Anda di halaman 1dari 39

Building a Minimum CVaR Portfolio

under Generalized Pareto Returns

by

Alexandre Martel

Department of Mathematics
King’s College London
The Strand, London WC2R 2LS
United Kingdom
Email: martelalex@hotmail.com
Tel: +44 (0)814 185 633
9 September 2010

Report submitted in partial fulfillment of


the requirements for the degree of MSc in
Financial Mathematics in the University of
London
Acknowledgements

I would like first of all to record a particular acknowledgement to Pr. W.T.


Shaw, who, as my supervisor, was able to take some time to answer my
persistent questions. Also, since this dissertation project was written as the
final work of a one year MSc in Financial Mathematics, I would like also to
thank the whole Department of Mathematics from King’s College London
for this enriching year : including the academic team, Dr. T. Di Matteo, Dr.
C. Albanese, Dr. C. Buescu, Dr. A. Charafi, Dr P. Emms and Dr A. Jack,
but also the administrative staff, Ms F. Benton, Miss J. Cooke and Miss S.
Rice.

1
Abstract

This paper examines portfolio construction in an extreme value framework


with Conditional Value at Risk as a risk measure. In chapter 2, some results
from Extreme Value and Copula theories are reviewed. To estimate return
distribution from a given asset in a consistent manner with the Peaks-over-
Threshold method, we create a semi parametric distribution function by
fitting a Generalized Pareto distribution on the lower and upper tails and a
Kernel distribution in the center. This is tested on the standardized daily
log return of the S&P 500 index. After some literature review on CVaR
optimization, an example of asset allocation is presented in chapter 4 using
5 stocks from the US market. A portfolio is generated from historical datas,
using the semi parametric distribution function and a copula function. The
optimal weights are found by minimizing CVaR for different level of return.
Finally, the results are compared with an optimal portfolio assuming Normal
returns.

2
Contents

1 INTRODUCTION 6

2 ASSET RETURN DISTRIBUTION 10


2.1 Extreme Value Theory . . . . . . . . . . . . . . . . . . . . . . 10
2.1.1 The Peaks-Over-Threshold methodology . . . . . . . . 11
2.1.2 Threshold selection . . . . . . . . . . . . . . . . . . . . 13
2.1.3 Parameters estimation . . . . . . . . . . . . . . . . . . 13
2.2 Copulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.1 Clayton Copula . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Data and Methodology . . . . . . . . . . . . . . . . . . . . . . 18

3 PORTFOLIO OPTIMIZATION 22
3.1 Portfolio Model . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 Conditional Value at Risk . . . . . . . . . . . . . . . . . . . . 23
3.3 Minimization of C-VaR . . . . . . . . . . . . . . . . . . . . . . 24

4 MARKET APPLICATION 27
4.1 Portfolio Simulation . . . . . . . . . . . . . . . . . . . . . . . 28
4.2 Minimization of CVaR . . . . . . . . . . . . . . . . . . . . . . 30

5 CONCLUSION 34

3
List of Tables

2.1 lower tail parameters for S&P500 : Threshold, Scale and Shape 20
2.2 upper tail parameters for S&P500: Threshold, Scale and Shape 20

4.1 lower tail parameters : Threshold, Shape and Scale . . . . . . 28


4.2 upper tail parameters : Threshold, Shape and Scale . . . . . . 29
4.3 Simulated Portfolios : Mean Return . . . . . . . . . . . . . . . 29
4.4 Simulated Portfolios : Skewness and Kurtosis . . . . . . . . . 29
4.5 EVT + Kernel Simulated Portfolio : Covariance Matrix . . . . 29
4.6 Normal Multivariate Simulated Portfolio : Covariance Matrix 30
4.7 EVT + Kernel Optimal Portfolios for β = 95%: Optimal
Portfolio, VaR and CVaR for different return level . . . . . . . 30
4.8 Multivariate Normal Optimal Portfolios for β = 95%: Optimal
Portfolio, VaR and CVaR for different return level . . . . . . . 31
4.9 EVT + Kernel Optimal Portfolios for β = 99%: Optimal
Portfolio, VaR and CVaR for different return level . . . . . . . 32
4.10 Multivariate Normal Optimal Portfolios for β = 99%: Optimal
Portfolio, VaR and CVaR for different return level . . . . . . . 32

4
List of Figures

1.1 QQ Plot of the S&P500 index standardized daily log returns


versus strandard Normal . . . . . . . . . . . . . . . . . . . . . 7
1.2 S&P500 standardized daily log returns vs Normal density . . . 8
1.3 S&P500 standardized daily log returns vs Normal density (from
-2% to -8%) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.1 Generalized Pareto density for different level of ξ . . . . . . . 12


2.2 Estimates of ξ for various thresholds on the left tail of S&P500
distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Estimates of σ for various thresholds on the left tail of S&P500
distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 3D-representation of Clayton copula for δ=10 . . . . . . . . . 17
2.5 Fitted Cumulative distribution of SP500 daily log returns . . . 19
2.6 Comparison of CDF for the lower tail . . . . . . . . . . . . . . 20
2.7 Comparison of CDF for the upper tail . . . . . . . . . . . . . 21

4.1 Daily Price Closings in base 100 . . . . . . . . . . . . . . . . . 27


4.2 Comparison of Efficient Frontier for β=95% . . . . . . . . . . 31
4.3 Comparison of Efficient Frontier for β=99% . . . . . . . . . . 33

5
Chapter 1

INTRODUCTION

“The process of selecting a portfolio may be divided in two stages. The


first stage starts with observation and experience and ends with beliefs about
the future performances of available securities. The second stage starts with
the relevant beliefs about future performances and ends with the choice of
portfolio.” Harry Markowitz, The Journal of Finance, 1952.
Despite this fundamental idea, the Modern Portfolio Theory, introduced
by H. Markowitz, has been questionned in many ways. The main criticisms
made by professionals and researchers have been on the assumptions underly-
ing the framework of this theory. To model returns from financial assets, the
well-known Normal distribution is assumed. Therefore, in a “normal” world
the variance of the portfolio is a good risk measure, and optimal portfolios
are drawn from the mean-variance criteria.
Unfortunately, historical observations show that the normal distribution
significantly underestimates the probability of having high returns. This sug-
gests an excess kurtosis, i.e. a fatter tail than with the Normal distribution.
To display this idea, a quantile-quantile plot compares the sample quantiles
versus theoretical quantiles from a normal distribution. If the sample dis-
tribtion is normal, the plot will be close to linear. In Figure 1.1, it is clear
that some deviations appear in both tails of the distribution of the S&P500
index1 , meaning that the Normal distribution fails to fit the data in both
tails. This is also visible by plotting the graph of the histogram of the re-
turns compared to the Normal density (see Figures 1.2 and 1.3).

