What is Basel II?
• 1988 Basel I Accord on Banking Supervision
- mainly CR- minimum risk capital (MRC) ≥ 8% of risk weighted assets
(Cooke Ratio)
• 1993 Birth of VaR
- “G-30 Report” addressing incorporation of off-balance sheetproducts (first time “VaR” appears)
- need for proper RM of these products
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
• 1996 Amendment to Basel I
- standardized model for MR- internal models allowed- legal implementation in 2000
• 2001 Initiation of consultative process for Basel II
- refined CR-approaches, IRB-models- consideration of new risk class: OR- implementation 2007+
I note Solvency I & II
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Risk Components (Basel II)
• Credit Risk
• Market Risk
• Operational Risk
• Business Risk
Operational Risk: The risk ofloss resulting from inadequateor failed internal processes,people and systems or fromexternal events. Including legalrisk, but excluding strategicand reputational risk.
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Risk Components (Basel II)
• Credit Risk
• Market Risk
• Operational Risk
• Business Risk
Operational Risk: The risk ofloss resulting from inadequateor failed internal processes,people and systems or fromexternal events. Including legalrisk, but excluding strategicand reputational risk.
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Some examples
• 1995: Nick Leeson/Barings Bank, £1.3b
• 2001: September 11
• 2001: Enron (largest US bankruptcy so far)
• ”Fat finger” errors
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Loss Distribution Approach (LDA)Loss Distribution Approach (LDA)
BL8
rrr
rrr
r r r r r r
BLi
BL1
RT1 RTk RT7
LT+1
LT+1i,k
���������������������'&
$%
type 2
1992 1994 1996 1998 2000 2002
05
1015
20
c©2006 (Embrechts & Neslehova) 13
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Basel II - Guidelines
• Risk measure: VaR
• Time horizon: 1 year
• Level: 99.9% (1 in 1000 year event!)
I Otherwise: Full methodological freedom (within LDA)
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
The Main LDA-Steps towards a Total Capital Charge
• Estimation of marginal VaR:
• Additional Aggregation:
• Diversification:
VaR1
α, . . . , VaRd
α
VaR+
α =d∑
k=1
VaRk
α
VaRrealα
?< VaR
+
α
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Reasonable Severity Distribution∗
• Good statistical fit of the data
• Loss distribution with realistic capital estimates
• Well specified: Are the characteristics of the fitted datasimilar to the loss data and logically consistent?
• Flexible: How well is the method able to reasonablyaccomodate a wide variety of empirical loss data?
• Simple: Is the method easy to apply in practice?
∗see Dutta and Perry (2006)
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Loss Distribution
EVT
• Moscadelli (2004):
- reasonable capital estimates(LDCE 2002)
- infinite mean models occur
• Well established theory:Peaks Over Threshold (POT)
• No specific underlying df
g-and-h
• Dutta and Perry (2006):
- EVT fails, propose g-and-h(LDCE 2004)
- finite mean g-and-h models
• No standard framework (yet)
• Specific parametric model
I Careful look at the g-and-h approach
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
g-and-h: Basic Properties
DefinitionLet Z ∼ N (0, 1) be a standard normal rv. A rv X is said to have ag-and-h distribution with parameters a, b, g , h ∈ R, if X satisfies
X = k(Z ) = a + begZ − 1
gehZ2/2
I g governs skewness
I h governs heavy-tailedness
I Distributional properties of F ∼ g-and-h?
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Theorem 1
Suppose F ∼ g-and-h, then:
• For g , h > 0, we have F ∈ RV−1/h, i.e. F (x) = x−1/hL(x)with L ∈ SV .
• For h = 0 and g > 0, we have F ∈ S\RV , where S denotesthe class of subexponential dfs.
I Well-known theory of (1st and 2nd order!) regular variation
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Theorem 2
The slowly varying function L asymptotically behaves like
exp(√
log x)
√log x
, x →∞.
I Difficult type of slowly varying function
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Pickands-Balkema-de Haan Theorem
First order property:
limu↑x0
supx∈(0,x0−u)
∣∣Fu(x)− Gξ,β(u)(x)∣∣︸ ︷︷ ︸
=:d(u)
= 0
• Fu(x) = P(X − u ≤ x |X > u): excess df
• Gξ,β(u): generalized Pareto distribution (GPD)
• x0 ≤ ∞: upper endpoint
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Pickands-Balkema-de Haan Theorem (continued)
• Theory: Under weak conditions d(u) converges to 0.(Maximum Domain of Attraction)
• Practice: No information on goodness of approximation.
