See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/24086160

On the Feasibility of Portfolio Optimization under Expected Shortfall

Article  in  Quantitative Finance · August 2007

DOI: 10.1080/14697680701422089 · Source: RePEc

CITATIONS READS
25 78

3 authors, including:

Imre Kondor Marc Mezard

Eötvös Loránd University Ecole Normale Supérieure de Paris
84 PUBLICATIONS   1,814 CITATIONS    240 PUBLICATIONS   16,009 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

analytic approach to portfolio optimization under various constraints View project

The measurement of shadow banking View project

All content following this page was uploaded by Marc Mezard on 05 June 2014.

The user has requested enhancement of the downloaded file.

 Quantitative Finance published in "Quantitative Finance 7, 04 (2007) 389-396"
Author manuscript,
DOI : 10.1080/14697680701422089

On the Feasibility of Portfolio Optimization under Expected

Shortfall
Fo

Journal: Quantitative Finance

rP
Manuscript ID: RQUF-2006-0108.R1

Manuscript Category: Research Paper

peer-00568435, version 1 - 23 Feb 2011

Date Submitted by the

28-Jan-2007
ee
Author:

Complete List of Authors: ciliberti, stefano; univ paris sud, LPTMS

Kondor, Imre
Mezard, Marc
rR

Portfolio Optimization, Risk Measures, Statistical Physics, Financial

Keywords:
Applications

JEL Code:
ev

Note: The following files were submitted by the author for peer review, but cannot be converted
ie

to PDF. You must view these files (e.g. movies) online.

resubmission.tgz
w
On
ly

E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf

 On the Feasibility of Portfolio Optimization under Expected Shortfall
Page 2 of 22 Quantitative Finance
Stefano Ciliberti,1 Imre Kondor,2 and Marc Mézard1
1
1 CNRS; Univ. Paris Sud, UMR8626, LPTMS, ORSAY CEDEX, F-91405
2
Collegium Budapest, 1014 Budapest, Szenthromsg u. 2
2
3 We address the problem of portfolio optimization under the simplest coherent risk measure, i.e.
4 the expected shortfall. As it is well known, one can map this problem into a linear programming
5 setting. For some values of the external parameters, when the available time series is too short,
6 the portfolio optimization is ill posed because it leads to unbounded positions, infinitely short on
some assets and infinitely long on some others. As first observed by Kondor and coworkers, this
7
phenomenon is actually a phase transition. We investigate the nature of this transition by means
8 of a replica approach.
9
10
11 I. INTRODUCTION
12
13 Among the several risk measures existing in the context of portfolio optimization, expected shortfall (ES) has
14 certainly gained increasing popularity in recent years. In several practical applications, ES is starting to replace the
Fo
15 classical Value-at-Risk (VaR). There are a number of reasons for this. For a given threshold probability β, the VaR
16 is defined so that with probability β the loss will be smaller than VaR. This definition only gives the minimum loss
17 one can reasonably expect but does not tell anything about the typical value of that loss that can be measured by the
18 conditional value-at-risk (CVaR, which is the same as ES for continuous distributions that we consider here [12]). We
rP
19 will be more precise on these definitions below. The point we want to stress here is that the VaR measure, lacking the
20 mandatory properties of subadditivity and convexity,is not coherent [1]. This means that summing VaR’s of individual
21 portfolios will not necessarily produce an upper bound for the VaR of the combined portfolio, thus contradicting the
peer-00568435, version 1 - 23 Feb 2011

22 holy principle of diversification in finance. A nice practical example of the inconsistency of VaR in credit portfolio
ee
23 management is reported in Ref. 3. On the other hand, it has been shown [2] that ES is a coherent measure with
24 interesting properties [4]. Moreover, the optimization of ES can be reduced to linear programming [5] (which allows
25 for a fast implementation) and leads to a good estimate for the VaR as a byproduct of the minimization process. To
26 summarize, the intuitive and simple character, together with the mathematical properties (coherence) and the fast
rR

27 algorithmic implementation (linear programming), are the main reasons behind the growing importance of ES as a
28 risk measure.
29 In this paper, we will focus on the feasibility of the portfolio optimization problem under the ES measure of risk.
30 The control parameters of this problem are (i) the imposed threshold in probability, β, and (ii) the ratio N/T between
ev

31 the number N of financial assets making up the portfolio and the time series length T used to sample the probability
32 distribution of returns. (It is curious that, albeit trivial, the scaling in N/T had not been explicitly pointed out before
33 [11]. It has been discovered by Kondor at al. Ref. 6 that, for certain values of these parameters, the optimization
34 problem does not have a finite solution because, even if convex, it is not bounded from below. Extended numerical
ie

35 simulations allowed these authors to determine the feasibility map of the problem. Here, in order to better understand
36 the root of the problem and to study the transition from a feasible regime to an unfeasible one (corresponding to an
37 ill-posed minimization problem) we address the same problem from an analytical point of view.
w

38 The paper is organized as follows. In Section II we briefly recall the basic definitions of β-VaR and β-CVaR and we
39 show how the portfolio optimization problem can be reduced to linear programming. We introduce a “cost function”
40 to be minimized under linear constraints and we discuss the rationale for a statistical mechanics approach. In Section
On

41 III we solve the problem of optimizing large portfolios under ES using the replica approach. Our results and the
42 comparison with numerics are reported in Section IV, and our conclusions are summarized in Section V.
43
44
45 II. THE OPTIMIZATION PROBLEM
ly

46
47 We consider a portfolio of N financial instruments w = {w1 , . . . wN }, where wi is the position of asset i. The global
48 budget constraint fixes the sum of these numbers: we impose for example
49 N
50 X
wi = N . (1)
51
i=1
52
53 We do not stipulate any constraint on short selling, so that wi can be any negative or positive number. This is, of
54 course, unrealistic for liquidity reasons, but considering this case allows us to show up the essence of the phenomenon.
55
56
57
58
59
60
E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf
 1 Quantitative Finance Page 3 of 22
β
1
2
3

P< (w; α)
4
5
6
7
8
9
10
11
12 0
13 β-VaR(w) α
14
Fo
15
16
17 FIG. 1: Schematic representation of the VaR measure of risk. P< (w) is the probability of a loss associated to the portfolio w
18 being smaller than α. The conditional VaR β-CVaR (or ES) is the average loss when this is constrained to be greater than the
rP
19 β-VaR.
20
21
peer-00568435, version 1 - 23 Feb 2011