1
datas correspond to the standardized daily log returns of the S&P500 index from
01/01/1970 to 01/01/2010 (historical prices were downloaded from Yahoo! Finance)

6
Figure 1.1: QQ Plot of the S&P500 index standardized daily log returns
versus strandard Normal

Also, since variance is a symmetric measure that takes into account neg-
ative returns as well as positive ones, one can argue that investors are only
concerned about losses and thus shall use asymetric risk measures or ”coher-
ent risk measures” instead. Thus, some new risk measures such as the Value
at Risk (VaR) or its extension, the Conditional Value at Risk (CVaR) have
been introduced in the emerging Risk Management industry back in the late
80’s.

Mathematically, given a confidence interval α between 0 and 1, and let L


be the loss function of the portfolio, the VaR at confidence level α is given
by the smallest value X such that the probability that the loss L exceeds X
is not larger than (1 − α) (see P. Jaurion [17] for a complete introduction).
Though VaR was very popular and efficient while estimating risk when re-
turns are normally distributed, it has recognized limitations in the case of
non-normal distributions. The main limit where its lack of subadditivity and

7
Figure 1.2: S&P500 standardized daily log returns vs Normal density

convexity, meaning for example that the VaR of a portfolio containing two
assets may be greater than the sum of the VaR of each asset. Another limit
is that in case the VaR is exceeded, no one knows how much the effective
loss would be.
However, the Conditional Value at Risk overcomes many of the draw-
backs of Variance and VaR as risk measures. Since it only evaluates risk on
the downside, it captures the incidence of heavy tailed distribution observed
from return distributions. Moreover, it is a coherent risk measure, and thus
would be more appropriate to incorporate in a portfolio optimization model.
This paper was written in the context of the MSc in Financial Mathemat-
ics at King’s College London. It also contributes to the growing literature on
risks from portfolio using non-gaussian distribution.Two main subjects will
be discussed: the use of Extreme Value Theory (EVT) in modeling extreme
returns and the related portfolio optimization using CVaR criteria.

8
This paper is organised as follows. Chapter 2 reviews the implications
of univariate Extreme Value Theory combined with Copulas to generate a
portfolio. In this section, we will also empirically examine and compare
how our distribution performs better in fitting historical datas. In Chapter
3, we will solved the CVaR optimization problem and show how it can be
applied to financial markets. In Chapter 4, a numerical application is given by
simulating an equity based portfolio and some comments are made. Section
5 concludes this paper.

Figure 1.3: S&P500 standardized daily log returns vs Normal density (from
-2% to -8%)

9
Chapter 2

ASSET RETURN
DISTRIBUTION

2.1 Extreme Value Theory


In Extreme Value Theory, two statistical methods are available to estimate
the tails of distribution : the Block Maxima (BM) method (developped by
Fisher and Tippett, 1928; Gnedenko, 1943) and the Peaks-Over-Threshold
(POT) method (Pickands, 1975; Balkema and de Haan, 1974). However, in
Finance, one method will be prefered to the other because of the clustering
phenomenom. Indeed, we will considered the estimation of tails distribu-
tion by the POT method with the Maximum Likelihood estimation which
possesses numerous advantages. First of all, POT method is pretty flexible
and realistic compared to BM method, which doesn’t take into account all
the possible extreme outcomes. The BM method extracts the maximum of
each period (block), previously defined (month, year, etc...). Thus, it can
miss some extreme values which could happen around the maximum of the
period (financial cycle or clustering phenomenom) whereas during the fol-
lowing period, the maximum could be relatively low. In the contrary, the
POT method avoids this kind of problem since it extracts the maximum
above a given threshold previously defined. Consequently, this method takes
into account the clustering phenomenom, representative of the financial as-
set returns. It is therefore particularly adapted to Finance, whereas, in other
domains, where cycles are absent, the BM method will be prefered (See P.
Embrechts in [11]).

10
2.1.1 The Peaks-Over-Threshold methodology

We suppose that we have some observations Z1 , ..., Zn independant and iden-


tically distributed (i.i.d.) from an unknown distribution F. We are interested
in the number Nu of extremes which exceed the high threshold u, and more
precisely on the excesses sample X1 , ..., XNu , assumed to be i.i.d., defined by
Xi = Zi − u.

Given the high threshold u on the observations, the distribution of ex-


cesses over u is given by

Fu (x) = P {Z − u ≤ x | Z > u} (2.1)

for 0 ≤ x < z0 − u where z0 ≤ ∞ is the right endpoint of F.


The excess distribution represents the probability that a loss exceeds the
threshold u by at most an amount x, given the information that it exceeds
the threshold. In terms of the underlying asset distribution

P {Z − u ≤ x, Z > u} P {Z ≤ x + u} − P {Z ≤ u}
Fu (x) = =
P {Z > u} P {Z > u}
Finally,
F (x + u) − F (u)
Fu (x) = (2.2)
1 − F (u)

We shall now introduce the Generalized Pareto distribution (GPD) by


the following essential theorem

Theorem 2.1.1. The Pickands-Balkema-de Haan Theorem Provided


that the underlying distribution F belongs to the domain of attraction1 of the
Generalised Extreme Value distribution, then ∃ a function β(u) such that
lim sup | Fu (y) − Gξ,β(u) (y) |= 0 where Gξ,β(u) denotes the Generalised
u→x0 0≤y≤x0 −u
Pareto distribution.

