101
GHICA - Risk Analysis with GH Distributions and Independent Components Ying Chen Wolfgang H¨ ardle Vladimir Spokoiny Institut f¨ ur Statistik and ¨ Okonometrie CASE - Center for Applied Statistics and Economics Humboldt-Universit¨ at zu Berlin Weierstraß-Institut ur Angewandte Analysis und Stochastik http://ise.wiwi.hu-berlin.de http://www.case.hu-berlin.de http://www.wias-berlin.de Weierstraß-Institut für Angewandte Analysis und Stochastik

GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

  • Upload
    others

  • View
    5

  • Download
    0

Embed Size (px)

Citation preview

Page 1: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

GHICA - Risk Analysis with GH Distributions andIndependent Components

Ying Chen

Wolfgang Hardle

Vladimir Spokoiny

Institut fur Statistik and OkonometrieCASE - Center for Applied Statistics and EconomicsHumboldt-Universitat zu Berlin

Weierstraß-Institutfur Angewandte Analysis und Stochastik

http://ise.wiwi.hu-berlin.dehttp://www.case.hu-berlin.dehttp://www.wias-berlin.de

W e ie rstra ß -In stitu t fü r A n g e w a n d te A n a ly s is u n d S to ch a stik

Page 2: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-1

Value at Risk (VaR) calculation

Rt = b>t xt = b>t Σ1/2t εt

Rt the portfolio returnbt the portfolio weightsxt ∈ Rd the individual returns with cov Σt

εt the stochastic term.

VaR at level 100a%: VaRa,t = F−1a,t (Rt),

F−1t equals the inverse of cdf of Rt .

Critical point: pdf of xt .

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 3: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-2

Value at Risk (VaR) calculation

Rt = b>t xt = b>t Σ1/2t εt

Rt the portfolio returnbt the portfolio weightsxt ∈ Rd the individual returns with cov Σt

εt the stochastic term.

RiskMetrics

εt ∼ N(0, Id)Σt = $Σt−1 + (1−$)xt−1x

>t−1 (Exponential Moving Average)

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 4: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-3

Value at Risk (VaR) calculation

Rt = b>t xt = b>t Σ1/2t εt

Rt the portfolio returnbt the portfolio weightsxt ∈ Rd the individual returns with cov Σt

εt the stochastic term.

T-deGARCH

εt ∼ t(df )Σt = $ + α1Σt−1 + β1xt−1x

>t−1 (GARCH(1,1))

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 5: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-4

A univariate story

Chen, Hardle and Jeong (2005), GHADA:Generalized Hyperbolic distribution + ADAptive volatilityThe Gaussian density (left) and GH density (right) fits of the devolatilized returns of

DEM/USD rates from 1979-12-01 to 1994-04-01 (black: nonparametric kernel density,

red: normal density, blue: GH density).

Gaussian density

-4 -2 0 2 4

X

00.

10.

20.

30.

40.

5

Y

GH(1,1.74,-0.02,0.78,0.01)

-4 -2 0 2 4

X

00.

10.

20.

30.

4

Y

GARCH(1,1) + t(15) density

-4 -2 0 2 4

X

00.

10.

20.

30.

40.

5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 6: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-5

A univariate story

Chen, Hardle and Jeong (2005), GHADA:Generalized Hyperbolic distribution + ADAptive volatilityThe t(15) density (right) and GH density (left) fits of the devolatilized returns of

DEM/USD rates (black: nonparametric kernel density, green: t(15) density, blue: GH

density). GARCH(1,1) process: σ2t = 1.65e − 06 + 0.07x2

t−1 + 0.89σ2t−1.

Gaussian density

-4 -2 0 2 4

X

00.

10.

20.

30.

40.

5

Y

GH(1,1.74,-0.02,0.78,0.01)

-4 -2 0 2 4

X

00.

10.

20.

30.

4

Y

GARCH(1,1) + t(15) density

-4 -2 0 2 4

X

00.

10.

20.

30.

40.

5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 7: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-6

Excursion to GHADA

X ∼ GH with density:

fGH(x ;λ, α, β, δ, µ) =(ι/δ)λ

√2πKλ(δι)

Kλ−1/2

{α√

δ2 + (x − µ)2}

{√δ2 + (x − µ)2

/α}1/2−λ

· eβ(x−µ)

Where ι2 = α2 − β2, Kλ(·) is the modified Bessel function of thethird kind with index λ: Kλ(x) = 1

2

∫∞0 yλ−1exp{− x

2 (y + y−1)} dy .

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 8: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-7

Subclass of GH distribution

The parameters (µ, δ, β, α)> can be interpreted as trend, riskiness,asymmetry and the likeliness of extreme events.Hyperbolic (HYP) distributions: λ = 1,

fHYP(x ;α, β, δ, µ) =ι

2αδK1(δι)e{−α

√δ2+(x−µ)2+β(x−µ)}, (1)

where x , µ ∈ R, 0 ≤ δ and |β| < α.Normal-inverse Gaussian (NIG) distributions: λ = −1/2,

fNIG (x ;α, β, δ, µ) =αδ

π

K1

{α√

δ2 + (x − µ)2}

√δ2 + (x − µ)2

e{δι+β(x−µ)}, (2)

where x , µ ∈ R, 0 < δ and |β| ≤ α.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 9: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-8

Adaptive Volatility Estimation

Volatility (and also θt = Cσγt ) is locally time homogeneous in a

short time interval [τ −m, τ), thus we can estimate:

σ2τ =

1

]{I}∑t∈I

R2t .

Time homogeneous: variables are constant and time independent.

1 2 3 τ −m τ − 1 T

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 10: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-9

Homogeneity test

TheoremIf R1, ...,Rτ obey the heteroscedastic model and volatilities σt satisfythe condition b ≤ σ2

t ≤ bB with b,B > 0, given a large value η:

P{|θI − θτ | >∆I (1 +ηsγ√]{I}

) + ηvI}

≤ 4√

eη(1 + logB) exp

{− η2

2aγ(1+ηsγ√]{I}

)2

}.

where ∆I measures the (local) bias: ∆2I = 1

]{I}∑

t∈I (θt − θτ )2 and

vI the volatility of θI .

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 11: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-10

Homogeneity test

Under homogeneity |θI − θτ | is bounded by ηvI provided that η issufficiently large.

|θI − θτ | < ηvI

|θI\J − θJ | < η(vI\J + vJ) = η′(

√θ2J/]{J}+

√θ2I\J/]{I\J}).

' $� �τ −m τ − 2/3m τ − 1/3m

J

J

τ − 1

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 12: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-11

Extension to the high-dimensional case

Independent component analysis (ICA)

Given xt = {x1t , · · · , xdt}>, find ICs yt :

yt = Wxt ,

where W is a d × d nonsingular matrix:

xt = W−1yt = Ayt

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 13: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-12

ICA example

IC1t : generalized hyperbolic variable GH(−0.5, 1.02,−0.11, 1, 0)IC2t : GH(−0.5, 1.29,−0.51, 1, 0)IC3t : GH(−0.5, 1.15,−0.35, 1, 0)

A = W−1 =

9 11 0.5−1 −8 212 −5 1

, xt = A

IC1t

IC2t

IC3t

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 14: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-13

ICA example

The Mahalanobis transformation:

cov(x)−1/2 =

0.19 0.22 −0.080.22 0.45 −0.14

−0.08 −0.14 0.12

6= W =

0.01 −0.04 0.080.07 0.01 −0.050.29 0.52 −0.18

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 15: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-14

Time plots of 3 independent GH variables (top), the linear transformation (middle)

and the estimated ICs (bottom).IC1: GH(-0.5,1.02,-0.11,1,0)

-4.00

0.00

4.00

8.00

250 500 750 1000

x1 = 9*y1 + 11*y2 +0.5*y3

-15.000.00

15.0030.00

250 500 750 1000

y1

-4.00

0.00

4.00

8.00

250 500 750 1000

IC2: GH(-0.5,1.66,-0.89,1,0)

-4.00

0.00

4.00

8.00

250 500 750 1000

x2 =-1*y1 - 8*y2 +2*y3

-15.000.00

15.0030.00

250 500 750 1000

y2

-4.00

0.00

4.00

8.00

250 500 750 1000

IC3: GH(-0.5,1.72,-0.95,1,0)

-4.00

0.00

4.00

8.00

250 500 750 1000

x3 = 12*y1 -5*y2 + y3

-15.000.00

15.0030.00

250 500 750 1000

y3

-4.00

0.00

4.00

8.00

250 500 750 1000

GHICAsim.xpl

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 16: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-15

What will ICA bring us?

The portfolio return Rt can be calculated as:

Rt = b>t xt (3)

= b>t W−1yt (4)

= b>t W−1D1/2t εt (5)

where covariance of yt : Dt = diag(σ21t , · · · , σ2

dt) and εt i.i.d. shocks.Marginal pdf fyj and volatility σjt of IC yj , j = 1, · · · , d can beestimated univariately.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 17: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-16

What will ICA bring us?

The portfolio return Rt can be calculated as:

Rt = b>t xt (3)

= b>t W−1yt (4)

= b>t W−1D1/2t εt (5)

� Equation (4) is ICA that gives us approximately ICs.

� Dt and distributional parameters of εt in (5) can be estimatedapplying GHADA.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 18: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Motivation 1-17

VaR estimation procedure: GHICA

1. ICA to get independent components.

2. Apply the GHADA to estimate the local constant vola and fitthe marginal pdf of εt .

3. Determine VaR at level a via MC (d samples, M observationsand N repetitions) .

VaRa,t =1

N

N∑n=1

F−1a,t (R

(n)t )

=1

N

N∑n=1

F−1a,t {b>t W−1D

1/2t ε

(n)t }

4. Perform backtesting.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 19: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Outline

1. Motivation: ICA + GHADA = GHICA X

2. ICA: properties and estimation

3. Simulation study

4. Empirical study

5. Conclusion

Page 20: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-1

Definition

ICA model: y1t...

ydt

=

w11 · · · w1d

· · · · ·wd1 · · · wdd

x1t

...xdt

yt = Wxt

equivalently xt = Ayt

where xt are d-dimensional observations, yt are ICs and W thenonsingular linear transforming matrix: W−1 = A.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 21: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-2

Properties of ICA

Scale identification: the scales of the ICs are not identifiable sinceboth yt and W are unknown.

x1t =d∑

j=1

a1jyjt =d∑

j=1

{ 1

cja1j}{cjyjt}

where cj are constants.Hence: prewhiten xt by the Mahalanobis transformation cov(x)−1/2

and assume that each IC has unit variance: E[y2j ] = 1. The matrix

W is then an orthogonal matrix.From now on xt is prewhitened!