22 If we imposed a constraint that would render the domain of the wi bounded (such as a ban on short selling, e.g.), this
would evidently prevent the weights from diverging, but a vestige of the transition would still remain in the form of
ee
23
24 large, though finite, fluctuations of the weights, and in a large number of them sticking to the ”walls” of the domain.
25 We denote the returns on the assets by x = {x1 , x2 , . . . xN } , and we assume that there exists an underlying proba-
bility distribution function p(x) of the returns. The loss of portfolio w given the returns x is ℓ(w|x) = − N
P
26 i=1 wi xi ,
rR

27 and the probability of that loss being smaller than a given threshold α is
28 Z


29 P< (w, α) = dx p(x)θ α − ℓ(w|x) , (2)
30
ev

31 where θ(·) is the Heaviside step function, equal to 1 if its argument is positive and 0 otherwise. The β-VaR of this
32 portfolio is formally defined by
33
34 β-VaR(w) = min{α : P< (w, α) ≥ β} , (3)
ie

35
36 (see Fig. 1), while the CVaR (or ES, in this case) associated with the same portfolio is the average loss on the tail of
the distribution,
37
w

38
Z


39 dx p(x)ℓ(w|x)θ ℓ(w|x) − β-VaR(w)
1
Z


40 β-CVaR(w) = Z = dx p(x)ℓ(w|x)θ ℓ(w|x) − β-VaR(w) . (4)
1−β
On

41 dx p(x)θ ℓ(w|x) − β-VaR(w)
42
43 The threshold β then represents a confidence level. In practice, the typical values of β which one considers are
44 β = 0.90, 0.95, and 0.99, but we will address the problem for any β ∈ [0, 1]. What is usually called “exceeding
45 probability” in previous literature would correspond here to (1 − β).
ly

46 As mentioned in the introduction, the ES measure can be obtained from a variational principle [5]. The minimization
47 of a properly chosen objective function leads directly to (4):
48
49 β-CVaR(w) = min Fβ (w, v) , (5)
v
50
Z
 +
51 Fβ (w, v) ≡ v + (1 − β)−1 dx p(x) ℓ(w|x) − v . (6)
52
53 Here, [a]+ ≡ (a + |a|)/2. The external parameter v over which one has to minimize is claimed to be relevant in
54 itself [5], since its optimal value may represent a good estimate for the actual value-at-risk of the portfolio. We will
55
56
57
58
59
60
E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf
 come back to this point as we discuss our results. We stress here that minimizing (6) over w and v is equivalent to
Page 4 of optimizing
22 (4) over the portfolio vectors w. Quantitative Finance
Of course, in practical cases the probability distribution of the loss is not known and must be inferred from the
past data. In other words, we need an “in-sample” estimate of the integral in (6), which would turn a well posed
1 (but useless) optimization problem into a practical approach. We thus approximate the integral by sampling the
2 probability distributions of returns. If we have a time series x(1) , . . . x(T ) , our objective function becomes simply
3
4 T T
" N
#+
1 X  (τ )
+ 1 X X
5 F̂β (w, v) = v + ℓ(w|x ) − v = v + −v − wi xiτ , (7)
6 (1 − β)T τ =1 (1 − β)T τ =1 i=1
7
8 where we denote by xiτ the return of asset i at time τ .
9 Minimizing this risk measure is the same as the following linear programming problem:
10 • given one data sample, i.e. a matrix xiτ , i = 1, . . . N , τ = 1, . . . T ,
11
12 • minimize the cost function
13
T
14     X
Eβ Y; {xiτ } = Eβ v, {wi }, {uτ }; {xiτ } = (1 − β)T v + uτ , (8)
Fo
15
t=τ
16
17
• over the (N + T + 1) variables Y ≡ {w1 , . . . wN u1 , . . . uT v,
18
rP
19 • under the (2T + 1) constraints
20
21 X N X N
peer-00568435, version 1 - 23 Feb 2011

22 uτ ≥ 0 , uτ + v + xiτ wi ≥ 0 ∀τ , and wi = N . (9)

ee
23 i=1 i=1
24
25 Since we allow short positions, not all the wi are positive, which makes this problem different from standard linear
26 programming. To keep the problem tractable, we impose the condition that wi ≥ −W , where W is a very large cutoff,
rR

27 and the optimization problem will be said to be ill-defined if its solution does not converge to a finite limit when
28 W → ∞. It is now clear why constraining all the wi to be non-negative would eliminate the feasibility problem: a finite
29 solution will always exist because the weights are by definition bounded, the worst case being an optimal portfolio
30 with only one non-zero weight taking care of the total budget. The control parameters that govern the problem are
the threshold β and the ratio N/T of assets to data points. The resulting “phase diagram” is then a line in the β-N/T
ev

31
32 plane separating a region in which, with high probability, the minimization problem is not bounded and thus does not
admit a finite solution, and another region in which a finite solution exists with high probability. These statements
33
are non-deterministic because of the intrinsic probabilistic nature of the returns. We will address this minimization
34
problem in the non-trivial limit where T → ∞, N → ∞, while N/T stays finite. In this “thermodynamic” limit, we
ie

35
shall assume that extensive quantities (like the average loss of the optimal portfolio, i.e. the minimum cost function) do
36
not fluctuate, namely that their probability distribution is concentrated around the mean value. This “self-averaging”
37
w

property has been proved for a wide range of similar statistical mechanics models [7]. Then, we will be interested
38
in the average value of the minimum of the cost function 8) over the distribution of returns. Given the similarity of
39
portfolio optimization with the statistical physics of disordered systems, this problem can be addressed analytically
40
by means of a replica approach [8].
On