The Generalized Pareto distribution (GPD) with two parameters ξ and


β is defined by
1
In the literature, the domain of attraction of the Generalized Extreme Value distri-
bution corresponds to a large class of underlying distributions, containing all the most
common distributions : normal, lognormal, χ2 , t, gamma, exponential, uniform, beta,
etc...

11
(
1 − (1 + ξx/σ)−1/ξ ξ=6 0,
Gξ,σ (x) = (2.3)
1 − exp(−x/σ) ξ = 0.

where σ > 0, and where x ≥ 0


The parameter ξ defines the important shape parameter of the distri-
bution and σ is an additional scaling parameter. If ξ > 0 then Gξ,σ is a
reparametrized version of the ordinary Pareto distribution, which has a long
history in actuarial mathematics as a model for large losses; ξ = 0 corre-
sponds to the exponential distribution and ξ < 0 is known as a Pareto type
II distribution. See Figure 2.1 for graphical example.
The first case is the most relevant for risk management purposes since the
GPD is heavy-tailed when ξ > 0. Whereas the normal distribution has mo-
ments of all orders, a heavy-tailed distribution does not possess a complete
set of moments.

Figure 2.1: Generalized Pareto density for different level of ξ

12
2.1.2 Threshold selection

Choosing the optimal threshold may be a cornelian compromise between se-


lecting a sufficiently high threshold so that the Pickands-Balkema-de Haan
theorem holds and a sufficiently low threshold so that we have enough obser-
vations for the estimation of our parameters ξ and σ. Indeed, choosing a high
value for u leads to few observations of extreme returns and implies inefficient
parameter estimates with large standard errors. On the other hand, a low
value for u leads to many observations of extreme returns but induces biased
parameter estimates as observations not belonging to the tails are included
in the estimation process.
In this paper, we follow a method explained by M. Sarma in [20]. The main
idea is simply to plot the graph of the estimated parameters for different
level of threshold. Then an optimal threshold is found when the estimates
start to stabilize. Such graphs are plotted in figures (2.2) and (2.3). As an
example, for the left tail of the S&P500 daily log return distribution, it can
be seen that from 500 observations beyond the lower threshold, the estimates
seem to stabilize in both graph. The corresponding lower threshold is the 5%
quantile of the S&P500 sample, which is equaled to -1.6238. The same pro-
cedure is applied for the upper tail : the associated optimal upper threshold
is 1.7096 which is the 96% quantile of the S&P500 sample.

2.1.3 Parameters estimation

In this section, we aim to find estimators of the shape ξ and the scale σ
parameters. We will introduce here the parametric approach in the case of
the POT method2 . Under the assumption that the limit distribution holds,
the maximum likelihood method gives unbiased and asymptotically normal,
and of minimum variance, estimators. The system of non-linear equations
can be solved numerically using the Newton-Raphson iterative method (See
F.M. Longin in [17] for details).

Suppose that our excess sample X = (X1 , ..., XNu ) is i.i.d. with cumulative
distribution function the Generalized Pareto distribution G. The probability
2
To estimate parameters of the GPD, various methods such as the method of moments,
the probability weighted moments, the L-moments, Maximum Likelihood, principle of
Maximum Entropy and Least Squares method have been used. A review of all this method
can be found in [19].

13
Figure 2.2: Estimates of ξ for various thresholds on the left tail of S&P500
distribution

density function g of G is
1 x
g(x) = exp − , f or ξ = 0,
σ σ

 − 1ξ −1
1 x
g(x) = 1+ξ , f or ξ 6= 0.
σ σ

The log-likelihood is then equal to


Nu
1X
l(0, σ; X) = −Nu ln σ − Xi , f or ξ = 0,
σ i=1
 Nu
X  
1 ξ
l(ξ, σ; X) = −Nu ln σ − +1 ln 1 + Xi , f or ξ 6= 0.
ξ i=1
σ

14
Figure 2.3: Estimates of σ for various thresholds on the left tail of S&P500
distribution

By deriving those expressions in ξ and σ, we obtain the maximization


equation from which we find the Maximum Likelihood estimators (ξˆNu , σ̂Nu ).
For ξ 6= 0, we use numerical methods such as Newton-Raphson method to
find the parameters. For ξ = 0 it is straightforward since
Nu
∂ −Nu 1 X
l(0, σ; X) = + 2 Xi
∂σ σ σ i=1

which leads to
Nu
1 X
σ̂Nu = X i = XN u
Nu i=1

15
2.2 Copulas
Given d random variables (X1 , ..., Xd ) with certain marginal distributions,
copulas provide a systematic way of building joint distributions that respect
these given marginal distributions. In this section we shall establish a link
between a general joint distribution and the joint distribution of d uniform
random variables. This study will help us in order to simulate dependent
random variables from a given copula.
A d-dimensional copula C is defined as the joint distribution function
C : [0, 1]d → [0, 1] of a vector (U1 , ..., Ud ) of uniform (0,1) random variables,
that is,
C(u, v) = P (U1 ≤ u1 , ..., Ud ≤ ud ), u1 , ..., ud ∈ [0, 1].
We introduce now the following fundamental theorem, which was first proved
by Sklar (1959), it states that for any joint distribution function H there exists
a copula C that ’couples’ H to its marginal distribution functions (X1 , ..., Xd ).

Theorem 2.2.1. Sklar’s Theorem Let F be a d-dimensional cdf with con-


tinuous margins F1 , ..., Fd .Then it has the following unique copula represen-
tation: F (x1 , ..., xd ) = C [ F1 (x1 ), ..., Fd (xd ) ]

From this theorem, we can see that for continuous multivariate distri-
bution functions, the univariate margins and the multivariate dependence
structure, i.e. the copula, can be separated. Before introducing a specific
copula in the following subsection, we define the family of Archimedean cop-
ulas.