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 22: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-3

Properties of ICA

Order identification: the order of the ICs is undetermined.

xt = Ayt = AP−1Pyt

where P is a permutation matrix and Pyt are the original ICs but ina different order.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 23: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-4

Properties of ICA

ICs are necessarily non-GaussianConsider two prewhitened Gaussian ICs y1 and y2 with pdf:

f (y1, y2) = |2πI|−12 exp(−y2

1 + y22

2) =

1

2πexp(−||y ||

2

2)

where ||y || is the norm of the vector (y1, y2)>.

The joint density of the observation x1 and x2 is given by:

f (x1, x2) = |2πI|−12 exp(−||Wx ||2

2)|detW | = 1

2πexp(−||x ||

2

2).

Since A is an orthogonal matrix after prewhitening.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 24: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-5

How to find the ICs?

CLT: Let (Y1, · · · ,Yd) be a set of d independent random variablesand each has an arbitrary probability distribution, their weightedsum has a limiting cdf that approaches a normal distribution.

yt = W xt = WW−1yt

yjt = w>j xt = w>

j W−1yt = (W−1>wj)>yt

LHS: yjt is expected to be identical to one IC (one variable)RHS: weighted sum of ICs (sum)

Find the maximal nongaussianity of yjt .

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 25: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-6

How to find the ICs?

Non-Gaussianity measurement: Negentropy

J(y) = H(ygauss)− H(y)

entropy: H(y) = −∫

fy (η) log fy (η)dη.

Here ygauss is a Gaussian variable with cov(ygauss) = cov(y).Recall that for given variance Gaussian variable has maximal entropy:

H(ygauss) =1

2log |detΣ|+ d

2(1 + log 2π)

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 26: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-7

How to find the ICs?

Remark: Compared to kurtosis and excess kurtosis, negentropy isless sensitive to outliers.

Negentropy calculation however requires the knowledge of the pdfof yt .

Therefore, Hyvarinen (1998) proposes approximations of negentropythat are computationally feasible:

J(y) ≈ c1[E{G1(y)}]2 + c2{E[G2(y)]− E[G2{N(0, 1)}]}2.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 27: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-8

How to find the ICs?

Negentropy approximationApproximation a: c1 = 36/(8

√3− 9) and ca

2 = 1/(2− 6/π)

J(y) ≈ c1[E{y exp(−y2/2)}]2 + ca2 [E{exp(−y2/2)} −

√1/2]2

G a1 (y) = y exp(−y2/2)

G a2 (y) = exp(−y2/2)

Approximation b: c1 = 36/(8√

3− 9) and cb2 = 24/(16

√3− 27/π)

J(y) ≈ c1[E{y exp(−y2/2)}]2 + cb2 [E{|y |} −

√2/π]2

Gb1 (y) = y exp(−y2/2)

Gb2 (y) = |y |

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 28: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-9

How to find the ICs?

Comparison of the true negentropy (black) and its approximations (a: red, b:

blue) of simulated Gaussian mixture variable: pN(0, 1) + (1 − p)N(1, 4) for p ∈[0, 1]. GHICAnegentropyapp.xpl

Negentropy comparison

0 0.5 1

p

05

10

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 29: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-10

Negentropy approximations and FastICA

Assumption on Negentropy approximation: variable is symmetric.In this sense, G1 as an odd function disappears.

J(y) ≈ C{E[G (y)]− E[G{N(0, 1)}]}2.

G (y) =1

slog cosh(sy), 1 ≤ s ≤ 2 (6)

g(y)def= G ′(y) = tanh(sy) (7)

g ′(y) = s{1− tanh2(sy)} (8)

very often, s = 1 is taken in this approximation.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 30: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-11

Negentropy approximations and FastICA

Assumption on Negentropy approximation: variable is symmetric.In this sense, G1 as an odd function disappears.

J(y) ≈ C{E[G (y)]− E[G{N(0, 1)}]}2.

G (y) = − exp(−y2/2) (9)

g(y)def= G ′(y) = y exp(−y2/2) (10)

g ′(y) = (1− y2) exp(−y2/2) (11)

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 31: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-12

Negentropy approximations and FastICA

Assumption on Negentropy approximation: variable is symmetric.In this sense, G1 as an odd function disappears.

J(y) ≈ C{E[G (y)]− E[G{N(0, 1)}]}2.

G (y) =1

4y4 (12)

g(y)def= G ′(y) = y3 (13)

g ′(y) = 3y2 (14)

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 32: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-13

FastICA

Objective function

{E{G (WX )} − E[G{N(0, 1)}]}E{Xg(WX )} = 0 (15)

A fast gradient method can be formulated under the constraintW>W = Id :

E{Xg(WX )}+ χW = 0 (16)

The iteration of wj with respect to yj :

w(n+1)j = E[Xg(w

(n)j X )− E{g ′(w (n)

j X )}w (n)j ] (17)

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 33: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-14

FastICA

Algorithm

1. Choose an initial vector wj of unit norm, W = (w1, · · · ,wd)>.

2. Let w(n)j = E[g(w

(n−1)j x)x ]− E[g ′(w

(n−1)j x)]w

(n−1)j . In

practice, the sample mean is used to calculate E[·].3. Orthogonalization (decorrelated):

w(n)j = w

(n)j −

∑k 6=j(w

(n)>j wk)wk .

4. Normalization: w(n)j = w

(n)j /||w (n)

j ||.

5. If not converged, i.e. ||w (n)j − w

(n−1)j || 6= 0, go back to 2.

6. Set j = j + 1. For j ≤ d , go back to step 1.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 34: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

ICA 2-15

VaR estimation procedure: GHICA

1. ICA to get independent components.

2. Apply the GHADA to estimate the local constant vola and fitthe marginal pdf of εt .

3. Determine VaR at level a via MC (d samples, M observationsand N repetitions) .

VaRa,t =1

N

N∑n=1

F−1a,t (R

(n)t )

=1

N

N∑n=1

F−1a,t {b>t W−1D

1/2t ε

(n)t }

4. Perform backtesting.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 35: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Simulation study 3-1

Simulation study on 50 NIG RVs

Simulation design

� NIG distributional parameters:λ = −0.5, α ∼ U[1, 2],

δ = 1, β = arg{σ2(y) = 1√α2−β2

α2

α2−β2 = 1},

µ = 0.

� d = 50 samples, each has T = 1000 observations

� An orthogonal matrix W−1 is achieved via the Jordandecomposition of a square matrix (elements ∼ N(0, 1))

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 36: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Simulation study 3-2

Simulation study on 50 NIG RVs

Simulation procedure

� Generate 50 independent NIG random variables yt and W−1

� Calculate the input of ICA: xt = W−1yt

� Estimate yt and their distributional parameters (αj , βj , δj , µj)for j = 1, · · · , 50

Simulation goalCompare the pdfs of the generated and the estimated ICs. Thelargest mean absolute error (MAE) is 0.054.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 37: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Simulation study 3-3

Simulation study on 50 NIG RVs

Graphical comparisons of the pdfs of the estimated IC and the trueIC1 (dotted line). GHICAsimpdfcomp.xpl

ICest1

-5 0 5

X

00.

10.

20.

30.

40.

50.

6

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 38: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Simulation study 3-4

Simulation study on 50 NIG RVs

Graphical comparisons of the pdfs of the estimated ICs and the trueICs (dotted line). GHICAsimpdfcomp.xpl

ICest1

-5 0 5

X

00.

20.

40.

6

Y

ICest11

-5 0 5

X

00.

20.

40.

6

Y

ICest21

-5 0 5

X

00.

20.

40.

6

Y

ICest31

-5 0 5

X

00.

5

Y

ICest41

-5 0 5

X

00.

5

Y

ICest2

-5 0 5

X

00.

20.

40.

6

Y

ICest12

-5 0 5

X

00.

20.

40.

6

Y

ICest22

-5 0 5

X

00.

20.

40.

6

Y

ICest32

-5 0 5

X

00.

20.

40.

6

Y

ICest42

-5 0 5

X

00.

5

Y

ICest3

-5 0 5

X

00.

20.

40.

6

Y

ICest13

-5 0 5

X

00.

20.

40.

60.

8

Y

ICest23

-5 0 5

X

00.

20.

40.

6

Y

ICest33

-5 0 5

X

00.

20.

40.

6

Y

ICest43

-5 0 5

X

00.

5

Y

ICest4

-5 0 5

X0

0.2

0.4

0.6

Y

ICest14

-5 0 5

X

00.

20.

40.

6

Y

ICest24

-5 0 5

X

00.

20.

40.

6

Y

ICest34

-5 0 5

X

00.

20.

40.

6

Y

ICest44

-5 0 5

X

00.

5

Y

ICest5

-5 0 5

X

00.

5

Y

ICest15

-5 0 5

X

00.

20.

40.

6

Y

ICest25

-5 0 5

X

00.

20.

40.

6

Y

ICest35

-5 0 5

X

00.

5

Y

ICest45

-5 0 5

X

00.

5

Y

ICest6

-5 0 5

X

00.

5

Y

ICest16

-5 0 5

X

00.

20.

40.

6

Y

ICest26

-5 0 5

X

00.

20.

40.

6

YICest36

-5 0 5

X

00.

20.

40.

6

Y

ICest46

-5 0 5

X

00.

5

Y

ICest7

-5 0 5

X

00.

20.

40.

6

Y

ICest17

-5 0 5

X

00.

5

Y

ICest27

-5 0 5

X

00.

20.

40.

60.

8

Y

ICest37

-5 0 5

X

00.

20.

40.

60.

8

Y

ICest47

-5 0 5

X

00.

20.

40.

6

Y

ICest8

-5 0 5

X

00.

20.

40.

6

Y

ICest18

-5 0 5

X

00.

20.

40.

6

Y

ICest28

-5 0 5

X

00.