41
42
43 III. THE REPLICA APPROACH
44
45
ly

For a given history of returns xit , one can compute the minimum of the cost function, minY Eβ [Y; {xit }]. In this
46
section we show how to compute analytically the expectation value of this quantity over the histories of return. For
47 simplicity we shall keep to the case in which the xit are independent identically distributed (iid) normal variables, so
48 that an history of returns xit is drawn from the distribution
49
50 2
Y
p({xit }) ∼ e−N xit /2 . (10)
51 it
52
53 This assumption of iid normal distribution of returns is very restrictive, but we would like to emphasize that the
54 method that we use can be generalized easily to iid variables with other distributions, and also in some cases to
55
56
57
58
59
60
E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf
 correlated variables. Certainly the precise location of the critical value of N/T separating an unfeasible phase from a
Quantitative
feasible one depends on the distribution of returns. But we Finance
expect that the broad features like the existence of Page
this 5 of 22
critical value, or the way the fluctuations in portfolio diverge when approaching the transition, should not depend on
this distribution. This property, called universality, has been one of the major discoveries of statistical mechanics in
1 the last fifty years.
2 Instead of focusing only on the the minimal cost, the statistical mechanics approach makes a detour: it consid-
3 ers, for a given hhistory of returnsx
4 i it , a probability distribution in the space of variables Y, defined by Pγ (Y) =
5 1/Zγ [{xit }] exp − γEβ [Y; {xit }] . The parameter γ is an auxiliary parameter: in physics it is the inverse of the
6 temperature, in the present case it is just one parameter that we introduce in order to have a probability distribution
7 on Y which interpolates between the uniform probability (γ = 0), and a probability which is peaked on the value of
8 Y which minimizes the cost Eβ [Y; {xit }] (case γ = ∞).
9 The normalization constant Zγ [{xit }] is called the partition function at inverse temperature γ; it is defined as
10 Z h i
11 Zγ [{xit }] = dY exp − γEβ [Y; {xit }] (11)
12 V
13
where V is the convex polytope defined by (9).
14
The partition function contains a lot of information on the problem. For instance the minimal cost can be expressed
Fo
15
as limγ→∞ (−1)/(N γ) log Zγ [{xit }]. We shall be interested in computing the large N limit of the minimal cost per
16 variable:
17
18 min E[{xit }] −1
ε[{xit }] = lim = lim lim log Zγ [{xit }] . (12)
rP
19 N →∞ N N →∞ γ→∞ N γ
20
21 In the following, we will compute the average value of this quantity over the choice of the sample xit . Using equation
peer-00568435, version 1 - 23 Feb 2011

22 (12), we can compute this average minimum cost if we are able to compute the average of the logarithm of Z. This is
a difficult problem which is usually circumvented by means of the so called “replica trick”: one computes the average
ee
23
24 of Z n , where n is an integer, and then the average of the logarithm is obtained by
25
26 ∂Z n
log Z = lim , (13)
rR

27 n→0 ∂n

28 thus assuming that Z n can be analytically continued to real values of n. The overline stands for an average over
29 different samples, i.e. over the probability distribution (10). This technique has a long history in the physics of spin
30 glasses [8]: the proof that it leads to the correct solution has been obtained [9] recently.
ev

31 The partition function (11) can be written more explicitly as

32
33 Z +∞ Z +∞ Y T Z +∞ Y N Z +i∞ " N
X
!#
34 Zγ [{xit }] = dv dut dwi dλ exp λ wi − N ×
ie

35 −∞ 0 t=1 −∞ i=1 −i∞ i=1

36 Z T
+∞ Y Z T
+i∞ Y
" T
X N
X
!# " T
X
#
37 × dµt dµ̂t exp µ̂t ut + v + xit wi − µt exp −γ(1 − β)T v − γ ut (14)
,
w

38 0 t=1 −i∞ t=1 t=1 i=1 t=1

39
40 where the constraints are imposed by means of the Lagrange multipliers λ, µ, µ̂. The replica trick is based on the
On

41 idea that the n-th power of the partition function appearing in (13), can be written as the partition function for n
42 independent replicas Y1 ,...,Yn of the system: all the replicas correspond to the hsame history of returns i{xit }, and
Pn
43 their joint probability distribution function is Pγ (Y1 , . . . , Yn ) = 1/Zγn [{xit }] exp − γ a=1 Eβ [Ya ; {xit }] . It is not
44 difficult to write down the expression for Z n and average it over the distribution of samples xit . One introduces the
45
ly

overlap matrix
46
47 N
ab 1 X a b
48 Q = w w , a, b = 1, . . . n , (15)
N i=1 i i
49
50
51 as well as its conjugate Q̂ab (the Lagrange multiplier imposing (15)), where a and b are replica indexes. This matrix
52 characterizes how the portfolios in different replicas differ: they provide some indication of how the measure Pγ (Y)
53
54
55
56
57
58
59
60
E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf
 is spread. After (several) Gaussian integrations, one obtains
Page 6 of 22 Quantitative Finance
n
Z +∞ Y Z +∞ Y Z +i∞ Y (
X X X
a ab ab
n
Zγ [{xit }] ∼ dv dQ dQ̂ exp N Qab Q̂ab − N Q̂ab − γ(1 − β)T va
1 −∞ a=1 −∞ a,b −i∞ a,b a,b a,b a
2
T
)
N nN
3 −T n log γ + T log Ẑγ {v a }, {Qab} − Tr log Q − Tr log Q̂ − log 2 , (16)
4 2 2 2
5
6 where
7
 
Z +∞ Y n n n
8 1 X X
Ẑγ {v a }, {Qab } ≡ dy a exp − (Q−1 )ab (y a − v a )(y b − v b ) + γ y a θ(−y a ) .


(17)
9 −∞ a=1 2 a=1
a,b=1
10
11 We now write T = tN and work at fixed t while N → ∞.
12 The most natural solution is obtained by realizing that all the replicas are identical. Given the linear character of
13 the problem, the symmetric solution should be the correct one. The replica-symmetric solution corresponds to the
14 ansatz
Fo
15 ( (
16 ab q1 if a = b ab q̂1 if a = b
Q = ; Q̂ = , (18)
17 q0 if a 6= b q̂0 if a 6= b
18
rP
19 and v a = v for any a. As we discuss in detail in appendix A, one can show that the optimal cost function, computed
20 from eq. (??), is the minimum of
21 Z +∞
peer-00568435, version 1 - 23 Feb 2011

 
22 1 q0 t 2 p
ε(v, q0 , ∆) = + ∆ t(1 − β)v − + √ dse−s g(v + s 2q0 ) , (19)
ee
23 2∆ 2 2 π −∞
24
25 where ∆ ≡ limγ→∞ γ∆q and the function g(·) is defined as
26
rR

27 0
 x≥0,
g(x) = x 2 (20)
28 −1 ≤ x < 0 ,
29

−2x − 1 x < −1 .
30
ev

31 Note that this function and its derivative are continuous. Moreover, v and q0 in (19) are solutions of the saddle point
32 equations
33
1
Z
2
34
p
1−β+ √ dse−s g ′ (v + s 2q0 ) = 0 , (21)
ie

35 2 π
t
Z
36 2 p
−1 + √ dse−s s g ′ (v + s 2q0 ) = 0 . (22)
37 2πq0
w

38
39 We require that the minimum of (19) occur at a finite value of ∆. In order to understand this point, we recall the
40 meaning of ∆ (see also (18)):
On