Definition 1. Archimedean Copulas are of the form

C(u1 , ..., un ) = ψ −1 [ ψ(u1 ) + ... + ψ(un ) ] f or all 0 ≤ u1 , ..., un ≤ 1

and where ψ is a function termed generator, satisfying:

1. ψ(1) = 0,

2. Decreasing. For all t ∈ (0, 1), ψ0 < 0,

3. Convex. For all t ∈ (0, 1), ψ00 ≥ 0,

16
Figure 2.4: 3D-representation of Clayton copula for δ=10

2.2.1 Clayton Copula

The Clayton copula, which is an asymmetric Archimedean copula, exhibits


greater dependence in the negative tail than in the positive. From a financial
point of view, this copula will express the trend that stocks have when they
depreciate simultaneously in times of financial stress in the markets.
This copula is given by
Cδ (u, v) = (u−δ + v −δ − 1)−1/δ ,
where 0 < δ < ∞ is a parameter controlling the dependence. Perfect depen-
dence is obtained if δ → ∞, while δ → 0 implies independance.
In her paper [1], K. Aas gives a simulation algorithm for the Clayton
copula.

• Simulate a gamma variate X ∼ Ga(1/δ, 1).

17
• Simulate d independent standard uniforms V1 , ..., Vd .
• Return U = (1 − logXV1 )− 1/δ, ..., (1 − logXVd )− 1/δ .


2.3 Data and Methodology


In order to build our multivariate distribution, we will use first of all the Clay-
ton copula to provide the correlation structure between our stock returns.
Next, for a given stock, we will use the following marginal distribution.

• a lower tail bounded by the threshold u1 (corresponding to the 5%


quantile), fitted with a Generalized Pareto distribution with parameters
(ξˆ1 , σ̂1 ),
• the interior of the distribution, between the threshold u1 and u2 (cor-
responding to the 95% quantile), estimated by a non-parametric kernel
distribution,
• an upper tail, from the threshold u2 , fitted with a Generalized Pareto
distribution with parameters (ξˆ2 , σ̂2 ),

More precisely, the marginal cumulative distribution function of our stock


return will be defined as (See Figure 2.5 for a graphical representation)

ˆ
GPD(x, u1 , ξ1 , σ̂1 ) if x ≤ u1

F (x) = Φ(x) if u1 ≤ x ≤ u2
GPD(x, u2 , ξˆ2 , σ̂2 ) if x ≥ u2

where GPD(x, u, ξ, σ) represents the cumulative distribution function of


the Generalized Pareto Distribution for a threshold u, with scale parameter
ξ and shape parameter σ. ξˆ and σ̂ are the estimated parameters (see section
2.2.2) from the historical datas for the Generalized Pareto Distribution.
Having estimated the three distinct regions of the composite semi-parametric
empirical CDF, we still need the corresponding inverse CDF which will be
of the form

−1 ˆ
GPD (v, u1 , ξ1 , σ̂1 ) if v ≤ v1

Q(v) = Φ−1 (v) if v1 ≤ v ≤ v2
GPD (v, u2 , ξˆ2 , σ̂2 ) if v ≥ v2

 −1

18
Figure 2.5: Fitted Cumulative distribution of SP500 daily log returns

where GPD−1 (v, u, ξ, σ) represents the inverse cumulative distribution


function of the Generalized Pareto Distribution for a threshold u, with scale
parameter ξ and shape parameter σ. ξˆ and σ̂ are the estimated parame-
ters (see section 2.2.2) from the historical datas for the Generalized Pareto
Distribution. Also, we have the following equalities
v1 = Φ(u1 ) = GP D(u1 , u1 , ξˆ1 , σ̂1 )
v2 = Φ(u2 ) = GP D(u2 , u2 , ξˆ2 , σ̂2 )

In tables 4.1 and 4.2 we give estimates of the scale and shape parameters
for the corresponding lower threshold and upper threshold for the S&P500
index. The positive shape estimate indicates clearly some thickness in both
left and right tail. In particular, the left tail with shape value 0.2949 is heavier
than the right tail with shape value 0.2024, showing some asymmetry in the
standardized S&P500 observations. This is confirmed by the calculation of
the skewness which has a negative value of -1.0780.

19
lower threshold scale parameter shape parameter
-1.4595 0.2949 0.6243

Table 2.1: lower tail parameters for S&P500 : Threshold, Scale and Shape

upper threshold scale parameter shape parameter


1.4159 0.2024 0.6081

Table 2.2: upper tail parameters for S&P500: Threshold, Scale and Shape

Figure 2.6: Comparison of CDF for the lower tail

Those results allow us to compute a visual representation of the CDF


and the inverse CDF for the S&P 500 sample. In particular, if we compare
the graph of the empirical CDF with those of the Normal CDF and our
constructed CDF, it can be clearly seen that our distribution fits much better
the tails (See Figures (2.6) and (2.7) for the lower and upper tail respectively

20
of the S&P500 index standardized daily log returns), and thus can performed
better risk management.

Figure 2.7: Comparison of CDF for the upper tail

In chapter 4, we use the Clayton copula for n=5 to obtain a uniform


dependant sample U = [U1 , U2 , U3 , U4 , U5 ]. For the dependance parameter, a
reasonable δ value of 0.05 is taken. By applying the inverse marginal CDF of
different asset to our uniform sample we obtain the multivariate distribution
which will be used to model a portfolio.

Y = [Y1 , Y2 , Y3 , Y4 , Y5 ] = [Q1 (U1 ), Q2 (U2 ), Q3 (U3 ), Q4 (U4 ), Q5 (U5 )]

where Qi represents the inverse marginal CDF based on parameters from


asset i.