20.

40.

6

Y

ICest38

-5 0 5

X

00.

20.

40.

6

Y

ICest48

-5 0 5

X

00.

5

Y

ICest9

-5 0 5

X

00.

20.

40.

6

Y

ICest19

-5 0 5

X

00.

20.

40.

6

Y

ICest29

-5 0 5

X

00.

20.

40.

6

Y

ICest39

-5 0 5

X

00.

5

Y

ICest49

-5 0 5

X

00.

51

Y

ICest10

-5 0 5

X

00.

20.

40.

6

Y

ICest20

-5 0 5

X

00.

20.

40.

6

Y

ICest30

-5 0 5

X

00.

5

Y

ICest40

-5 0 5

X

00.

20.

40.

6

Y

ICest50

-5 0 5

X

00.

51

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 39: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-1

Foreign exchange rate portfolio

Data: (DEM/USD, GBP/USD) daily rates from 1979-12-01 to1994-04-01 (3720 observations), available at FEDC.Trading strategies

b1 = (1, 1)>

b2 = (1, 2)>

b3 = (−1, 2)>

b4 = (−2, 1)>

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 40: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-2

Foreign exchange rate portfolio

ICA

W =

(207.93 −213.6377.72 73.29

)A =

(2.30e − 3 6.71e − 3

−2.44e − 3 6.53e − 0

)GH parameters estimation

IC1:HYP( 1.71, −0.17, 0.55, 0.13)NIG ( 1.22, −0.18, 1.11, 0.13)

IC2:HYP( 1.78, 0.03, 0.72, −0.02)NIG ( 1.37, 0.03, 1.28, −0.03)

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 41: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-3

Foreign exchange rate portfolio

Temporal dependence: ACF plots of the log returns: DEM/USDon the left and GBP/USD on the right.GHICAfxdescriptive.xpl

ACF plot of DEM/USD

0 5 10 15 20 25 30

lag

00.5

1

acf

ACF plot of GBP/USD

0 5 10 15 20 25 30

lag

00.5

1

acf

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 42: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-4

Foreign exchange rate portfolio

Temporal dependence: ACF plots of the 2 independent series: IC1left and IC2 right. GHICAfxdescriptive.xpl

ACF plot of IC1

0 5 10 15 20 25 30

lag

00.5

1

acf

ACF plot of IC2

0 5 10 15 20 25 30

lag

00.5

1

acf

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 43: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-5

Foreign exchange rate portfolio

Linear dependenceScatterplots of the FX returns (left) and ICs (right) of the FXportfolio. GHICAfxdescriptive.xpl

FX DEM/USD & GBP/USD

-4 -2 0 2

DEM/USD*E-2

-4-2

02

4

GB

P/U

SD*E

-2

FX IC1 & IC2

-5 0 5

IC1

-50

IC2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 44: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-6

Foreign exchange rate portfolio

Joint density: Comparison of the nonparametric joint density(black) of the returns of the exchange rates and the product (blue)of the HYP marginal densities of two ICs. The red surface is theGaussian fitting with the same covariance as the returns of the ex-change rates.GHICAfxjointdensitycomp.xpl

Joint pdf of FX returns

-0.03

-0.01 0.00

0.02 0.04-0.04

-0.03-0.01

0.00 0.02

1610.16

3220.31

4830.47

6440.63

8050.79

Joint pdf of ICs

-3.99

-1.86 0.27

2.39 4.52-8.61

-5.71-2.82

0.07 2.97

0.06

0.12

0.18

0.24

0.30

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 45: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-7

Foreign exchange rate portfolio

Adaptive vola: GHICAfxdescriptive.xpl

FX Adaptive Vola of IC1

5 10 15 20 25 30 35X*E2

12

3

Y

FX Adaptive Vola of IC2

5 10 15 20 25 30 35X*E2

0.5

11.

52

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 46: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-8

Foreign exchange rate portfolio

MC quantile estimation: Descriptive statistics of quantile esti-mates: b = (1, 1)>, MC simulation with M = 10000, N = 100,T = 1000. GHICAstatMCq.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.01 -0.02 0.04 -0.02 0.06 -0.02 0.08 -0.02 0.14 1 -0.01 0.02 -0.02 0.05 -0.02 0.06 -0.02 0.09 -0.03 0.14

2 -0.01 0.01 -0.01 0.04 -0.02 0.05 -0.02 0.08 -0.02 0.13 2 -0.01 0.01 -0.02 0.04 -0.02 0.06 -0.02 0.09 -0.02 0.15

· · · · · · · · · · · · · · · · · · · · · ·424 -0.03 0.07 -0.06 0.15 -0.07 0.21 -0.08 0.34 -0.09 0.50 423 -0.04 0.06 -0.07 0.15 -0.07 0.23 -0.08 0.33 -0.10 0.45

425 -0.03 0.06 -0.06 0.14 -0.07 0.22 -0.08 0.27 -0.09 0.44 425 -0.04 0.06 -0.07 0.14 -0.07 0.24 -0.08 0.28 -0.10 0.46

426 -0.03 0.04 -0.05 0.11 -0.06 0.18 -0.06 0.25 -0.08 0.38 426 -0.03 0.05 -0.06 0.12 -0.06 0.18 -0.07 0.27 -0.08 0.53

597 -0.03 0.06 -0.06 0.16 -0.06 0.22 -0.07 0.30 -0.09 0.54 607 -0.04 0.06 -0.06 0.15 -0.07 0.21 -0.08 0.32 -0.10 0.46

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.02 -0.02 0.04 -0.02 0.06 -0.02 0.09 -0.03 0.14 622 -0.03 0.07 -0.06 0.12 -0.07 0.18 -0.08 0.26 -0.09 0.48

1000 -0.01 0.02 -0.02 0.04 -0.02 0.07 -0.02 0.10 -0.03 0.16 1000 -0.01 0.02 -0.02 0.05 -0.02 0.07 -0.03 0.11 -0.03 0.18

Tab. 4: Descriptive statistics of quantile estimates: b = (1, 1)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.02 -0.03 0.06 -0.03 0.09 -0.03 0.13 -0.04 0.23 1 -0.01 0.02 -0.03 0.06 -0.03 0.09 -0.04 0.14 -0.04 0.25

2 -0.01 0.02 -0.02 0.05 -0.03 0.08 -0.03 0.11 -0.04 0.18 2 -0.01 0.02 -0.03 0.06 -0.03 0.09 -0.03 0.12 -0.04 0.19

· · · · · · · · · · · · · · · · · · · · · ·424 -0.05 0.10 -0.09 0.23 -0.10 0.31 -0.12 0.49 -0.14 0.74 325 -0.06 0.11 -0.10 0.22 -0.11 0.30 -0.13 0.39 -0.15 0.71

425 -0.05 0.10 -0.10 0.21 -0.10 0.33 -0.12 0.40 -0.14 0.66 423 -0.06 0.11 -0.10 0.25 -0.11 0.32 -0.12 0.46 -0.15 0.75

597 -0.05 0.09 -0.09 0.24 -0.10 0.33 -0.11 0.45 -0.13 0.81 424 -0.06 0.10 -0.10 0.22 -0.11 0.32 -0.13 0.45 -0.15 0.88

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.03 -0.03 0.07 -0.03 0.08 -0.04 0.13 -0.04 0.20 597 -0.05 0.09 -0.10 0.22 -0.10 0.34 -0.12 0.46 -0.14 0.71

1000 -0.02 0.03 -0.03 0.06 -0.03 0.10 -0.04 0.14 -0.05 0.26 1000 -0.02 0.03 -0.03 0.08 -0.04 0.11 -0.04 0.16 -0.05 0.27

Tab. 5: Descriptive statistics of quantile estimates: b = (1, 2)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

19

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 47: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-9

Foreign exchange rate portfolio

MC quantile estimation: Descriptive statistics of quantile esti-mates: b = (1, 1)>, MC simulation with M = 10000, N = 100,T = 1000. GHICAstatMCq.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.01 -0.02 0.04 -0.02 0.06 -0.02 0.08 -0.02 0.14 1 -0.01 0.02 -0.02 0.05 -0.02 0.06 -0.02 0.09 -0.03 0.14

2 -0.01 0.01 -0.01 0.04 -0.02 0.05 -0.02 0.08 -0.02 0.13 2 -0.01 0.01 -0.02 0.04 -0.02 0.06 -0.02 0.09 -0.02 0.15

· · · · · · · · · · · · · · · · · · · · · ·424 -0.03 0.07 -0.06 0.15 -0.07 0.21 -0.08 0.34 -0.09 0.50 423 -0.04 0.06 -0.07 0.15 -0.07 0.23 -0.08 0.33 -0.10 0.45

425 -0.03 0.06 -0.06 0.14 -0.07 0.22 -0.08 0.27 -0.09 0.44 425 -0.04 0.06 -0.07 0.14 -0.07 0.24 -0.08 0.28 -0.10 0.46

426 -0.03 0.04 -0.05 0.11 -0.06 0.18 -0.06 0.25 -0.08 0.38 426 -0.03 0.05 -0.06 0.12 -0.06 0.18 -0.07 0.27 -0.08 0.53

597 -0.03 0.06 -0.06 0.16 -0.06 0.22 -0.07 0.30 -0.09 0.54 607 -0.04 0.06 -0.06 0.15 -0.07 0.21 -0.08 0.32 -0.10 0.46

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.02 -0.02 0.04 -0.02 0.06 -0.02 0.09 -0.03 0.14 622 -0.03 0.07 -0.06 0.12 -0.07 0.18 -0.08 0.26 -0.09 0.48

1000 -0.01 0.02 -0.02 0.04 -0.02 0.07 -0.02 0.10 -0.03 0.16 1000 -0.01 0.02 -0.02 0.05 -0.02 0.07 -0.03 0.11 -0.03 0.18

Tab. 4: Descriptive statistics of quantile estimates: b = (1, 1)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.02 -0.03 0.06 -0.03 0.09 -0.03 0.13 -0.04 0.23 1 -0.01 0.02 -0.03 0.06 -0.03 0.09 -0.04 0.14 -0.04 0.25

2 -0.01 0.02 -0.02 0.05 -0.03 0.08 -0.03 0.11 -0.04 0.18 2 -0.01 0.02 -0.03 0.06 -0.03 0.09 -0.03 0.12 -0.04 0.19