41
N N
42 1 X (1)
2 1 X (1) (2)
∆/γ ∼ ∆q = (q1 − q0 ) = wi − w wi ∼ w2 − w2 , (23)
43 N i=1 N i=1 i
44
45
ly

where the superscripts (1) and (2) represent two generic replicas of the system. We then find that ∆ is proportional
46 to the fluctuations in the distribution of the w’s. An infinite value of ∆ would then correspond to a portfolio which is
47 infinitely short on some particular positions and, because of the global budget constraint (1), infinitely long on some
48 other ones.
49 Given (19), the existence of a solution at finite ∆ translates into the following condition:
50 Z +∞
51 q0 t 2 p
t(1 − β)v − + √ dse−s g(v + s 2q0 ) ≥ 0 , (24)
52 2 2 π −∞
53
54 which defines, along with eqs. (21) and (22), our phase diagram.
55
56
57
58
59
60
E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf
 5
0.6 10
Quantitative Finance β = 0.95 Page 7 of 22
β = 0.90
0.5
4 β = 0.80
1 10 √
1/ x
0.4 non-feasible
2
3 3

∆
0.3
N

10
T

4
5
0.2
6
7 feasible 102
0.1
8
9
10 0 101
0 0.2 0.4 0.6 0.8 1 10-8 10-7 10-6» 10-5
„ «∗ –10
-4
10-3 10-2
11 β
N
−
N
12 T T
13
14 FIG. 2: Left panel: The phase diagram of the feasibility problem for expected shortfall. Right panel: The order parameter ∆
Fo
15 diverges with an exponent 1/2 as the transition line is approached. A curve of slope −1/2 is also shown for comparison.
16
17
IV. THE FEASIBLE AND UNFEASIBLE REGIONS
18
rP
19
20 We can now chart the feasibility map of the expected shortfall problem. Following the notation of [6], we will use
21 as control parameters N/T ≡ 1/t and β. The limiting case β → 1 can be worked out analytically and one can show
that the critical value t∗ is given by
peer-00568435, version 1 - 23 Feb 2011

22
ee
23 1 1 h i
= − O (1 − β)3 e−(4π(1−β) )
2 −1

24 . (25)
t ∗ 2
25
26 This limit corresponds to the over-pessimistic case of maximal loss, in which the single worst loss contribute to the
rR

27 risk measure. The optimization problem is the following:

28   X 
29 min max − wi xit . (26)
w t
30 i
ev

31 A simple “geometric” argument by Kondor et al. [6] leads to the critical value 1/t∗ = 0.5 in this extreme case. The
32 idea is the following. According to eq. (26), one has to look for the minimum of a polytope made by a large number
33 of planes, whose normal vectors (the xit ) are drawn from a symmetric distribution. The simplex is convex, but with
34 some probability it can be unbounded from below and then the optimization problem is ill defined. Increasing T
ie

35 means that the probability of this event decreases, because there are more planes and thus it is more likely that for
36 large values of the wi the max over t has a positive slope in the i-th direction. The exact law for this probability
37 can be obtained by induction on N and T [6] and, as we said, it jumps in the thermodynamic limit from 1 to 0 at
w

38 N/T = 0.5. The example of the max-loss risk measure is also helpful because it allows to stress two aspects of the
39 problem: 1) even for finite N and T there is a finite chance that the risk measure is unbounded from below in some
40 samples, and 2) the phase transition occurs in the thermodynamic limit when N/T is strictly smaller than 1, i.e.
On

41 much before the covariance matrix develops zero modes. The very nature of the problem is that the risk measure
42 there is simply not bounded from below. As for the ES risk measure, the threshold value N/T = 0.5 can be thought
43 of as a good approximation of the actual value for many cases of practical interest (i.e. β & 0.9), since the corrections
44 to this limit case are exponentially small (eq. (25)).
45 For finite values of β we have solved numerically eqs. (21), (22) and (24) using the following procedure. We first
ly

46 solve the two equations (21) and (22), which always admit a solution for (v, q0 ). We then plot the l.h.s. of eq. (24) as
47 a function of 1/t for a fixed value of β. This function is positive at small 1/t and becomes negative beyond a threshold
48 1/t∗ . By keeping track of 1/t∗ (numerically obtaining via linear interpolations) for each value of β we build up the
49 phase diagram (Fig. 2, left). This diagram is in agreement with the numerical results obtained in ref. 6. We show in
50 the right panel of Fig. 2 the divergence of the order parameter ∆ versus 1/t − 1/t∗ . The critical exponent is found to
51 be 1/2:
52  −1/2
53 1 1
∆∼ − ∗ , (27)
54 t t (β)
55
56
57
58
59
60
E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf
 1 1
Page 8 of 22 Quantitative Finance

0.8 0.8
1
2

probability
probability 0.6 0.6
3
4 N = 32 N = 32
5 0.4 N = 64 0.4 N = 64
6 N = 128 N = 128
7 0.2 N = 256 0.2 N = 256
8 N = 512 N = 512
9 N = 1024 N = 1024
0 0
10 0.4 0.45 0.5 0.55 0.6 -1.5 -1 -0.5
» „ 0«∗ – 0.5 1 1.5
11 N N
−
N
· N 0.5
12 T T T
13
14 FIG. 3: Left: The probability of finding a finite solution as obtained from linear programming at increasing values of N and
Fo
15 with β = 0.8. Right: Scaling plot of the same data. The critical value is set equal to the analytical one, N/T = 0.4945 and the
critical exponent is 1/2, i.e. the one obtained by Kondor et al. [6] for the limit case β → 1. The data do not collapse perfectly,
16
and better results can be obtained by slightly changing either the critical value or the exponent.
17
18
rP
19 again in agreement with the scaling found in ref. 6. We have performed extensive numerical simulations in order
20 to check the validity of our analytical findings. For a given realization of the time series, we solve the optimization
21 problem (8) by standard linear programming [10]. We impose a large negative cutoff for the w’s, that is wi > −W ,
peer-00568435, version 1 - 23 Feb 2011

22 and we say that a feasible solution exists if it stays finite for W → ∞. We then repeat the procedure for a certain
ee
23 number of samples, and then average our final results (optimal cost, optimal v, and the variance of the w’s in the
24 optimal portfolio) over those of them which produced a finite solution. In Fig. 3 we show how the probability of
25 finding a finite solution depends on the size of the problem. Here, the probability is simply defined in terms of the
26 frequency. We see that the convergence towards the expected 1-0 law is fairly slow, and a finite size scaling analysis
rR

27 is shown in the right panel. Without loss of generality, we can summarize the finite-N numerical results by writing
28 the probability of finding a finite solution as
29   
30 1 1
p(N, T, β) = f − ∗ · N α(β) , (28)
ev

31 t t (β)
32
33 where f (x) → 1 if x ≫ 1 and f (x) → 0 if x ≪ 1, and where α(1) = 1/2. It is interesting to note that these results do
34 not depend on some initial conditions of the algorithm we used to solve the problem: for a given sample, the algorithm
ie