Second order property:
I How fast does d(u) converge to 0?
I Determined by L ∈ SV
I Highly relevant for practical applications
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Pickands-Balkema-de Haan Theorem (continued)
• Theory: Under weak conditions d(u) converges to 0.(Maximum Domain of Attraction)
• Practice: No information on goodness of approximation.
Second order property:
I How fast does d(u) converge to 0?
I Determined by L ∈ SV
I Highly relevant for practical applications
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Rate of convergence to the GPD for different distributions, as afunction of the threshold u
Distribution Parameters F d(u)
Exponential(λ) λ > 0 e−λx 0Pareto(α) α > 0 x−α 0
Double exp. parent e−ex
O(e−u)Student t ν > 0 tν(x) O( 1
u2 )
Normal(0, 1) Φ(x) O( 1u2 )
Weibull(τ, c) τ ∈ R+\ {1} , c > 0 e−(cx)τ O( 1uτ )
Lognormal(µ, σ) µ ∈ R, σ > 0 Φ( log x−µσ
) O( 1log u
)
Loggamma(γ, α) α > 0, γ 6= 1 Γα,γ(x) O( 1log u
)
g-and-h g , h > 0 Φ(k−1(x)) O( 1√log u
)
I If data are well modeled by a g-and-h, EVT-based estimationconverges very slowly
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Rate of convergence to the GPD for different distributions, as afunction of the threshold u
Distribution Parameters F d(u)
Exponential(λ) λ > 0 e−λx 0Pareto(α) α > 0 x−α 0
Double exp. parent e−ex
O(e−u)Student t ν > 0 tν(x) O( 1
u2 )
Normal(0, 1) Φ(x) O( 1u2 )
Weibull(τ, c) τ ∈ R+\ {1} , c > 0 e−(cx)τ O( 1uτ )
Lognormal(µ, σ) µ ∈ R, σ > 0 Φ( log x−µσ
) O( 1log u
)
Loggamma(γ, α) α > 0, γ 6= 1 Γα,γ(x) O( 1log u
)
g-and-h g , h > 0 Φ(k−1(x)) O( 1√log u
)
I If data are well modeled by a g-and-h, EVT-based estimationconverges very slowly
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Rate of convergence to the GPD for different distributions, as afunction of the threshold u
Distribution Parameters F d(u)
Exponential(λ) λ > 0 e−λx 0Pareto(α) α > 0 x−α 0
Double exp. parent e−ex
O(e−u)Student t ν > 0 tν(x) O( 1
u2 )
Normal(0, 1) Φ(x) O( 1u2 )
Weibull(τ, c) τ ∈ R+\ {1} , c > 0 e−(cx)τ O( 1uτ )
Lognormal(µ, σ) µ ∈ R, σ > 0 Φ( log x−µσ
) O( 1log u
)
Loggamma(γ, α) α > 0, γ 6= 1 Γα,γ(x) O( 1log u
)
g-and-h g , h > 0 Φ(k−1(x)) O( 1√log u
)
I If data are well modeled by a g-and-h, EVT-based estimationconverges very slowly
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Tail Index Estimation
• Xiiid∼ F ∈ RV−1/ξ
• Hk,n :=1
k
k∑j=1
(log Xn−j+1,n − log Xn−k,n) (Hill estimator)
• Hk,n very sensitive to choice of threshold k
• “optimal” k often s.t. AMSE of Hk,n minimal
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Tail Index Estimation - Simulation Study
heavy-tailednesssk
ewn
ess
g\h 0.1 0.2 0.5 0.7 1 2
0.1 142 82 33 23 18 110.2 165 97 42 32 25 200.5 224 132 49 38 27 190.7 307 170 63 44 29 20
1 369 218 86 58 36 262 696 385 151 108 74 313 1097 613 243 163 115 54
Empirical SRMSE (in %) of the Hill estimator hHillkopt of h for
g-and-h data for different parameter values of g and h
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Hill Plotsg
hsmall
2 119 275 431 587 743 899 1075 1270 1465 1660 1855 2050 2245 2440 2635
0.6
0.8
1.0
1.2
1.4
1.6
214000 1750 778 513 388 303 248 211 183 163 148
Order Statistics
xi (
CI,
p =
0.95
)
Threshold
I Hill plot works fine(g = 0.1, h = 1)
g
hlarge
15 73 141 219 297 375 453 531 609 687 765 843 921 999 1086 1183 1280 1377
0.6
0.8
1.0
1.2
1.4
1.6
1.8
21900000 1250000 551000 332000 232000 176000 137000 113000
Order Statistics
xi (
CI,
p =
0.95
)
Threshold
I Hill plot misleadingly indicatesinfinite mean model!