· · · · · · · · · · · · · · · · · · · · · ·424 -0.05 0.10 -0.09 0.23 -0.10 0.31 -0.12 0.49 -0.14 0.74 325 -0.06 0.11 -0.10 0.22 -0.11 0.30 -0.13 0.39 -0.15 0.71

425 -0.05 0.10 -0.10 0.21 -0.10 0.33 -0.12 0.40 -0.14 0.66 423 -0.06 0.11 -0.10 0.25 -0.11 0.32 -0.12 0.46 -0.15 0.75

597 -0.05 0.09 -0.09 0.24 -0.10 0.33 -0.11 0.45 -0.13 0.81 424 -0.06 0.10 -0.10 0.22 -0.11 0.32 -0.13 0.45 -0.15 0.88

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.03 -0.03 0.07 -0.03 0.08 -0.04 0.13 -0.04 0.20 597 -0.05 0.09 -0.10 0.22 -0.10 0.34 -0.12 0.46 -0.14 0.71

1000 -0.02 0.03 -0.03 0.06 -0.03 0.10 -0.04 0.14 -0.05 0.26 1000 -0.02 0.03 -0.03 0.08 -0.04 0.11 -0.04 0.16 -0.05 0.27

Tab. 5: Descriptive statistics of quantile estimates: b = (1, 2)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

19

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 48: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-10

Foreign exchange rate portfolio

MC quantile estimation: Descriptive statistics of quantile esti-mates: b = (1, 2)>, MC simulation with M = 10000, N = 100,T = 1000. GHICAstatMCq.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.01 -0.02 0.04 -0.02 0.06 -0.02 0.08 -0.02 0.14 1 -0.01 0.02 -0.02 0.05 -0.02 0.06 -0.02 0.09 -0.03 0.14

2 -0.01 0.01 -0.01 0.04 -0.02 0.05 -0.02 0.08 -0.02 0.13 2 -0.01 0.01 -0.02 0.04 -0.02 0.06 -0.02 0.09 -0.02 0.15

· · · · · · · · · · · · · · · · · · · · · ·424 -0.03 0.07 -0.06 0.15 -0.07 0.21 -0.08 0.34 -0.09 0.50 423 -0.04 0.06 -0.07 0.15 -0.07 0.23 -0.08 0.33 -0.10 0.45

425 -0.03 0.06 -0.06 0.14 -0.07 0.22 -0.08 0.27 -0.09 0.44 425 -0.04 0.06 -0.07 0.14 -0.07 0.24 -0.08 0.28 -0.10 0.46

426 -0.03 0.04 -0.05 0.11 -0.06 0.18 -0.06 0.25 -0.08 0.38 426 -0.03 0.05 -0.06 0.12 -0.06 0.18 -0.07 0.27 -0.08 0.53

597 -0.03 0.06 -0.06 0.16 -0.06 0.22 -0.07 0.30 -0.09 0.54 607 -0.04 0.06 -0.06 0.15 -0.07 0.21 -0.08 0.32 -0.10 0.46

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.02 -0.02 0.04 -0.02 0.06 -0.02 0.09 -0.03 0.14 622 -0.03 0.07 -0.06 0.12 -0.07 0.18 -0.08 0.26 -0.09 0.48

1000 -0.01 0.02 -0.02 0.04 -0.02 0.07 -0.02 0.10 -0.03 0.16 1000 -0.01 0.02 -0.02 0.05 -0.02 0.07 -0.03 0.11 -0.03 0.18

Tab. 4: Descriptive statistics of quantile estimates: b = (1, 1)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.02 -0.03 0.06 -0.03 0.09 -0.03 0.13 -0.04 0.23 1 -0.01 0.02 -0.03 0.06 -0.03 0.09 -0.04 0.14 -0.04 0.25

2 -0.01 0.02 -0.02 0.05 -0.03 0.08 -0.03 0.11 -0.04 0.18 2 -0.01 0.02 -0.03 0.06 -0.03 0.09 -0.03 0.12 -0.04 0.19

· · · · · · · · · · · · · · · · · · · · · ·424 -0.05 0.10 -0.09 0.23 -0.10 0.31 -0.12 0.49 -0.14 0.74 325 -0.06 0.11 -0.10 0.22 -0.11 0.30 -0.13 0.39 -0.15 0.71

425 -0.05 0.10 -0.10 0.21 -0.10 0.33 -0.12 0.40 -0.14 0.66 423 -0.06 0.11 -0.10 0.25 -0.11 0.32 -0.12 0.46 -0.15 0.75

597 -0.05 0.09 -0.09 0.24 -0.10 0.33 -0.11 0.45 -0.13 0.81 424 -0.06 0.10 -0.10 0.22 -0.11 0.32 -0.13 0.45 -0.15 0.88

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.03 -0.03 0.07 -0.03 0.08 -0.04 0.13 -0.04 0.20 597 -0.05 0.09 -0.10 0.22 -0.10 0.34 -0.12 0.46 -0.14 0.71

1000 -0.02 0.03 -0.03 0.06 -0.03 0.10 -0.04 0.14 -0.05 0.26 1000 -0.02 0.03 -0.03 0.08 -0.04 0.11 -0.04 0.16 -0.05 0.27

Tab. 5: Descriptive statistics of quantile estimates: b = (1, 2)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

19

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 49: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-11

Foreign exchange rate portfolio

MC quantile estimation: Descriptive statistics of quantile esti-mates: b = (1, 2)>, MC simulation with M = 10000, N = 100,T = 1000. GHICAstatMCq.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.01 -0.02 0.04 -0.02 0.06 -0.02 0.08 -0.02 0.14 1 -0.01 0.02 -0.02 0.05 -0.02 0.06 -0.02 0.09 -0.03 0.14

2 -0.01 0.01 -0.01 0.04 -0.02 0.05 -0.02 0.08 -0.02 0.13 2 -0.01 0.01 -0.02 0.04 -0.02 0.06 -0.02 0.09 -0.02 0.15

· · · · · · · · · · · · · · · · · · · · · ·424 -0.03 0.07 -0.06 0.15 -0.07 0.21 -0.08 0.34 -0.09 0.50 423 -0.04 0.06 -0.07 0.15 -0.07 0.23 -0.08 0.33 -0.10 0.45

425 -0.03 0.06 -0.06 0.14 -0.07 0.22 -0.08 0.27 -0.09 0.44 425 -0.04 0.06 -0.07 0.14 -0.07 0.24 -0.08 0.28 -0.10 0.46

426 -0.03 0.04 -0.05 0.11 -0.06 0.18 -0.06 0.25 -0.08 0.38 426 -0.03 0.05 -0.06 0.12 -0.06 0.18 -0.07 0.27 -0.08 0.53

597 -0.03 0.06 -0.06 0.16 -0.06 0.22 -0.07 0.30 -0.09 0.54 607 -0.04 0.06 -0.06 0.15 -0.07 0.21 -0.08 0.32 -0.10 0.46

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.02 -0.02 0.04 -0.02 0.06 -0.02 0.09 -0.03 0.14 622 -0.03 0.07 -0.06 0.12 -0.07 0.18 -0.08 0.26 -0.09 0.48

1000 -0.01 0.02 -0.02 0.04 -0.02 0.07 -0.02 0.10 -0.03 0.16 1000 -0.01 0.02 -0.02 0.05 -0.02 0.07 -0.03 0.11 -0.03 0.18

Tab. 4: Descriptive statistics of quantile estimates: b = (1, 1)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.02 -0.03 0.06 -0.03 0.09 -0.03 0.13 -0.04 0.23 1 -0.01 0.02 -0.03 0.06 -0.03 0.09 -0.04 0.14 -0.04 0.25

2 -0.01 0.02 -0.02 0.05 -0.03 0.08 -0.03 0.11 -0.04 0.18 2 -0.01 0.02 -0.03 0.06 -0.03 0.09 -0.03 0.12 -0.04 0.19

· · · · · · · · · · · · · · · · · · · · · ·424 -0.05 0.10 -0.09 0.23 -0.10 0.31 -0.12 0.49 -0.14 0.74 325 -0.06 0.11 -0.10 0.22 -0.11 0.30 -0.13 0.39 -0.15 0.71

425 -0.05 0.10 -0.10 0.21 -0.10 0.33 -0.12 0.40 -0.14 0.66 423 -0.06 0.11 -0.10 0.25 -0.11 0.32 -0.12 0.46 -0.15 0.75

597 -0.05 0.09 -0.09 0.24 -0.10 0.33 -0.11 0.45 -0.13 0.81 424 -0.06 0.10 -0.10 0.22 -0.11 0.32 -0.13 0.45 -0.15 0.88

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.03 -0.03 0.07 -0.03 0.08 -0.04 0.13 -0.04 0.20 597 -0.05 0.09 -0.10 0.22 -0.10 0.34 -0.12 0.46 -0.14 0.71

1000 -0.02 0.03 -0.03 0.06 -0.03 0.10 -0.04 0.14 -0.05 0.26 1000 -0.02 0.03 -0.03 0.08 -0.04 0.11 -0.04 0.16 -0.05 0.27

Tab. 5: Descriptive statistics of quantile estimates: b = (1, 2)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

19

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 50: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-12

Foreign exchange rate portfolio

MC quantile estimation: Descriptive statistics of quantile esti-mates: b = (−1, 2)>, MC simulation with M = 10000, N = 100,T = 1000. GHICAstatMCq.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.10 1 -0.01 0.01 -0.01 0.03 -0.02 0.05 -0.02 0.07 -0.02 0.10

2 -0.01 0.01 -0.01 0.03 -0.01 0.05 -0.02 0.07 -0.02 0.12 2 -0.01 0.01 -0.01 0.03 -0.02 0.05 -0.02 0.08 -0.02 0.13

· · · · · · · · · · · · · · · · · · · · · ·601 -0.03 0.05 -0.06 0.15 -0.06 0.22 -0.07 0.28 -0.08 0.41 601 -0.03 0.05 -0.06 0.15 -0.07 0.19 -0.08 0.30 -0.09 0.48

607 -0.04 0.06 -0.06 0.13 -0.07 0.18 -0.08 0.26 -0.09 0.47 606 -0.03 0.06 -0.06 0.15 -0.07 0.23 -0.08 0.30 -0.09 0.55