21
Chapter 3

PORTFOLIO OPTIMIZATION

3.1 Portfolio Model


We start by considering n assets available in the financial market. Let
x = {x1 , x2 , ..., xn }T ∈ X represents the weight of stock i allocated to the
portfolio, where X is a subset of IRn and corresponds to the set of alloca-
tion constraints. For example, if we do not allow short postions X = {x1 ≥
0, x2 ≥ 0, ..., xn ≥ 0}. We also define the uncertain returns of the n assets,
from time 0 to a fixed time T, by r = {r1 , r2 , ..., rn }T ∈ IRn , with proba-
bility density function p(r). A static portfolio allocation model aims to find
the optimal portfolio x to be constructed at time 0, in order to optimize a
given criteria. In this paper, we will focus in finding the optimal portfolio x∗
minimizing the Conditional Value at Risk for a given level of portfolio return
under some allocation constraints.
The return on the portfolio is the random variable
Rp = x1 r1 + x2 r2 + ... + xn rn = xT r
The weight constraint condition is written as
n
X
xi = 1 (3.1)
i=1

We suppose that short positions are not allowed


xi ≥ 0, f or i = 1, ..., n (3.2)
Also, we impose the portfolio return to reach a given value
Rp = xT r = R∗ (3.3)

22
Finally, we can define the set of allocation constraints as

X = {xi , i = 1, ..., n, such that (3.1), (3.2) and (3.3) are verif ied} (3.4)

Finally, we define the loss function f : IRn → IR of the portfolio as

f (x/r) = −rT x (3.5)

3.2 Conditional Value at Risk


The Conditional Value at Risk with probability level β is the expected return
on the portfolio in the worst β% of the cases. This is a coherent measure
and moreover a spectral measure of financial portfolio risk. It represents the
average loss value given that the loss already exceeded the VaR, i.e. the
average loss on the tail of the distribution of the portfolio.
Definition 2. Coherent Risk Measure A real valued function ρ of a ran-
dom variable is a coherent risk measure if it satisfies the following properties,

1. Subadditivity. For any two random variables X and Y, ρ(X + Y ) ≤


ρ(X) + ρ(Y ),
2. Monotonicity. For any two random variables X ≥ Y, ρ(X) ≥ ρ(Y ),
3. Positive homogeneity. For λ ≥ 0, ρ(λX) = λρ(X),
4. Translation invariance. For any a ∈ IR, ρ(a + X) = a + ρ(X).

The recent theory of spectral risk measures have been introduced by C.


Acerbi in [2]. Spectral risk measures define a class of measures based on
integrals of the quantile function of the portfolio return. It consists in a
weighted average of the quantiles of the distribution of the returns, using
a non-increasing weight function called the spectrum. It can be viewed as
weighted averages of Values at Risk.
Definition 3. Spectral Risk Measure Consider a portfolio X. There are
S equiprobable outcomes with the corresponding payoffs given by the order
statistics X1:S , ..., XS:S . Let φ ∈ IRS . The measure Mφ : IRS → IR defined
by Mφ (X) = −δ Ss=1 φs Xs:S is a spectral measure of risk if φ ∈ IRS satisfies
P
the conditions

1. Nonnegativity. φs ≥ 0 for all s = 1, ..., S,

23
PS
2. Normalization. s=1 φs = 1,
3. Monotonicity. φs is non-increasing, that is φs1 ≥ φs2 if s1 ≤
s2 and s1, s2 ∈ {1, ..., S}.

3.3 Minimization of C-VaR


This technique was first introduced by S. Uryasev and R. Tyrrell Rockafellar
in 1999 [23]. For (x, α) ∈ X × IR, the probability that the loss function
f (x/r) does not exceed some threshold value α is given by
Z
ψ(x, α) = p(r)dr (3.6)
f (x/r)≤α

In what follows, ψ(x, α) is assumed to be everywhere continuous with


respect to α
The associated Value at Risk (VaR) with probability level β ∈ (0, 1) is
V aRβ (x) = min{α : ψ(x, α) ≥ β} (3.7)

While the associated Conditional-VaR (CVaR) with probability level β ∈


[0, 1] is
CV aRβ (x) = E[f (x/r)/f (x/r) ≥ V aRβ (x)] (3.8)
The calculation leads to
R
V aRβ (x)≤f (x/r)
f (x/r)p(r)dr
CV aRβ (x) = R
V aRβ (x)≤f (x/r)
p(r)dr

where the denominator is the probability that the loss function exceeds the
V aRβ (x) which is equaled by definition to 1 − β.
Z
1
CV aRβ (x) = f (x/r)p(r)dr
1 − β V aRβ (x)≤f (x/r)
Which can be written as
Z
1
CV aRβ (x) = V aRβ (x) + (f (x/r) − V aRβ (x))+ p(r)dr (3.9)
1−β
Since: Z
V aRβ (x)
V aRβ (x) = p(r)dr
1−β V aRβ (x)≤f (x/r)

24
(
u if u > 0
where (u)+ =
0 if u ≤ 0

Next, from Equation (3.9) we can defined the following objective function
Fβ (x, α) on X × IR, with alpha a parameter:
Z
1
Fβ (x, α) = α + (f (x/r) − α)+ p(r)dr (3.10)
1−β

The idea behind this objective function is drawn from the two following
theorems. For proves, see [23].
Theorem 3.3.1. As a function of α, Fβ (x, α) is convex and continuously
differentiable. The CV aRβ of the loss associated with any x ∈ X can be
determined from the formula

φβ (x) = min Fβ (x, α)


α∈IR

In this formula the set consisting of the values of α for which the minimum
is attained, namely
Aβ (x) = arg min Fβ (x, α)
α∈IR

is a nonempty, colsed, bounded interval (perhaps reducing to a single point),


and the V aRβ of the loss is given by

αβ (x) = lef t endpoint of Aβ (x).

In particular, on always has

αβ (x) ∈ arg min Fβ (x, α) and φβ (x) = Fβ (x, αβ (x)).