35 finds in a linear time the minimum of the polytope by looking at all its vertexes exhaustively. The statistics is taken
36 by repeating such a deterministic procedure on a large number of samples chosen at random.
37 In Fig. 4 (left panel) we plot, for a given value of β, the optimal cost found numerically for several values of the size
w

38 N compared to the analytical prediction at infinite N . One can show that the cost vanishes as ∆−1 ∼ (1/t − 1/t∗)1/2 .
39 The right panel of the same figure shows the behavior of the value of v which leads to the optimal cost versus N/T ,
40 for the same fixed value of β. Also in this case, the analytical (N → ∞ limit) is plotted for comparison. We note that
On

41 this quantity was suggested [5] to be a good approximation of the VaR of the optimal portfolio: We find here that
42 vopt diverges at the critical threshold and becomes negative at an even smaller value of N/T .
43
44
45 V. CONCLUSIONS
ly

46
47 We have shown that the problem of optimizing a portfolio under the expected shortfall measure of risk by using
48 empirical distributions of returns is not well defined when the ratio N/T of assets to data points is larger than
49 a certain critical value. This value depends on the threshold β of the risk measure in a continuous way and this
50 defines a phase diagram. The lower the value of β, the larger the length of the time series needed for the portfolio
51 optimization. The analytical approach we have discussed in this paper allows us to have a clear understanding of this
52 phase transition. The mathematical reason for the non-feasibility of the optimization problem is that, with a certain
53 probability p(N, T, β), the linear constraints in (9) define a simplex which is not bounded from below, thus leading
54 to a solution which is not finite (∆q → ∞ in our language), in the same way as it happens in the extreme case β → 1
55
56
57
58
59
60
E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf
 1 0.4
N = 32 Quantitative Finance Page 9 of 22
0.2
N = 64
0.8 N = 128 0
1 N = 256 -0.2
2 cost function 0.6 analytic
3 -0.4

vopt
4 -0.6
5 0.4
-0.8 N = 32
6 N = 64
-1
7 0.2 N = 128
8 -1.2 N = 256
9 -1.4 analytic
0
10 0.25 0.3 0.35 0.4 0.45 0.5 0.25 0.3 0.35 0.4 0.45 0.5
11 N N
12 T T
13
14 FIG. 4: Numerical results from linear programming and comparison with analytical predictions at large N . Left: The minimum
Fo
15 cost of the optimization problem vs N/T , at increasing values of N . The thick line is the analytical solution (19). Here β = 0.7,
(N/T )∗ ≃ 0.463. Right: The optimal value of v as found numerically for several values of N is compared to the analytical
16
solution.
17
18
rP
19 discussed in [6]. From a more physical point of view, it is reasonable that the feasibility of the problem depend on
20 the number of data points we take from the time series with respect to the number of financial instruments of our
21 portfolio. The probabilistic character of the time series is reflected in the probability p(N, T, β). Interestingly, this
peer-00568435, version 1 - 23 Feb 2011

22 probability becomes a threshold function at large N if N/T ≡ 1/t is finite, and its general form is given in (28).
ee
23 These results have a practical relevance in portfolio optimization. The order parameter discussed in this paper is
24 tightly related to the relative estimation error [6]. The fact that this order parameter has been found to diverge
25 means that in some regions of the parameter space the estimation error blows up which makes the task of portfolio
26 optimization completely meaningless. The divergence of estimation error is not limited to the case of expected
rR

27 shortfall: as shown in [6], it happens in the case of variance and absolute deviation as well, but the noise sensitivity
28 of expected shortfall turns out to be even greater than that of these more conventional risk measures.
29 There is nothing surprising about the fact that if there are no sufficient data, the estimation error is large and
30 we cannot make a good decision. What is surprising is the fact that there is a sharply defined threshold where the
ev

31 estimation error actually diverges.

32 For a given portfolio size, it is important to know that a minimum amount of data points is required in order to
33 perform an optimization based on empirical distributions. We also note that the divergence of the parameter ∆ at
34 the phase transition, which is directly related to the fluctuations of the optimal portfolio, may play a dramatic role
ie

35 in practical cases. To stress this point, we can define a sort of “susceptibility” with respect to the data,
36
37 ∂hwj i
χtij =
w

, (29)
38 ∂xit
39
and one can show that this quantity diverges at the critical point, since χij ∼ ∆. A small change (or uncertainty) in
40
xit becomes increasingly relevant as the transition is approached, and the portfolio optimization could then be very
On

41
unstable even in the feasible region of the phase diagram. We stress that the susceptibility we have introduced might
42
be considered as a measure of the effect of the noise on portfolio selection and is very reminiscent of the measure
43
proposed in [11].
44
In order to present a clean, analytic picture, we have made several simplifying assumptions in this work. We have
45
ly

omitted the constraint on the returns, liquidity constraints, correlations between the assets, non-stationary effects,
46
etc. Some of these can be systematically taken into account and we plan to return to these finer details in a subsequent
47 work.
48 Acknowledgments. We thank O. C. Martin, and M. Potters for useful discussions, and particularly J. P. Bouchaud
49 for a critical reading of the manuscript. S. C. is supported by EC through the network MTR 2002-00319, STIPCO,
50 I.K. by the National Office of Research and Technology under grant No. KCKHA005.
51
52
53
54
55
56
57
58
59
60
E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf
 APPENDIX A: THE REPLICA SYMMETRIC SOLUTION
Page 10 of 22 Quantitative Finance
We show in this appendix how the minimum cost function corresponding to the replica-symmetric ansatz is obtained.
The ‘TrLogQ’ term in (16) is computed by realizing that the eigenvalues of such a symmetric matrix are (q1 + (n −
1
1)q0 ) (with multiplicity 1) and (q1 − q0 ) with multiplicity n − 1. Then,
2
3
 
q1
4 Tr log Q = log detQ = log(q1 + (n − 1)q0 ) + (n − 1) log(q1 − q0 ) = n log ∆q + + O(n2 ) , (A1)
∆q
5
6 where ∆q ≡ q1 − q0 . The effective partition function in (17) depends on Q−1 , whose elements are:
7 (
(∆q − q0 )/(∆q)2 + O(n) if a = b
8 (Q−1 )ab = (A2)
9 −q0 /(∆q)2 + O(n) if a =
6 b
10 ds 2

11 By introducing a Gaussian measure dPq0 (s) ≡ √2πq 0

e−s /2q0 , one can show that
12 (Z )
1 1
Z
Y 1 a 2 a a s
P a
13
P P
log Ẑ(v, q1 , q0 ) = log dxa e− 2∆q a (x ) +γ a (x +v)θ(−x −v) dPq0 (s)e ∆q a x
14 n n a
Fo
15 q0
Z
16 = + dPq0 (s) log Bγ (s, v, ∆q) + O(n) (A3)
2∆q
17
18 where we have defined
rP
19 (x − s)2
Z  
20 Bγ (s, v, ∆q) ≡ dx exp − + γ(x + v)θ(−x − v) . (A4)
2∆q
21
peer-00568435, version 1 - 23 Feb 2011