611 -0.03 0.06 -0.06 0.13 -0.06 0.20 -0.07 0.28 -0.08 0.43 607 -0.04 0.07 -0.07 0.13 -0.07 0.19 -0.08 0.26 -0.10 0.48

613 -0.03 0.05 -0.06 0.15 -0.07 0.20 -0.08 0.27 -0.09 0.39 613 -0.04 0.06 -0.06 0.12 -0.07 0.18 -0.08 0.27 -0.09 0.49

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.09 999 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.07 -0.02 0.11

1000 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.09 1000 -0.01 0.01 -0.01 0.03 -0.01 0.05 -0.02 0.07 -0.02 0.10

Tab. 6: Descriptive statistics of quantile estimates: b = (−1, 2)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.07 -0.02 0.11 1 -0.01 0.01 -0.01 0.03 -0.02 0.04 -0.02 0.07 -0.02 0.11

2 -0.01 0.01 -0.01 0.03 -0.01 0.05 -0.02 0.06 -0.02 0.10 2 -0.01 0.01 -0.01 0.03 -0.02 0.05 -0.02 0.08 -0.02 0.12

· · · · · · · · · · · · · · · · · · · · · ·607 -0.04 0.06 -0.06 0.13 -0.07 0.18 -0.08 0.25 -0.09 0.41 601 -0.03 0.06 -0.06 0.15 -0.07 0.22 -0.08 0.30 -0.09 0.55

608 -0.03 0.05 -0.06 0.13 -0.06 0.19 -0.07 0.28 -0.09 0.40 605 -0.04 0.06 -0.06 0.14 -0.07 0.19 -0.08 0.30 -0.09 0.48

612 -0.03 0.05 -0.06 0.14 -0.07 0.21 -0.08 0.27 -0.09 0.47 612 -0.04 0.07 -0.07 0.14 -0.07 0.19 -0.08 0.27 -0.09 0.38

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.10 999 -0.01 0.01 -0.01 0.02 -0.01 0.04 -0.02 0.05 -0.02 0.11

1000 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.10 1000 -0.01 0.01 -0.01 0.03 -0.02 0.04 -0.02 0.06 -0.02 0.10

Tab. 7: Descriptive statistics of quantile estimates: b = (−2, 1)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

20

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 51: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-13

Foreign exchange rate portfolio

MC quantile estimation: Descriptive statistics of quantile esti-mates: b = (−1, 2)>, MC simulation with M = 10000, N = 100,T = 1000. GHICAstatMCq.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.10 1 -0.01 0.01 -0.01 0.03 -0.02 0.05 -0.02 0.07 -0.02 0.10

2 -0.01 0.01 -0.01 0.03 -0.01 0.05 -0.02 0.07 -0.02 0.12 2 -0.01 0.01 -0.01 0.03 -0.02 0.05 -0.02 0.08 -0.02 0.13

· · · · · · · · · · · · · · · · · · · · · ·601 -0.03 0.05 -0.06 0.15 -0.06 0.22 -0.07 0.28 -0.08 0.41 601 -0.03 0.05 -0.06 0.15 -0.07 0.19 -0.08 0.30 -0.09 0.48

607 -0.04 0.06 -0.06 0.13 -0.07 0.18 -0.08 0.26 -0.09 0.47 606 -0.03 0.06 -0.06 0.15 -0.07 0.23 -0.08 0.30 -0.09 0.55

611 -0.03 0.06 -0.06 0.13 -0.06 0.20 -0.07 0.28 -0.08 0.43 607 -0.04 0.07 -0.07 0.13 -0.07 0.19 -0.08 0.26 -0.10 0.48

613 -0.03 0.05 -0.06 0.15 -0.07 0.20 -0.08 0.27 -0.09 0.39 613 -0.04 0.06 -0.06 0.12 -0.07 0.18 -0.08 0.27 -0.09 0.49

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.09 999 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.07 -0.02 0.11

1000 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.09 1000 -0.01 0.01 -0.01 0.03 -0.01 0.05 -0.02 0.07 -0.02 0.10

Tab. 6: Descriptive statistics of quantile estimates: b = (−1, 2)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.07 -0.02 0.11 1 -0.01 0.01 -0.01 0.03 -0.02 0.04 -0.02 0.07 -0.02 0.11

2 -0.01 0.01 -0.01 0.03 -0.01 0.05 -0.02 0.06 -0.02 0.10 2 -0.01 0.01 -0.01 0.03 -0.02 0.05 -0.02 0.08 -0.02 0.12

· · · · · · · · · · · · · · · · · · · · · ·607 -0.04 0.06 -0.06 0.13 -0.07 0.18 -0.08 0.25 -0.09 0.41 601 -0.03 0.06 -0.06 0.15 -0.07 0.22 -0.08 0.30 -0.09 0.55

608 -0.03 0.05 -0.06 0.13 -0.06 0.19 -0.07 0.28 -0.09 0.40 605 -0.04 0.06 -0.06 0.14 -0.07 0.19 -0.08 0.30 -0.09 0.48

612 -0.03 0.05 -0.06 0.14 -0.07 0.21 -0.08 0.27 -0.09 0.47 612 -0.04 0.07 -0.07 0.14 -0.07 0.19 -0.08 0.27 -0.09 0.38

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.10 999 -0.01 0.01 -0.01 0.02 -0.01 0.04 -0.02 0.05 -0.02 0.11

1000 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.10 1000 -0.01 0.01 -0.01 0.03 -0.02 0.04 -0.02 0.06 -0.02 0.10

Tab. 7: Descriptive statistics of quantile estimates: b = (−2, 1)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

20

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 52: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-14

Foreign exchange rate portfolio

MC quantile estimation: Descriptive statistics of quantile esti-mates: b = (−2, 1)>, MC simulation with M = 10000, N = 100,T = 1000. GHICAstatMCq.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.10 1 -0.01 0.01 -0.01 0.03 -0.02 0.05 -0.02 0.07 -0.02 0.10

2 -0.01 0.01 -0.01 0.03 -0.01 0.05 -0.02 0.07 -0.02 0.12 2 -0.01 0.01 -0.01 0.03 -0.02 0.05 -0.02 0.08 -0.02 0.13

· · · · · · · · · · · · · · · · · · · · · ·601 -0.03 0.05 -0.06 0.15 -0.06 0.22 -0.07 0.28 -0.08 0.41 601 -0.03 0.05 -0.06 0.15 -0.07 0.19 -0.08 0.30 -0.09 0.48

607 -0.04 0.06 -0.06 0.13 -0.07 0.18 -0.08 0.26 -0.09 0.47 606 -0.03 0.06 -0.06 0.15 -0.07 0.23 -0.08 0.30 -0.09 0.55

611 -0.03 0.06 -0.06 0.13 -0.06 0.20 -0.07 0.28 -0.08 0.43 607 -0.04 0.07 -0.07 0.13 -0.07 0.19 -0.08 0.26 -0.10 0.48

613 -0.03 0.05 -0.06 0.15 -0.07 0.20 -0.08 0.27 -0.09 0.39 613 -0.04 0.06 -0.06 0.12 -0.07 0.18 -0.08 0.27 -0.09 0.49

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.09 999 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.07 -0.02 0.11

1000 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.09 1000 -0.01 0.01 -0.01 0.03 -0.01 0.05 -0.02 0.07 -0.02 0.10

Tab. 6: Descriptive statistics of quantile estimates: b = (−1, 2)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.07 -0.02 0.11 1 -0.01 0.01 -0.01 0.03 -0.02 0.04 -0.02 0.07 -0.02 0.11

2 -0.01 0.01 -0.01 0.03 -0.01 0.05 -0.02 0.06 -0.02 0.10 2 -0.01 0.01 -0.01 0.03 -0.02 0.05 -0.02 0.08 -0.02 0.12

· · · · · · · · · · · · · · · · · · · · · ·607 -0.04 0.06 -0.06 0.13 -0.07 0.18 -0.08 0.25 -0.09 0.41 601 -0.03 0.06 -0.06 0.15 -0.07 0.22 -0.08 0.30 -0.09 0.55

608 -0.03 0.05 -0.06 0.13 -0.06 0.19 -0.07 0.28 -0.09 0.40 605 -0.04 0.06 -0.06 0.14 -0.07 0.19 -0.08 0.30 -0.09 0.48

612 -0.03 0.05 -0.06 0.14 -0.07 0.21 -0.08 0.27 -0.09 0.47 612 -0.04 0.07 -0.07 0.14 -0.07 0.19 -0.08 0.27 -0.09 0.38

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.10 999 -0.01 0.01 -0.01 0.02 -0.01 0.04 -0.02 0.05 -0.02 0.11

1000 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.10 1000 -0.01 0.01 -0.01 0.03 -0.02 0.04 -0.02 0.06 -0.02 0.10

Tab. 7: Descriptive statistics of quantile estimates: b = (−2, 1)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

20

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 53: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-15

Foreign exchange rate portfolio

MC quantile estimation: Descriptive statistics of quantile esti-mates: b = (−2, 1)>, MC simulation with M = 10000, N = 100,T = 1000. GHICAstatMCq.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.10 1 -0.01 0.01 -0.01 0.03 -0.02 0.05 -0.02 0.07 -0.02 0.10

2 -0.01 0.01 -0.01 0.03 -0.01 0.05 -0.02 0.07 -0.02 0.12 2 -0.01 0.01 -0.01 0.03 -0.02 0.05 -0.02 0.08 -0.02 0.13

· · · · · · · · · · · · · · · · · · · · · ·601 -0.03 0.05 -0.06 0.15 -0.06 0.22 -0.07 0.28 -0.08 0.41 601 -0.03 0.05 -0.06 0.15 -0.07 0.19 -0.08 0.30 -0.09 0.48

607 -0.04 0.06 -0.06 0.13 -0.07 0.18 -0.08 0.26 -0.09 0.47 606 -0.03 0.06 -0.06 0.15 -0.07 0.23 -0.08 0.30 -0.09 0.55

611 -0.03 0.06 -0.06 0.13 -0.06 0.20 -0.07 0.28 -0.08 0.43 607 -0.04 0.07 -0.07 0.13 -0.07 0.19 -0.08 0.26 -0.10 0.48