α∈IR

Theorem 3.3.2. Minimizing the CV aRβ of the loss associated with x over
all x ∈ X is equivalent to minimizing Fβ (x, α) over all (x, α) ∈ X × IR, in
the sense that
min φβ (x) = min Fβ (x, α),
x∈X (x,α)∈X×IR

where moreover a pair (x∗ , α∗ ) achieves the second minimum if and only if
x∗ achieves the first minimum and α∗ ∈ Aβ (x∗ ). In particular, therefore,
in circumstances where the interval Aβ (x∗ ) reduces to a single point (as is
typical), the minimization of Fβ (x, α) over (x, α) ∈ X × IR produces a pair
(x∗ , α∗ ) not necessarily unique, such that x∗ minimizes the CV aRβ and α∗

25
gives the corresponding V aRβ . Furthermore, Fβ (x, α) is convex with respect
to (x, α), and φβ (x) is convex with respect to x, when f (x/r) is convex with
respect to x, in which case, if the constraints are such that X is a convex set,
the joint minimization is an instance of convex programming.

Assuming that the loss function f (x/r) is convex with respect to (w.r.t.)
x, the function Fβ (x, α) is convex w.r.t. x. Also, it can be verified that
Equation (3.10) is linear and convex w.r.t. α, see [23] for proof. Finally, we
see that Equation (3.10) is convex with respect to (x, α) if the loss f (x, y)
is convex with respect to x. Therefore, those two theorems explain how the
CVaR minimization problem can be reduced to a simpler continuously dif-
ferentiable and convex function optimization problem.

Moreover, it appears that, by using the objective function Fβ , it is not needed


to calculate first the V aRβ , which would be more complicated.

Next, for practical cases, the integral in Equation (3.10) can be approxi-
mated using scenari rj , j = 1, ..., J which are sampled using the density func-
tion p. The set of the generated scenari can be represented from J vectors
r1 , r2 , ..., rJ . Hence, Equation (3.10) becomes:
J
1 1X
F̃β (x, α) ≈ α + (f (x/ri ) − α)+ (3.11)
1 − β J j=1

Now, assuming that the set of allocation constraints X is convex, we


shall solve the convex optimization problem

min F̃β (x, α) (3.12)


(x,α)∈X×IR

Following the two previous theorems, there exists (x∗ , α∗ ) solution to


(3.12). x∗ represents the optimal portfolio vector and the corresponding
VaR is equaled to α∗ .

26
Chapter 4

MARKET APPLICATION

Figure 4.1: Daily Price Closings in base 100

In this section, we aim to construct a portfolio of 5 stocks and find the


optimal weights in order to minimize the CVaR, by following the procedure

27
seen in the previous chapters. We shall start by simulating our portfolio,
that is finding the multivariate distribution using the Clayton copula and
the semi-parametric distribution based on historical prices, as introduced in
chapter 2. Then, we will run the linear optimization, studied in chapter 3,
under the associated portfolio allocation constraints. Some comments will
be made by comparing the results with a minimum CVaR portfolio using a
multivariate normal distribution for the stock returns.

4.1 Portfolio Simulation


Our portfolio will be based on 5 stocks : Coca Cola Company (ticker KO),
Ebay Inc.(EBAY), Intel Corpotation (INTC), Goldman Sachs Group Inc.
(GS) and Exxon Mobil Corp.(XOM). The data corresponds to the period
01/01/2003 to 01/01/2010 for a length of N=1762 daily log returns and were
downloaded from Yahoo! Finance. For subsequent examination, figure (4.1)
illustrates the price movements of each stock. The initial level of each stock
has been normalized in base 100 to facilitate the comparison.
In tables (4.1) and (4.2), we give estimates of the scale and shape pa-
rameters for the corresponding lower threshold and upper threshold for each
stock. Positive shape estimates clearly indicate some heavy-tailed distribu-
tion for each stock. Then, we simulate two portfolios of 5 stocks of size
J=5000. The first portfolio is based on the Clayton copula and the semi
parametric distribution and the second one is based on the Clayton copula
and normal distribution. For the Clayton copula, a reasonable delta value of
0.05 is used. In tables (4.3), (4.4), (4.5) and (4.6) some basic statistics are
computed for the simulated portfolios.
The MATLAB codes for the inverse semi parametric CDF and the Clay-
ton copula are availlable in the Appendix.

Asset lower threshold shape parameter scale parameter


KO -1.4595 0.1496 0.7455
EBAY -1.3984 0.4065 0.4899
INTC -1.5574 0.1428 0.7048
GS -1.3602 0.2790 0.7323
XOM -1.4293 0.2099 0.7306

Table 4.1: lower tail parameters : Threshold, Shape and Scale

28
Asset upper threshold shape parameter scale parameter
KO 1.4159 0.2409 0.7043
EBAY 1.5521 0.2164 0.6173
INTC 1.5666 0.0744 0.6237
GS 1.3373 0.4065 0.6504
XOM 1.2821 0.4938 0.4601

Table 4.2: upper tail parameters : Threshold, Shape and Scale

Stock EVT + Kernel Multivariate Normal


Mean Return Mean Return
KO 0.0073 0.0147
EBAY -0.0481 -0.0526
INTC -0.0406 -0.0451
GS -0.0008 0.0374
XOM 0.0857 0.0750

Table 4.3: Simulated Portfolios : Mean Return


Stock EVT + Kernel EVT + Kernel Multivariate Normal Multivariate Normal
Skewness Kurtosis Skewness Kurtosis
KO -0.0510 14.7517 -0.0398 3.1084
EBAY -1.1042 29.5859 0.0190 2.9521
INTC -0.1315 6.2294 0.0058 2.8875
GS -1.6640 128.2658 -0.0645 3.1883
XOM 1.2944 26.1583 0.0377 3.1171

Table 4.4: Simulated Portfolios : Skewness and Kurtosis


KO EBAY INTC GS XOM
KO 1.7565 0.1173 0.1340 0.0994 0.1556
EBAY 0.1173 6.8625 0.2930 0.2798 0.2992
INTC 0.1340 0.2930 4.5394 0.1513 0.2185
GS 0.0994 0.2798 0.1513 12.2104 0.2950
XOM 0.1556 0.2992 0.2185 0.2950 3.3882