22 The exponential in (16) now reads expN n[S(q0 , ∆q, q̂0 , ∆q̂) + O(n)], where
ee
23 Z
24 S(q0 , ∆q, q̂0 , ∆q̂) = q0 ∆q̂ + q̂0 ∆q + ∆q∆q̂ − ∆q̂ − γt(1 − β)v − t log γ + t
dPq0 (s) log Bγ (s, v, ∆q)
25   (A5)
t 1 q̂0 log 2
26 − log ∆q − log ∆q̂ + − .
rR

27 2 2 ∆q̂ 2
28 The saddle point equations for q̂0 and ∆q̂ allow then to simplify this expression. The free energy (−γ)fγ =
29 limn→0 ∂Zγn /∂n is given by
30
1 1−t q0 − 1
Z
ev

31 −γfγ (v, q0 , ∆q) = − t log γ + log ∆q + − γt(1 − β)v + t dPq0 (s) log Bγ (s, v, ∆q) , (A6)
32 2 2 2∆q
33 where the actual values of v, q0 and ∆q are fixed by the saddle point equations
34
ie

35 ∂fγ ∂fγ ∂fγ

= = =0. (A7)
36 ∂v ∂q0 ∂∆q
37 A close inspection of these saddle point equations allows one to perform the low temperature γ → ∞ limit by assuming
w

38 that ∆q = ∆/γ while v and q0 do not depend on the temperature. In this limit one can show that
39 
40
1 s + v + ∆/2
 s < −v − ∆
On

41 lim log Bγ (s, v, ∆/γ) = −(v + s)2 /2∆ −v − ∆ ≤ s < −v (A8)

γ→∞ γ
42 
0 s ≥ −v
43
44 If we plug this expression into eq. (A6) and perform the large-γ limit we get the minimum cost:
45
ly

Z −∆   Z 0
q0 − 1 dx − (x−v) 2
∆ t dx − (x−v)2
46 E = lim fγ = − + t(1 − β)v − t √ e 2q0 x+ + √ e 2q0 x2 . (A9)
47 γ→∞ 2∆ −∞ 2πq0 2 2∆ −∆ 2πq0
48 We rescale x → x∆, v → v∆, and q0 → q0 ∆2 , and after some algebra we obtain eq. (19).
49
50
51
52
53 [1] see P. Artzner, F. Delbaen, J. M. Eber, and D. Heath, Mathematical Finance 9, 203–228 (1999), for an axiomatic definition
54 of coherence.
55
56
57
58
59
60
E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf
 [2] C. Acerbi, and D. Tasche, Journal of Banking and Finance 26, 1487–1503 (2002).
[3] R. Frey, and A. J. McNeil, (2002). Quantitative Finance Page 11 of 22
[4] G. C. Pflug, in “Probabilistic Constrained Optimization: Methodology and Applications”, S. Uryasev (ed.), Kluwer Aca-
demic Publisher (2000).
1 [5] R. Rockafellar, and S. Uryasev, The Journal of Risk 2, 21–41 (2000).
2 [6] I. Kondor, Noise sensitivity of portfolio selection under various risk measures, Lectures given at the Summer School on
3 “Risk Measurement and Control”, Rome, June 9-17, 2005; I. Kondor, S. Pafka, and G. Nagy, Noise sensitivity of portfolio
4 selection under various risk measures, submitted to J. of Banking and Finance.
[7] see e.g. F. Guerra, F.L. Toninelli, J. Stat. Phys. 115, 531 (2004) and references therein.
5
[8] M. Mézard, G. Parisi, and M. .A. .Virasoro, “Spin Glass theroy and Beyond”, World Scientific Lecture Notes in Physics
6 Vol. 9, Singapore (1987).
7 [9] M. Talagrand, Spin Glasses: a Challenge for Mathematicians, Spinger-Verlag (2002).
8 [10] W. H. Press, S. H. Teukolsky, W. T. Wetterling, and B. P. Flannery, “Numerical Recipes in C”, Cambridge University
9 Press (Cambridge, UK, 1992).
10 [11] S. Pafka, and I. Kondor, Europ. Phys. J. B27, 277 (2002).
11 [12] see [2] for the subtleties related to a discrete distribution
12
13
14
Fo
15
16
17
18
rP
19
20
21
peer-00568435, version 1 - 23 Feb 2011

22
ee
23
24
25
26
rR

27
28
29
30
ev

31
32
33
34
ie

35
36
37
w

38
39
40
On

41
42
43
44
45
ly

46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf
Page 12 of 22 Quantitative Finance

1
2
3
4
5
6
7
8
9
10
11
12 On the Feasibility of Portfolio Optimization under Expected Shortfall
13
14
Fo
15
16 Journal: Quantitative Finance
17 Manuscript ID: RQUF-2006-0108.R1
18
rP
19 Manuscript Category: Research Paper
20
21 Date Submitted by the
28-Jan-2007
Author:
peer-00568435, version 1 - 23 Feb 2011

22
ee
23 Complete List of Authors: ciliberti, stefano; univ paris sud, LPTMS
24 Kondor, Imre
25 Mezard, Marc
26
r

27 Portfolio Optimization, Risk Measures, Statistical Physics, Financial

Keywords:
Applications
28
Re

29 Portfolio Optimization, Risk Measures, Statistical Physics, Financial

30 JEL Code:
Applications
31
32
vi

33
34 Note: The following files were submitted by the author for peer review, but cannot be converted
to PDF. You must view these files (e.g. movies) online.
35
ew

36 resubmission.tgz
37
38
39
40
On

41
42
43
44
45
ly

46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf
Page 1 of 10 Quantitative Finance Page 13 of 22

1
2 On the Feasibility of Portfolio Optimization under Expected Shortfall
3
4 Stefano Ciliberti,1 Imre Kondor,2 and Marc Mézard1
1
5 CNRS; Univ. Paris Sud, UMR8626, LPTMS, ORSAY CEDEX, F-91405
2
6 Collegium Budapest, 1014 Budapest, Szenthromsg u. 2
7 We address the problem of portfolio optimization under the simplest coherent risk measure, i.e.
8 the expected shortfall. As it is well known, one can map this problem into a linear programming
9 setting. For some values of the external parameters, when the available time series is too short,
10 the portfolio optimization is ill posed because it leads to unbounded positions, infinitely short on
11 some assets and infinitely long on some others. As first observed by Kondor and coworkers, this
12 phenomenon is actually a phase transition. We investigate the nature of this transition by means
13 of a replica approach.
14
15
I. INTRODUCTION
16
Fo