613 -0.03 0.05 -0.06 0.15 -0.07 0.20 -0.08 0.27 -0.09 0.39 613 -0.04 0.06 -0.06 0.12 -0.07 0.18 -0.08 0.27 -0.09 0.49

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.09 999 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.07 -0.02 0.11

1000 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.09 1000 -0.01 0.01 -0.01 0.03 -0.01 0.05 -0.02 0.07 -0.02 0.10

Tab. 6: Descriptive statistics of quantile estimates: b = (−1, 2)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.07 -0.02 0.11 1 -0.01 0.01 -0.01 0.03 -0.02 0.04 -0.02 0.07 -0.02 0.11

2 -0.01 0.01 -0.01 0.03 -0.01 0.05 -0.02 0.06 -0.02 0.10 2 -0.01 0.01 -0.01 0.03 -0.02 0.05 -0.02 0.08 -0.02 0.12

· · · · · · · · · · · · · · · · · · · · · ·607 -0.04 0.06 -0.06 0.13 -0.07 0.18 -0.08 0.25 -0.09 0.41 601 -0.03 0.06 -0.06 0.15 -0.07 0.22 -0.08 0.30 -0.09 0.55

608 -0.03 0.05 -0.06 0.13 -0.06 0.19 -0.07 0.28 -0.09 0.40 605 -0.04 0.06 -0.06 0.14 -0.07 0.19 -0.08 0.30 -0.09 0.48

612 -0.03 0.05 -0.06 0.14 -0.07 0.21 -0.08 0.27 -0.09 0.47 612 -0.04 0.07 -0.07 0.14 -0.07 0.19 -0.08 0.27 -0.09 0.38

· · · · · · · · · · · · · · · · · · · · · ·999 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.10 999 -0.01 0.01 -0.01 0.02 -0.01 0.04 -0.02 0.05 -0.02 0.11

1000 -0.01 0.01 -0.01 0.03 -0.01 0.04 -0.02 0.06 -0.02 0.10 1000 -0.01 0.01 -0.01 0.03 -0.02 0.04 -0.02 0.06 -0.02 0.10

Tab. 7: Descriptive statistics of quantile estimates: b = (−2, 1)>, MC simulation with M = 10000, N = 100, T = 1000.GHICAmcqstat.xpl

20

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 54: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-16

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (1, 1)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(16)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0500)

5 10

X*E2

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 55: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-17

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (1, 1)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(16)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0100)

5 10

X*E2

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 56: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-18

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (1, 1)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(16)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0050)

5 10

X*E2

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 57: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-19

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (1, 1)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(16)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0025)

5 10

X*E2

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 58: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-20

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (1, 1)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(16)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0010)

5 10

X*E2

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 59: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-21

Backtesting

LR1 is the likelihood ratio w.r.t. the risk level test and LR2 w.r.t. the independence

test. * indicates the rejection at 99% level.

GHICA (HYP fit) GHICA (NIG fit) RiskMetrics Method T-deGARCH Methodb> a(10−2) n/T% LR1 LR2 n/T% LR1 LR2 n/T% LR1 LR2 n/T% LR1 LR2

(1,1) 5 6.7 5.52 *27.60 5.8 1.28 *28.82 10.8 *53.95 *19.84 11.0 *57.33 *26.05(1,1) 1 1.9 6.47 *16.29 1.4 1.43 *11.14 5.5 *99.59 *21.70 5.3 *92.67 *22.37(1,1) 0.5 0.9 2.59 5.58 0.7 0.71 0.00 5.0 *142.33 *23.44 4.2 *106.16 *18.96(1,1) 0.25 0.4 0.76 0.00 0.3 0.09 0.00 3.8 *137.10 *22.04 2.5 *70.64 *26.70(1,1) 0.1 0.2 0.77 0.00 0.1 0.00 0.00 2.3 *100.72 *27.74 1.2 *37.75 0.00

(1,2) 5 6.8 6.16 *27.23 5.4 0.32 *30.50 10.9 *55.63 *24.78 11.0 *57.33 *24.48(1,2) 1 1.6 3.07 *10.51 1.4 1.43 *11.14 5.3 *92.67 *22.37 4.9 *79.30 *23.81(1,2) 0.5 0.9 2.59 0.00 0.8 1.52 0.00 4.7 *128.43 *21.77 4.4 *114.93 *22.83(1,2) 0.25 0.5 1.93 0.00 0.5 1.93 0.00 4.0 *148.23 *24.40 2.8 *84.94 *22.67(1,2) 0.1 0.3 2.59 0.00 0.2 0.77 0.00 2.5 *113.53 *18.05 1.5 *53.43 *7.11

(-1,2) 5 6.4 3.80 *38.56 5.5 0.51 *28.17 12.0 *75.40 *35.66 12.0 *75.40 *35.66(-1,2) 1 1.4 1.43 *11.14 1.2 0.37 *11.86 6.0 *117.58 *25.86 5.5 *99.59 *23.40(-1,2) 0.5 1.2 *7.06 *11.86 1.1 5.38 *15.84 4.4 *114.93 *18.33 3.2 *65.54 *20.90(-1,2) 0.25 1.0 *12.78 *7.97 0.7 5.43 *8.76 3.1 *99.91 *21.26 2.5 *70.64 *23.81(-1,2) 0.1 0.6 *11.52 *9.11 0.3 2.59 *10.80 2.5 *113.53 *23.81 1.7 *64.58 *28.58

(-2,1) 5 6.5 4.34 *22.35 5.0 0.00 *20.34 12.7 *89.18 *18.41 12.7 *89.18 *18.41(-2,1) 1 1.6 3.07 0.00 1.1 0.09 0.00 5.6 *103.12 *14.79 4.7 *72.87 4.49(-2,1) 0.5 0.9 2.59 0.00 0.8 1.52 0.00 4.2 *106.16 2.40 3.2 *65.54 5.54(-2,1) 0.25 0.5 1.93 0.00 0.3 0.09 0.00 3.2 *105.05 5.54 2.3 *61.50 6.22(-2,1) 0.1 0.1 0.00 0.00 0.1 0.00 0.00 2.3 *100.72 6.22 1.6 *58.94 0.00

Tab. 8: Backtesting of the VaR forecast of the exchange portfolios: MC simulation with M = 10000, N = 100, T = 1000. * indicates themodel is rejected at 99% level.

GHICAfxVaRplot.xpl

22

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 60: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-22

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (1, 2)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(16)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0500)

5 10

X*E2

-20

-15

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 61: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-23

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (1, 2)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(16)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0100)

5 10

X*E2

-20

-15

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 62: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-24

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (1, 2)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(16)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0050)

5 10

X*E2

-20

-15

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 63: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-25

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (1, 2)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(16)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0025)

5 10

X*E2

-20

-15

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 64: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-26

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (1, 2)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(16)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0010)

5 10

X*E2

-20

-15

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 65: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-27

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (−1, 2)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(13)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0500)

5 10

X*E2

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 66: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-28

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (−1, 2)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(13)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0100)

5 10

X*E2

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 67: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-29

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (−1, 2)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(13)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0050)

5 10

X*E2

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 68: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-30

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (−1, 2)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(13)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0025)

5 10

X*E2

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 69: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-31

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (−1, 2)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(13)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0010)

5 10

X*E2

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 70: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-32

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (−2, 1)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(14)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0500)

5 10

X*E2

-15

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 71: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-33

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (−2, 1)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(14)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0100)

5 10

X*E2

-15

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 72: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-34

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (−2, 1)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(14)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0050)

5 10

X*E2

-15

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 73: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-35

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (−2, 1)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(14)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0025)

5 10

X*E2

-15

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 74: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-36

VaR forecasts of exchange rate portfolio (DEM/USD and GBP/USD) for 1000 days to

1994-04-01, b = (−2, 1)>. VaR forecasts and exceedances of the GHICA methodology

are marked in blue while those based on the RiskMetrics and T(14)-deGARCH are

marked in red and green respectively. GHICAfxVaRplot.xplVaR plot (hyp-alpha = 0.0010)

5 10

X*E2

-15

-10

-50

5

Y*E

-2

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 75: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-37

German stock portfolio

Data: 20 components (ALLIANZ, BASF, BAYER, BMW,COBANK, DAIMLER, DEUTSCHE BANK, DEGUSSA, DRESD-NER, HOECHST, KARSTADT, LINDE, MAN, MANNESMANN,PREUSSAG, RWE, SCHERING, SIEMENS, THYSSEN, VOLKSW)5748 observations from 1974-01-02 to 1996-12-30. Downloadedfrom MD*Base.Trading strategy: b = (1, · · · , 1)> ∈ IR20

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 76: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-38

German stock portfolio

Descriptive statistics: GHICASFMdescriptive.xpl

Returns 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Mean 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

Variance 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

Skewness 0.07 -0.17 0.04 -0.01 -0.24 0.03 -0.30 -0.39 0.12 -0.12 -0.40 -0.24 -0.59 -0.27 0.13 -0.19 -0.03 -0.54 -0.05 -0.32

Kurtosis 32.40 8.65 9.60 17.02 10.03 26.67 13.77 19.12 8.82 9.98 20.43 14.56 18.03 13.69 10.34 16.72 9.57 10.30 6.10 10.23

ρ1 -0.06 -0.02 -0.02 -0.01 0.04 -0.02 0.03 0.02 0.06 0.01 0.00 -0.03 0.02 0.05 0.08 0.04 0.06 0.06 0.04 0.05

ρ2 -0.02 -0.04 -0.04 -0.01 -0.04 -0.01 -0.03 -0.01 -0.03 -0.04 0.00 -0.01 -0.01 -0.01 -0.05 -0.02 -0.01 -0.01 -0.05 -0.01

ICs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Mean 0.01 -0.00 0.01 0.01 0.00 0.00 0.00 0.01 -0.00 0.01 0.01 -0.00 -0.01 -0.03 -0.00 -0.01 -0.00 -0.02 -0.00 0.00

Variance 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00

Skewness -1.44 -0.81 0.71 0.8 1.01 -0.57 0.68 0.38 -0.84 0.48 0.52 -0.04 -0.04 -0.24 -0.17 -0.51 -0.05 -0.13 0.06 -0.32