Table 4.5: EVT + Kernel Simulated Portfolio : Covariance Matrix

29
KO EBAY INTC GS XOM
KO 1.6794 0.1465 0.1223 0.0304 0.1125
EBAY 0.1465 6.3459 0.2631 0.3247 0.2927
INTC 0.1223 0.2631 4.5908 0.1665 0.1771
GS 0.0304 0.3247 0.1665 7.5439 0.1652
XOM 0.1125 0.2927 0.1771 0.1652 3.0105

Table 4.6: Normal Multivariate Simulated Portfolio : Covariance Matrix

4.2 Minimization of CVaR


From the results (see tables (4.7), (4.8), (4.9) and (4.10)), it can be seen that
stocks with low volatility and positive skewness are prefered. Indeed Coca
Cola and Exxon Mobil represent the most important weights in the optimal
portfolios but show the smallest volatility and the higher skewness. Since
kurtosis depends on the behaviour of both the peakedness of the tails and
the center of the distribution, no relation can be conclude.
Since the left tail of the distribution is underestimated in the Normal
case, it is obvious that the CVaR and VaR should be smaller than in the
semi-parametric case. While for the confidence level β=95%, the CVaR and
VaR are almost identical for both portfolios, for β=99% there is evidence of
a clear inflation of the CVaR while going from the Normal distribution to
the hybrid distribution. This can also be seen from the efficient frontiers in
(4.2) and (4.3).
For the numerical experiment, the MATLAB script for the CVaR port-
folio optimization is available from M. Vogiatzoglou in [24].

Return KO EBAY INTEL GSAC EXXON VaR CVaR


-0.02 0.4339 0.1027 0.1673 0.0572 0.2390 1.4660 2.2267
0 0.4322 0.1027 0.1679 0.0572 0.2399 1.4661 2.2267
0.02 0.4274 0.0755 0.1411 0.0495 0.3065 1.4651 2.2503
0.04 0.3924 0.0215 0.0744 0.0306 0.4810 1.5595 2.5036
0.06 0.3231 0 0 0.0041 0.6728 1.8022 2.9778
0.07 0.1999 0 0 0 0.8001 1.9865 3.3258

Table 4.7: EVT + Kernel Optimal Portfolios for β = 95%: Optimal Portfolio,
VaR and CVaR for different return level

30
Return KO EBAY INTEL GSAC EXXON VaR CVaR
-0.02 0.4176 0.1010 0.1481 0.1019 0.2315 1.4793 1.9017
0 0.4174 0.1010 0.1482 0.1019 0.2316 1.4791 1.9017
0.02 0.4264 0.0892 0.1172 0.1026 0.2647 1.4728 1.9093
0.04 0.4476 0.0007 0.0315 0.1105 0.4097 1.6210 2.1053
0.06 0.1907 0 0 0.0935 0.7158 2.1196 2.7058
0.07 0.0315 0 0 0.0827 0.8858 2.4300 3.2009

Table 4.8: Multivariate Normal Optimal Portfolios for β = 95%: Optimal


Portfolio, VaR and CVaR for different return level

Figure 4.2: Comparison of Efficient Frontier for β=95%

31
Return KO EBAY INTEL GSAC EXXON VaR CVaR
-0.02 0.4204 0.0790 0.2422 0.0177 0.2408 2.7668 3.4541
0 0.4227 0.0764 0.2426 0.0177 0.2406 2.7712 3.4539
0.02 0.3915 0.0689 0.1987 0.0080 0.3329 2.7346 3.5552
0.04 0.4112 0.0139 0.0896 0.031 0.4821 3.0511 4.1458
0.06 0.3276 0 0 0 0.6724 3.7722 5.1115
0.07 0.1999 0 0 0 0.8001 4.1744 5.8798

Table 4.9: EVT + Kernel Optimal Portfolios for β = 99%: Optimal Portfolio,
VaR and CVaR for different return level

Return KO EBAY INTEL GSAC EXXON VaR CVaR


-0.02 0.4365 0.0949 0.1622 0.0793 0.2271 2.1429 2.5219
0 0.4364 0.0959 0.1613 0.0804 0.2260 2.1458 2.5219
0.02 0.4262 0.0716 0.1348 0.0743 0.2731 2.1889 2.5451
0.04 0.4593 0.0003 0.0380 0.0726 0.4299 2.4183 2.8414
0.06 0.2112 0 0 0.0606 0.7282 3.1206 3.6157
0.07 0.0328 0 0 0.0806 0.8865 3.6347 4.2465

Table 4.10: Multivariate Normal Optimal Portfolios for β = 99%: Optimal


Portfolio, VaR and CVaR for different return level

32
Figure 4.3: Comparison of Efficient Frontier for β=99%

33
Chapter 5

CONCLUSION

The paper considered the well-studied optimal portfolio problem. In particu-


lar, we decided to minimize the coherent CVaR risk measure of the portfolio.
In addition, we used a semi-parametric distribution to model the stock re-
turns, based on a mixture of two distributions : the Generalized Pareto
distribution and the Kernel distribution following the Extreme Value theory.
In chapter 2, the Peaks-over-Threshold method was explained, showing how
the tails can be estimated by finding first a consistent threshold, and then
the shape and scale parameters by a Maximum Likelihood method. We il-
lustrated how our hybrid distribution fits historical data much better than
the normal distribution. We used the standardized daily-log return of the
S&P 500 as an example. Chapter 3 explained the CVaR optimization process
following the paper from S. Uryasev and R.T. Rockafellar in [23]. This was
done by introducing an objective function and some portfolio constraints un-
der which the optimization problem reduces to a simpler linear programming
problem. Finally, we conducted a numerical experiment based on historical
datas from 5 stocks of the US market. We constructed two portfolios, one
using the semi parametric distribution and one using a normal distribution.
We also proposed the inclusion of the Clayton copula, introduced in chapter
2, into the simulation process to model the correlation structure between the
stocks. After optimizing our portfolios, some optimal weights were found
for each portfolio. A slightly different capital allocation was obtained, with
higher values of portfolio CVaR in the semi parametric case. Indeed, by
incorporating non-normality in the portfolio allocation process, one shall be
closer to the reality and thus observes greater downside risk than in the ’nor-
mal’ world. Also, using a coherent risk measure such as CVaR may help in
a better understanding of Risk Management.