17
18 Among the several risk measures existing in the context of portfolio optimization, expected shortfall (ES) has
19 certainly gained increasing popularity in recent years. In several practical applications, ES is starting to replace the
20 classical Value-at-Risk (VaR). There are a number of reasons for this. For a given threshold probability β, the VaR
rP

21 is defined so that with probability β the loss will be smaller than VaR. This definition only gives the minimum loss
22 one can reasonably expect but does not tell anything about the typical value of that loss that can be measured by the
23 conditional value-at-risk (CVaR, which is the same as ES for continuous distributions that we consider here [12]). We
will be more precise on these definitions below. The point we want to stress here is that the VaR measure, lacking the
peer-00568435, version 1 - 23 Feb 2011

24
mandatory properties of subadditivity and convexity,is not coherent [1]. This means that summing VaR’s of individual
ee

25
portfolios will not necessarily produce an upper bound for the VaR of the combined portfolio, thus contradicting the
26
holy principle of diversification in finance. A nice practical example of the inconsistency of VaR in credit portfolio
27
management is reported in Ref. 3. On the other hand, it has been shown [2] that ES is a coherent measure with
28
interesting properties [4]. Moreover, the optimization of ES can be reduced to linear programming [5] (which allows
rR

29
for a fast implementation) and leads to a good estimate for the VaR as a byproduct of the minimization process. To
30
summarize, the intuitive and simple character, together with the mathematical properties (coherence) and the fast
31 algorithmic implementation (linear programming), are the main reasons behind the growing importance of ES as a
32 risk measure.
33
ev

In this paper, we will focus on the feasibility of the portfolio optimization problem under the ES measure of risk.
34 The control parameters of this problem are (i) the imposed threshold in probability, β, and (ii) the ratio N/T between
35 the number N of financial assets making up the portfolio and the time series length T used to sample the probability
36 distribution of returns. (It is curious that, albeit trivial, the scaling in N/T had not been explicitly pointed out before
iew

37 [11]. It has been discovered by Kondor at al. Ref. 6 that, for certain values of these parameters, the optimization
38 problem does not have a finite solution because, even if convex, it is not bounded from below. Extended numerical
39 simulations allowed these authors to determine the feasibility map of the problem. Here, in order to better understand
40 the root of the problem and to study the transition from a feasible regime to an unfeasible one (corresponding to an
41 ill-posed minimization problem) we address the same problem from an analytical point of view.
42 The paper is organized as follows. In Section II we briefly recall the basic definitions of β-VaR and β-CVaR and we
43
On

show how the portfolio optimization problem can be reduced to linear programming. We introduce a “cost function”
44 to be minimized under linear constraints and we discuss the rationale for a statistical mechanics approach. In Section
45 III we solve the problem of optimizing large portfolios under ES using the replica approach. Our results and the
46 comparison with numerics are reported in Section IV, and our conclusions are summarized in Section V.
47
ly

48
49 II. THE OPTIMIZATION PROBLEM
50
51 We consider a portfolio of N financial instruments w = {w1 , . . . wN }, where wi is the position of asset i. The global
52 budget constraint fixes the sum of these numbers: we impose for example
53
N
X
54
55 wi = N . (1)
56 i=1

57 We do not stipulate any constraint on short selling, so that wi can be any negative or positive number. This is, of
58 course, unrealistic for liquidity reasons, but considering this case allows us to show up the essence of the phenomenon.
59
60

E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf

Page 14 of 22 Quantitative Finance Page 2 of 10
2
1
2 1
3
4 β
5
6
7

P< (w; α)
8
9
10
11
12
13
14
15
16 0
Fo

17 β-VaR(w) α
18
19
20
rP

21
FIG. 1: Schematic representation of the VaR measure of risk. P< (w) is the probability of a loss associated to the portfolio w
22 being smaller than α. The conditional VaR β-CVaR (or ES) is the average loss when this is constrained to be greater than the
23 β-VaR.
peer-00568435, version 1 - 23 Feb 2011

24
ee

25
26 If we imposed a constraint that would render the domain of the wi bounded (such as a ban on short selling, e.g.), this
27 would evidently prevent the weights from diverging, but a vestige of the transition would still remain in the form of
28 large, though finite, fluctuations of the weights, and in a large number of them sticking to the ”walls” of the domain.
rR

29 We denote the returns on the assets by x = {x1 , x2 , . . . xN } , and we assume that there exists an underlying proba-
P
30 bility distribution function p(x) of the returns. The loss of portfolio w given the returns x is `(w|x) = − N i=1 wi xi ,
31 and the probability of that loss being smaller than a given threshold α is
32 Z
33

ev

P< (w, α) = dx p(x)θ α − `(w|x) , (2)

34
35
where θ(·) is the Heaviside step function, equal to 1 if its argument is positive and 0 otherwise. The β-VaR of this
36
portfolio is formally defined by
iew

37
38 β-VaR(w) = min{α : P< (w, α) ≥ β} , (3)
39
40 (see Fig. 1), while the CVaR (or ES, in this case) associated with the same portfolio is the average loss on the tail of
41 the distribution,
42 Z


43 dx p(x)`(w|x)θ `(w|x) − β-VaR(w) Z
On

1

44 β-CVaR(w) = Z = dx p(x)`(w|x)θ `(w|x) − β-VaR(w) . (4)

1−β
45 dx p(x)θ `(w|x) − β-VaR(w)
46
47 The threshold β then represents a confidence level. In practice, the typical values of β which one considers are
ly

48 β = 0.90, 0.95, and 0.99, but we will address the problem for any β ∈ [0, 1]. What is usually called “exceeding
49 probability” in previous literature would correspond here to (1 − β).
50 As mentioned in the introduction, the ES measure can be obtained from a variational principle [5]. The minimization
51 of a properly chosen objective function leads directly to (4):
52
53 β-CVaR(w) = min Fβ (w, v) , (5)
v
54 Z
−1
 +
55 Fβ (w, v) ≡ v + (1 − β) dx p(x) `(w|x) − v . (6)
56
57 Here, [a]+ ≡ (a + |a|)/2. The external parameter v over which one has to minimize is claimed to be relevant in
58 itself [5], since its optimal value may represent a good estimate for the actual value-at-risk of the portfolio. We will
59
60