Kurtosis 59.71 46.51 57.51 43.72 81.20 13.92 25.31 27.28 14.84 14.30 14.40 9.12 11.89 6.26 6.94 9.34 7.08 6.48 5.62 6.15

ρ1 0.00 -0.26 -0.26 -0.11 -0.12 0.10 -0.02 -0.13 -0.01 0.06 -0.12 0.08 0.04 0.04 0.08 0.08 0.08 0.05 -0.01 0.07

ρ2 -0.01 0.02 0.02 0.00 0.00 0.01 0.00 -0.01 -0.01 0.00 -0.01 0.01 0.02 -0.01 0.00 -0.03 0.02 0.00 0.00 0.00

Tab. 9: Descriptive statistics of the 20-dimensional German stock portfolio.GHICAsfmdescriptive.xpl

ICs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

HYP α 2.01 1.54 1.63 1.63 1.59 2.19 1.64 1.57 1.59 1.83 1.78 1.89 1.96 1.76 1.70 1.95 1.76 1.85 1.91 1.79

HYP β 0.07 -0.04 0.07 0.11 0.13 0.09 0.14 0.02 -0.10 0.04 0.07 0.09 -0.00 -0.12 -0.11 -0.21 -0.03 -0.07 -0.00 -0.20

HYP δ 1.12 0.42 0.38 0.35 0.30 1.34 0.38 0.30 0.41 0.77 0.61 0.91 0.96 0.72 0.53 0.91 0.65 0.81 0.90 0.72

HYP µ -0.05 0.02 -0.04 -0.10 -0.12 -0.08 -0.12 -0.01 0.10 -0.03 -0.06 -0.09 -0.01 0.09 0.11 0.17 0.02 0.04 -0.00 0.20

NIG α 1.63 0.97 0.98 0.95 0.90 1.83 0.95 0.90 1.00 1.34 1.21 1.45 1.54 1.29 1.15 1.55 1.25 1.35 1.44 1.34

NIG β 0.05 -0.03 0.06 0.10 0.11 0.06 0.12 0.02 -0.09 0.04 0.07 0.06 0.00 -0.10 -0.10 -0.22 -0.02 -0.06 -0.00 -0.19

NIG δ 1.56 1.00 0.90 0.86 0.83 1.73 0.86 0.85 0.97 1.24 1.07 1.37 1.42 1.23 1.05 1.38 1.15 1.26 1.34 1.23

NIG µ -0.03 0.02 -0.04 -0.09 -0.10 -0.04 -0.11 -0.01 0.09 -0.03 -0.06 -0.06 -0.02 0.07 0.10 0.18 0.01 0.02 -0.00 0.19

Tab. 10: HYP and NIG parameters estimates of the German stock portfolio.GHICAsfmdescriptive.xpl

25

Linear dependence: the smallest correlation (ALLIANZ -PREUSSAG) is 0.37

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 77: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-39

German stock portfolio

ICA and GH: HYP and NIG parameters estimated of the ICs onthe basis of the German stock portfolio.GHICASFMdescriptive.xpl

ICs 1 2 3 4 5 6 7 8 9 10 · · ·HYP α 2.01 1.54 1.63 1.63 1.59 2.19 1.64 1.57 1.59 1.83 · · ·HYP β 0.07 -0.04 0.07 0.11 0.13 0.09 0.14 0.02 -0.10 0.04 · · ·HYP δ 1.12 0.42 0.38 0.35 0.30 1.34 0.38 0.30 0.41 0.77 · · ·HYP µ -0.05 0.02 -0.04 -0.10 -0.12 -0.08 -0.12 -0.01 0.10 -0.03 · · ·NIG α 1.63 0.97 0.98 0.95 0.90 1.83 0.95 0.90 1.00 1.34 · · ·NIG β 0.05 -0.03 0.06 0.10 0.11 0.06 0.12 0.02 -0.09 0.04 · · ·NIG δ 1.56 1.00 0.90 0.86 0.83 1.73 0.86 0.85 0.97 1.24 · · ·NIG µ -0.03 0.02 -0.04 -0.09 -0.10 -0.04 -0.11 -0.01 0.09 -0.03 · · ·

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 78: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-40

German stock portfolio

Density estimations: pdfs of the 20 ICs on the basis of the Germanstock portfolio. The HYP fit is displayed as a straight curve and thenonparametric density estimation is plotted by a dotted curve.GHICASFMdescriptive.xpl

Empirical density & hyp fitting - IC1

-5 0

X

00.1

0.20.3

0.4

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 79: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-41

German stock portfolio

Density estimation: pdfs of the 20 ICs on the basis of the Germanstock portfolio. The NIG fit is displayed as a straight curve and thenonparametric density estimation is plotted by a dotted curve.GHICASFMdescriptive.xpl

Empirical density & nig fitting - IC1

-5 0

X

00.1

0.20.3

0.4

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 80: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-42

German stock portfolio

MC quantile estimation: Descriptive statistics of quantile esti-mates: MC simulation with M = 10000, N = 100, T = 1000.GHICAstatMCq.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.28 0.36 -0.43 0.66 -0.45 0.92 -0.50 1.33 -0.56 1.91 1 -0.30 0.42 -0.47 0.82 -0.49 1.13 -0.55 1.55 -0.62 2.15

2 -0.27 0.35 -0.41 0.76 -0.43 0.93 -0.48 1.25 -0.54 2.04 2 -0.29 0.40 -0.44 0.79 -0.47 0.93 -0.52 1.35 -0.59 1.96

· · · · · · · · · · · · · · · · · · · · · ·993 -0.38 0.48 -0.59 1.11 -0.63 1.41 -0.70 1.98 -0.80 3.34 993 -0.40 0.58 -0.64 1.17 -0.68 1.55 -0.76 2.15 -0.87 3.75

· · · · · · · · · · · · · · · · · · · · · ·997 -0.4 0.54 -0.64 1.08 -0.67 1.58 -0.75 2.15 -0.84 3.18 997 -0.44 0.71 -0.69 1.55 -0.73 2.14 -0.81 2.61 -0.92 3.42

998 -0.38 0.55 -0.59 0.96 -0.63 1.56 -0.70 2.08 -0.79 3.22 998 -0.41 0.59 -0.64 1.20 -0.68 1.58 -0.76 1.91 -0.86 3.49

999 -0.32 0.46 -0.49 0.89 -0.52 1.39 -0.58 1.90 -0.65 2.75 999 -0.34 0.50 -0.53 0.96 -0.56 1.18 -0.62 1.71 -0.70 2.42

1000 -0.35 0.49 -0.54 0.94 -0.57 1.33 -0.63 1.75 -0.72 2.89 1000 -0.38 0.53 -0.59 0.99 -0.62 1.38 -0.69 1.90 -0.78 2.87

Tab. 11: Descriptive statistics of quantile estimates: MC simulation with M = 10000, N = 100, T = 1000.GHICAstatMCq.xpl

28

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 81: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-43

German stock portfolio

MC quantile estimation: Descriptive statistics of quantile esti-mates: MC simulation with M = 10000, N = 100, T = 1000.GHICAstatMCq.xpl

HYP a 5% a 1% a 0.5% a 0.25% a 0.1% NIG a 5% a 1% a 0.5% a 0.25% a 0.1%

Day mean STD% mean STD% mean STD% mean STD% mean STD% Day mean STD% mean STD% mean STD% mean STD% mean STD%

1 -0.28 0.36 -0.43 0.66 -0.45 0.92 -0.50 1.33 -0.56 1.91 1 -0.30 0.42 -0.47 0.82 -0.49 1.13 -0.55 1.55 -0.62 2.15

2 -0.27 0.35 -0.41 0.76 -0.43 0.93 -0.48 1.25 -0.54 2.04 2 -0.29 0.40 -0.44 0.79 -0.47 0.93 -0.52 1.35 -0.59 1.96

· · · · · · · · · · · · · · · · · · · · · ·993 -0.38 0.48 -0.59 1.11 -0.63 1.41 -0.70 1.98 -0.80 3.34 993 -0.40 0.58 -0.64 1.17 -0.68 1.55 -0.76 2.15 -0.87 3.75

· · · · · · · · · · · · · · · · · · · · · ·997 -0.4 0.54 -0.64 1.08 -0.67 1.58 -0.75 2.15 -0.84 3.18 997 -0.44 0.71 -0.69 1.55 -0.73 2.14 -0.81 2.61 -0.92 3.42

998 -0.38 0.55 -0.59 0.96 -0.63 1.56 -0.70 2.08 -0.79 3.22 998 -0.41 0.59 -0.64 1.20 -0.68 1.58 -0.76 1.91 -0.86 3.49

999 -0.32 0.46 -0.49 0.89 -0.52 1.39 -0.58 1.90 -0.65 2.75 999 -0.34 0.50 -0.53 0.96 -0.56 1.18 -0.62 1.71 -0.70 2.42

1000 -0.35 0.49 -0.54 0.94 -0.57 1.33 -0.63 1.75 -0.72 2.89 1000 -0.38 0.53 -0.59 0.99 -0.62 1.38 -0.69 1.90 -0.78 2.87

Tab. 11: Descriptive statistics of quantile estimates: MC simulation with M = 10000, N = 100, T = 1000.GHICAstatMCq.xpl

28

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 82: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-44

VaR time plots of the German stock portfolio with the equal weights. The dots are

the real portfolio returns, the RiskMetrics (red) and T(19)-deGRACH (green) results

are plotted as a dotted line and the GHICA as a straight line. The exceedances w.r.t.

the risk management models are displayed as cross on the bottom.

GHICASFMVaRplot.xplVaR plot (hyp-a =0.0500)

5 10

X*E2

-1.5

-1-0

.50

0.5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 83: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-45

VaR time plots of the German stock portfolio with the equal weights. The dots are

the real portfolio returns, the RiskMetrics (red) and T(19)-deGRACH (green) results

are plotted as a dotted line and the GHICA as a straight line. The exceedances w.r.t.

the risk management models are displayed as cross on the bottom.

GHICASFMVaRplot.xplVaR plot (hyp-a =0.0100)

5 10

X*E2

-1.5

-1-0

.50

0.5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 84: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-46

VaR time plots of the German stock portfolio with the equal weights. The dots are

the real portfolio returns, the RiskMetrics (red) and T(19)-deGRACH (green) results

are plotted as a dotted line and the GHICA as a straight line. The exceedances w.r.t.

the risk management models are displayed as cross on the bottom.

