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
1
Abstract
2
Contents
1 INTRODUCTION 6
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
List of Figures
5
Chapter 1
INTRODUCTION
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.
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
10
2.1.1 The Peaks-Over-Threshold methodology
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)
11
(
1 − (1 + ξx/σ)−1/ξ ξ=6 0,
Gξ,σ (x) = (2.3)
1 − exp(−x/σ) ξ = 0.
12
2.1.2 Threshold selection
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
14
Figure 2.3: Estimates of σ for various thresholds on the left tail of S&P500
distribution
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 ).
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.
1. ψ(1) = 0,
16
Figure 2.4: 3D-representation of Clayton copula for δ=10
17
• Simulate d independent standard uniforms V1 , ..., Vd .
• Return U = (1 − logXV1 )− 1/δ, ..., (1 − logXVd )− 1/δ .
18
Figure 2.5: Fitted Cumulative distribution of SP500 daily log returns
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
Table 2.2: upper tail parameters for S&P500: Threshold, Scale and Shape
20
of the S&P500 index standardized daily log returns), and thus can performed
better risk management.
21
Chapter 3
PORTFOLIO OPTIMIZATION
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)
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}.
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
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
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.
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
26
Chapter 4
MARKET APPLICATION
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.
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
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 4.7: 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
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
32
Figure 4.3: Comparison of Efficient Frontier for β=99%
33
Chapter 5
CONCLUSION
34
Bibliography
35
[12] P. Embrechts, F. Lindskog and A. McNeil Modelling Depen-
dence with Copulas and Applications to Risk Management, 2001.
[15] P. Jorion Value at Risk, Second edition, McGraw Hill, New York,
2001.
[18] A. J. Neil and T. Saladin The Peaks over Thresholds Method for
Estimating High Quantiles of Loss Distributions, 2007.
36
Appendix
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
38