(g = 4, h = 0.2)
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Aggregation
Dutta-Perry: “We have not mathematically verified thesubadditivity property for g-and-h, but in all caseswe have observed empirically that enterprise levelcapital is less than or equal to the sum of thecapitals from business lines or event types.”
Question: COpRiskα < VaR
+
αdef=
d∑k=1
VaRk
α ?
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Subadditivity of VaR typically fails for:
• Skewness
• Heavy-Tailedness
• Dependence
RemarkIn the space Lp, 0 < p < 1, there exist no convex open sets otherthan the empty set and Lp itself.
I No reasonable risk measures existI Diversification goes the wrong way
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
g = 2.4, h = 0.2 (iid case)
99.4%
Proposition [Danıelsson et al.]
Suppose that the non-degenerate vector (X1,X2) is regularlyvarying with extreme value index ξ < 1. Then VaRα is subadditivefor α sufficiently large.
alpha
delta
0.90 0.92 0.94 0.96 0.98 1.00
-20
-10
010
2030
40
diversification benefit:delta = VaRα(X1) + VaRα(X2)− VaRα(X1 + X2)
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
RemarkThis proposition is only an asymptotic statement - It does notguarantee subadditivity for a broad range of high quantiles
I of no use for practical assessment of subadditivityI Basel II: 1-year 99.9% VaR - which choices of g and h yieldsubadditive models?
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Subadditivity of VaR at 99.9%
skewness
hea
vy-t
aile
dn
ess
g
h
1.8 1.9 2.0 2.1 2.2 2.3 2.4
0.1
0.2
0.3
0.4
0.5
-800
-400
-200
200
0
• Entire parameter rectangle within subadditivity range
• Small changes of parameters ⇒ superadditivity
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
What happens when we go deeper in the data?
• VaR-estimation at 99.9% and higher: difficult!
• Estimate at lower level (90%, say) and scale: how?
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Subadditivity of VaR at 99%
skewness
hea
vy-t
aile
dn
ess
g
h
1.8 1.9 2.0 2.1 2.2 2.3 2.4
0.1
0.2
0.3
0.4
0.5
-25
-50
-100
-150
-200
0
• Substantial fraction of parameter rectangle switched regime
• Far from diversification!
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Dependence matters
Gauss-Copula
alpha
delta
0.90 0.92 0.94 0.96 0.98 1.00
-40
-20
020
40
corr=0corr=0.5corr=0.7
Increasing correlation ⇒ superadditivity range extends
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
Conclusion
• Very slow convergence of g-and-h excess df to the GPD wheng , h > 0
• Optimal threshold selection for an EVT based POT approachbecomes very difficult (unreliable risk capital estimates)
• Small changes of g and/or h may lead VaR to switch(sub-/superadditivity) regime
• g-and-h is subexponential → one claim causes ruin
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
Basel II LDA g-and-h Aggregation Conclusion and References
References
Degen, M., Embrechts, P. and Lambrigger, D. (2006) Thequantitative modeling of operational risk: between g-and-hand EVT. ASTIN Bulletin 2007, to appear.
Dutta, K. and Perry, J. (2006) A tale of tails: an empiricalanalysis of loss distribution models for estimating operationalrisk capital. Federal Reserve Bank of Boston, Working PaperNo 06-13.
Moscadelli, M. (2004) The modelling of operational risk:experiences with the analysis of the data collected by theBasel Committee. Bank of Italy, Working Paper No 517.
P. Embrechts ETH Zurich
Quantitative Modeling of Operational Risk
top related