GHICASFMVaRplot.xplVaR plot (hyp-a =0.0050)

5 10

X*E2

-1.5

-1-0

.50

0.5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 85: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-47

VaR time plots of the German stock portfolio with the equal weights. The dots are

the real portfolio returns, the RiskMetrics (red) and T(19)-deGRACH (green) results

are plotted as a dotted line and the GHICA as a straight line. The exceedances w.r.t.

the risk management models are displayed as cross on the bottom.

GHICASFMVaRplot.xplVaR plot (hyp-a =0.0025)

5 10

X*E2

-1.5

-1-0

.50

0.5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 86: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-48

VaR time plots of the German stock portfolio with the equal weights. The dots are

the real portfolio returns, the RiskMetrics (red) and T(19)-deGRACH (green) results

are plotted as a dotted line and the GHICA as a straight line. The exceedances w.r.t.

the risk management models are displayed as cross on the bottom.

GHICASFMVaRplot.xplVaR plot (hyp-a =0.0010)

5 10

X*E2

-1.5

-1-0

.50

0.5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 87: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-49

VaR time plots of the German stock portfolio with the equal weights. The dots are

the real portfolio returns, the RiskMetrics (red) and T(19)-deGARCH (green) results

are plotted as a dotted line and the GHICA as a straight line. The exceedances w.r.t.

the risk management models are displayed as cross on the bottom.

GHICASFMVaRplot.xplVaR plot (nig-a =0.0500)

5 10

X*E2

-1.5

-1-0

.50

0.5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 88: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-50

VaR time plots of the German stock portfolio with the equal weights. The dots are

the real portfolio returns, the RiskMetrics (red) and T(19)-deGARCH (green) results

are plotted as a dotted line and the GHICA as a straight line. The exceedances w.r.t.

the risk management models are displayed as cross on the bottom.

GHICASFMVaRplot.xplVaR plot (nig-a =0.0100)

5 10

X*E2

-1.5

-1-0

.50

0.5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 89: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-51

VaR time plots of the German stock portfolio with the equal weights. The dots are

the real portfolio returns, the RiskMetrics (red) and T(19)-deGARCH (green) results

are plotted as a dotted line and the GHICA as a straight line. The exceedances w.r.t.

the risk management models are displayed as cross on the bottom.

GHICASFMVaRplot.xplVaR plot (nig-a =0.0050)

5 10

X*E2

-1.5

-1-0

.50

0.5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 90: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-52

VaR time plots of the German stock portfolio with the equal weights. The dots are

the real portfolio returns, the RiskMetrics (red) and T(19)-deGARCH (green) results

are plotted as a dotted line and the GHICA as a straight line. The exceedances w.r.t.

the risk management models are displayed as cross on the bottom.

GHICASFMVaRplot.xplVaR plot (nig-a =0.0025)

5 10

X*E2

-1.5

-1-0

.50

0.5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 91: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-53

VaR time plots of the German stock portfolio with the equal weights. The dots are

the real portfolio returns, the RiskMetrics (red) and T(19)-deGARCH (green) results

are plotted as a dotted line and the GHICA as a straight line. The exceedances w.r.t.

the risk management models are displayed as cross on the bottom.

GHICASFMVaRplot.xplVaR plot (nig-a =0.0010)

5 10

X*E2

-1.5

-1-0

.50

0.5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 92: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-54

Backtesting

LR1 is the likelihood ratio w.r.t. the risk level test and LR2 w.r.t. the independence

test. * indicates the rejection at 99% level.

a = 5% a = 1% a = 0.5% a = 0.25% a = 0.1%n/T% LR1 LR2 n/T% LR1 LR2 n/T% LR1 LR2 n/T% LR1 LR2 n/T% LR1 LR2

HYP 4.9 0.02 *16.76 1.7 4.09 0.13 1.2 *7.06 0.04 0.8 *7.64 0.03 0.7 *15.27 0.00NIG 3.8 3.29 *7.14 1.4 1.43 0.05 0.8 1.52 0.03 0.7 5.43 0.00 0.4 5.09 0.00RiskMetr. 10.4 *47.46 *18.29 5.4 *96.11 *7.28 3.5 *77.12 3.92 2.8 *84.94 0.42 2.0 *82.19 0.15T4-deGARCG 12.7 *89.18 *12.28 3.2 *30.93 4.27 1.7 *17.75 0.13 1.2 *18.73 0.04 0.2 0.77 0.00T5-deGARCH 12.2 *79.24 *14.25 3.3 *33.33 4.15 2.0 *25.67 0.15 1.3 *21.97 0.05 0.6 *11.52 0.00T6-deGARCH 11.7 *69.77 *16.40 3.5 *38.33 3.92 2.1 *28.53 0.16 1.3 *21.97 0.05 0.9 *23.61 0.00T7-deGARCH 11.4 *64.32 *13.82 3.7 *43.56 6.50 2.2 *31.48 0.17 1.4 *25.37 0.10 1.1 *32.85 0.04T8-deGARCH 11.2 *60.78 *15.24 4.1 *54.68 5.88 2.4 *37.65 0.18 1.6 *32.58 0.12 1.2 *37.75 0.04T9-deGARCH 11.0 *57.33 *15.71 4.2 *57.59 5.74 2.5 *40.87 0.19 1.7 *36.38 0.13 1.3 *42.83 0.05T14-deGARCH 10.6 *50.66 *16.70 4.5 *66.61 5.34 2.9 *54.53 0.44 2.0 *48.48 0.15 1.3 *42.83 0.05T19-deGARCH 10.6 *50.66 *16.70 4.9 *79.30 *8.04 2.9 *54.53 0.44 2.1 *52.73 0.16 1.5 *53.43 0.11T24-deGARCH 10.6 *50.66 *16.70 4.9 *79.30 *8.04 3.2 *65.54 4.27 2.2 *57.07 0.17 1.6 *58.94 0.12

Tab. 13: Backtesting of the VaR forecasts of the German stock portfolios: MC simulation with M = 10000, N = 100, T = 1000. *indicates the rejection at 99% level.

GHICASFMVaRplot.xpl

31

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 93: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Empirical Study 4-55

Backtesting

LR1 is the likelihood ratio w.r.t. the risk level test and LR2 w.r.t. the independence

test. * indicates the rejection at 99% level.

T(4)-deGARCH (green).

VaR plot (nig-a =0.0010)

5 10

X*E2

-1.5

-1-0

.50

0.5

Y

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 94: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

Conclusion 5-1

Conclusion and Outlook

� GHICA X

� Advanced ICA 1:Gaussian ICs (∈ IRG ) + non-Gaussian ICs (∈ IRNG ) withG >> NG

� Advanced ICA 2:Localization of ICA: yt = Wtxt

� Further GHICA application: Portfolio including bond andderivatives

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 95: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

References 6-1

Reference

Back, A. and Weigend, A.A first application of independent component analysis toextracting structure from stock returns.International Journal of Neural Systems, 8: 473-484, 1998

Barndorff-Nielsen, O.Exponentially decreasing distributions for the logarithm ofparticle size.Proceedings of the Royal Society of London, A 353: 401-419,1977.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 96: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

References 6-2

Reference

Barndorff-Nielsen, O. and Blæsild, P.Hyperbolic distribution and ramifications: Contributions totheory and applications.Statistical Distributions in Scientific Work, C. Taillie, P. Patiland A. Baldessari (eds).Vol. 4, D. Reidel, pp. 19-44, 1981.

Barndorff-Nielsen, O. and and Shephard, N.Modelling by levy processes for financial econometrics.Levy Processes : Theory and Applications, Barndorff-Nielsen,T. Mikosch and S. Resnik (eds).Birkhauser Boston, 2001.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 97: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

References 6-3

Reference

Chen, Y., Hardle, W. and Jeong, S.Nonparametric risk management with generalized hyperbolicdistributions.SFB 649, discussion paper, 2005

Christoffersen, P. F.Evaluating interval forecast.International Economic Review, 39: 841-862, 1998.

Cover, T. and Thomas, J.Elements of Information TheoryJohn Wiley & Sons, 1991.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 98: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

References 6-4

Reference

Duann, J., Jung, T., Kuo, W., Yeh, T., Makeig, S., Hsieh, J.and Sejnowski, T.Single-trial variability in event-related bold signals.NeuroImage, 15: 823-835, 2002.

Eberlein, E. and Keller, U.Hyperbolic distributions in finance.Bernoulli, 1: 281-299, 1995.

Franke, J., Hardle, W. and Hafner, C.Statistics of Financial MarketsSpringer- Verlag Berlin Heidelberg New York, 2004.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 99: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

References 6-5

Reference

Hardle, W., Kleinow, T. and Stahl, G.Applied Quantitative FinanceSpringer- Verlag Berlin Heidelberg New York, 2002.

Hyvarinen, A.New Approximations of Differential Entropy for IndependentComponent Analysis and Projection PursuitMIT Press, pp. 273-279, 1998.

Hyvarinen, A. and Oja, E.Independent component analysis: Algorithms and applicationsNeural Networks, 13: 411-430, 1999.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 100: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

References 6-6

Reference

Hyvarinen, A., Karhunen, J. and Oja, E.Independent Component AnalysisJohn Wiley & Sons, Inc. 2001.

Jaschke, S. and Jiang, Y.Approximating value at risk in conditional gaussian modelsApplied Quantitative Finance, W. Hardle, T. Kleinow and G.Stahl (eds).Springer Verlag, 2002.

Jorion, P.Value at RiskMcGraw-Hill, 2001.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy

Page 101: GHICA - Risk Analysis with GH Distributions and ...sfb649.wiwi.hu-berlin.de/fedc/events/Motzen/YChen-paper.pdf · GHICA - Risk Analysis with GH Distributions and Independent Components

References 6-7

Reference

Mercurio, D. and Spokoiny, V.Statistical inference for time inhomogeneous volatility models.The Annals of Statistics, 32: 577-602, 2004.

GHICA

Distribution comparison

-5 0 5

X

00.

10.

20.

30.

40.

5

Y

NIGLaplaceNormalt(5)Cauchy