34
Bibliography

[1] K. Aas Modelling the dependence structure of financial assets : A sur-


vey of four copulas, Norwegian Computing Center, 2004.
[2] C. Acerbi Spectral measures of risk: A coherent representation of sub-
jective risk aversion, Journal of Banking and Finance (Elsevier) 26:
15051518, 2002.
[3] M. Bohdaova and O. Nanasiova A Note to Copula Functions, 2006.
[4] Y. Bensalah Asset Allocation using Extreme Value Theory, Bank of
Canada, ISSN 1192-5434, 2002.
[5] B. O. Bradley and M. S. Taqqu An Extreme Value Theory Ap-
proach to the Allocation of Multiple Assets, International Journal of
Theoretical and Applied Finance, Vol. 7, No. 8 (2004) 10311068, 2004.
[6] U. Cherubini, E. Luciano and W. Vecchiato Copula methods in
finance, Wiley, 2004.
[7] S. Ciliberti, I. Kondor and M. Mezard On the Feasibility of
Portfolio Optimization under Expected Shortfall, 2006.
[8] E. De Giorgi A Note on Portfolio Selection under Various Risk Mea-
sures, FINRISK, 2002.
[9] A. Di Clemente and C. Romano Beyond Markowitz : Building Op-
timal Portfolio Using Non-Elliptical Asset Return Distributions, 2003.
[10] Z. Eksi, I. Yildirim and K. Yildirak Alternative Risk Measures
and Extreme Value Theory in Finance : Implementation on ISE 100
Index, 1999.
[11] P. Embrechts, C. Kluppelberg and T. Mikosch Modelling ex-
tremal events for insurance and finance, Springer, ISBN-10 3540609318,
1997.

35
[12] P. Embrechts, F. Lindskog and A. McNeil Modelling Depen-
dence with Copulas and Applications to Risk Management, 2001.

[13] A. A. Jobst Loss Distribution Modelling of a Credit Portfolio through


Extreme Value Theory, 2002.

[14] E. Jondeau, S. Poon and M. Rockinger Financial Modeling Under


Non-Gaussian Distributions, Springer, 2007.

[15] P. Jorion Value at Risk, Second edition, McGraw Hill, New York,
2001.

[16] S. Kotz and S. Nadarajah Extreme Value Distributions, Theory and


Applications, Imperial College Press, 2000.

[17] F.M. Longin The Asymptotic Distribution of Extreme Stock Market


Returns, The Journal of Business, Vol. 69, No. 3, pp. 383-408, Jul. 1996.

[18] A. J. Neil and T. Saladin The Peaks over Thresholds Method for
Estimating High Quantiles of Loss Distributions, 2007.

[19] T. Oztekin Comparison of Parameter Estimation Methods for the


Three-Parameter Generalized Pareto Distribution, 2004.

[20] M. Sarma Characterisation of the tail behaviour of financial returns :


studies from India, 2005.

[21] R. L. Smith Measuring Risk with Extreme Value Theory, 1998.

[22] R. Suarez Araya Improving Modeling of Extreme Events using Gen-


eralized Extreme Value Distribution or Generalized Pareto Distribution
with Mixing Uncondtional Disturbances, 2009.

[23] S. Uryasev and R. Tyrrell Rockafellar Optimization of Con-


ditional Value at Risk, 1999.

[24] M. Vogiatzoglou CVaR Optimization, MATLAB Central, 2002.


http: // www. mathworks. com/ matlabcentral/ fileexchange/
19907-cvar-optimization

36
Appendix

inverse of semi parametric CDF

function [x] = invGepadiCDF(u,index)


x=zeros(numel(u),1);
upperQuantile=quantile(index,.95);
lowerQuantile=quantile(index,.05);
z=sort(index);
isInUpperTail=z>upperQuantile;
isInLowerTail=z<lowerQuantile;
upperTail=z(isInUpperTail);
lowerTail=z(isInLowerTail);

lowerTailParam=gpfit(abs(lowerTail-lowerQuantile));
upperTailParam=gpfit(upperTail-upperQuantile);
vi=linspace(0,1, numel(index));
g = ksdensity(index,vi,’function’,’icdf’);

for i=1:numel(u)
if (u(i)<0.05) % lower Tail : Generalized pareto distribution
x(i)=lowerQuantile-gpinv(1-u(i)/.05,lowerTailParam(1),lowerTailParam(2));
end
if ((u(i)>=0.05)&&(u(i)<=0.95)) % center : kernel cdf
x(i) = interp1(vi,g,u(i));
end
if (u(i)>0.95) % upper Tail : Generalized pareto distribution
x(i)=upperQuantile+gpinv((u(i)-.95)/.05,upperTailParam(1),upperTailParam(2));
end
end
end

37
Clayton copula for n=5

function [u] = claycop(delta)


x=gamrnd(1/delta,1);
v1=unifrnd(0,1);
v2=unifrnd(0,1);
v3=unifrnd(0,1);
v4=unifrnd(0,1);
v5=unifrnd(0,1);
u1=(1 − log(v1)/x)ˆ(−1/delta);
u2=(1 − log(v2)/x)ˆ(−1/delta);
u3=(1 − log(v3)/x)ˆ(−1/delta);
u4=(1 − log(v4)/x)ˆ(−1/delta);
u5=(1 − log(v5)/x)ˆ(−1/delta);
u=[u1,u2,u3,u4,u5];
end

38

Anda mungkin juga menyukai