E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf

Page 3 of 10 Quantitative Finance Page 15 of 22
3
1
2 come back to this point as we discuss our results. We stress here that minimizing (6) over w and v is equivalent to
3 optimizing (4) over the portfolio vectors w.
4 Of course, in practical cases the probability distribution of the loss is not known and must be inferred from the
5 past data. In other words, we need an “in-sample” estimate of the integral in (6), which would turn a well posed
6 (but useless) optimization problem into a practical approach. We thus approximate the integral by sampling the
7 probability distributions of returns. If we have a time series x(1) , . . . x(T ) , our objective function becomes simply
8 T T
" N
#+
1 X  + 1 X X
9 F̂β (w, v) = v + (τ )
`(w|x ) − v = v + −v − wi xiτ , (7)
10 (1 − β)T τ =1 (1 − β)T τ =1 i=1
11
12 where we denote by xiτ the return of asset i at time τ .
13 Minimizing this risk measure is the same as the following linear programming problem:
14
• given one data sample, i.e. a matrix xiτ , i = 1, . . . N , τ = 1, . . . T ,
15
16 • minimize the cost function
Fo

17
18 T
X
   
19 Eβ Y; {xiτ } = Eβ v, {wi }, {uτ }; {xiτ } = (1 − β)T v + uτ , (8)
20 t=τ
rP

21
22 • over the (N + T + 1) variables Y ≡ {w1 , . . . wN u1 , . . . uT v,
23 • under the (2T + 1) constraints
peer-00568435, version 1 - 23 Feb 2011

24
ee

25 N
X N
X
26 uτ ≥ 0 , uτ + v + xiτ wi ≥ 0 ∀τ , and wi = N . (9)
27 i=1 i=1
28
rR

29 Since we allow short positions, not all the wi are positive, which makes this problem different from standard linear
30 programming. To keep the problem tractable, we impose the condition that wi ≥ −W , where W is a very large cutoff,
31 and the optimization problem will be said to be ill-defined if its solution does not converge to a finite limit when
32 W → ∞. It is now clear why constraining all the wi to be non-negative would eliminate the feasibility problem: a finite
33 solution will always exist because the weights are by definition bounded, the worst case being an optimal portfolio
ev

34 with only one non-zero weight taking care of the total budget. The control parameters that govern the problem are
35 the threshold β and the ratio N/T of assets to data points. The resulting “phase diagram” is then a line in the β-N/T
36 plane separating a region in which, with high probability, the minimization problem is not bounded and thus does not
iew

37 admit a finite solution, and another region in which a finite solution exists with high probability. These statements
38 are non-deterministic because of the intrinsic probabilistic nature of the returns. We will address this minimization
39 problem in the non-trivial limit where T → ∞, N → ∞, while N/T stays finite. In this “thermodynamic” limit, we
40 shall assume that extensive quantities (like the average loss of the optimal portfolio, i.e. the minimum cost function) do
41 not fluctuate, namely that their probability distribution is concentrated around the mean value. This “self-averaging”
42 property has been proved for a wide range of similar statistical mechanics models [7]. Then, we will be interested
43 in the average value of the minimum of the cost function 8) over the distribution of returns. Given the similarity of
On

44 portfolio optimization with the statistical physics of disordered systems, this problem can be addressed analytically
45 by means of a replica approach [8].
46
47
III. THE REPLICA APPROACH
ly

48
49
50 For a given history of returns xit , one can compute the minimum of the cost function, minY Eβ [Y; {xit }]. In this
51 section we show how to compute analytically the expectation value of this quantity over the histories of return. For
52 simplicity we shall keep to the case in which the xit are independent identically distributed (iid) normal variables, so
53 that an history of returns xit is drawn from the distribution
54 Y 2

55 p({xit }) ∼ e−N xit /2 . (10)

it
56
57 This assumption of iid normal distribution of returns is very restrictive, but we would like to emphasize that the
58 method that we use can be generalized easily to iid variables with other distributions, and also in some cases to
59
60

E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf

Page 16 of 22 Quantitative Finance Page 4 of 10
4
1
2 correlated variables. Certainly the precise location of the critical value of N/T separating an unfeasible phase from a
3 feasible one depends on the distribution of returns. But we expect that the broad features like the existence of this
4 critical value, or the way the fluctuations in portfolio diverge when approaching the transition, should not depend on
5 this distribution. This property, called universality, has been one of the major discoveries of statistical mechanics in
6 the last fifty years.
7 Instead of focusing only on the the minimal cost, the statistical mechanics approach makes a detour: it consid-
8 i it , a probability distribution in the space of variables Y, defined by Pγ (Y) =
ers, for a given hhistory of returnsx
9 1/Zγ [{xit }] exp − γEβ [Y; {xit }] . The parameter γ is an auxiliary parameter: in physics it is the inverse of the
10 temperature, in the present case it is just one parameter that we introduce in order to have a probability distribution
11 on Y which interpolates between the uniform probability (γ = 0), and a probability which is peaked on the value of
12 Y which minimizes the cost Eβ [Y; {xit }] (case γ = ∞).
13 The normalization constant Zγ [{xit }] is called the partition function at inverse temperature γ; it is defined as
14 Z h i
15 Zγ [{xit }] = dY exp − γEβ [Y; {xit }] (11)
16
Fo
V
17
18 where V is the convex polytope defined by (9).
19 The partition function contains a lot of information on the problem. For instance the minimal cost can be expressed
20 as limγ→∞ (−1)/(N γ) log Zγ [{xit }]. We shall be interested in computing the large N limit of the minimal cost per
rP

21 variable:
22 min E[{xit }] −1
23 ε[{xit }] = lim = lim lim log Zγ [{xit }] . (12)
N →∞ N N →∞ γ→∞ Nγ
peer-00568435, version 1 - 23 Feb 2011

24
ee

25 In the following, we will compute the average value of this quantity over the choice of the sample xit . Using equation
26 (12), we can compute this average minimum cost if we are able to compute the average of the logarithm of Z. This is
27 a difficult problem which is usually circumvented by means of the so called “replica trick”: one computes the average
28 of Z n , where n is an integer, and then the average of the logarithm is obtained by
rR

29
30 ∂Z n
log Z = lim , (13)
31 n→0 ∂n
32
33 thus assuming that Z n can be analytically continued to real values of n. The overline stands for an average over
ev

34 different samples, i.e. over the probability distribution (10). This technique has a long history in the physics of spin
35 glasses [8]: the proof that it leads to the correct solution has been obtained [9] recently.
36 The partition function (11) can be written more explicitly as
" !#
iew

37 Z +∞ Z +∞ Y T Z +∞ Y N Z +i∞ N
X
38 Zγ [{xit }] = dv dut dwi dλ exp λ wi − N ×
39 −∞ 0 t=1 −∞ i=1 −i∞ i=1
40 Z T Z T
" T N
!# " T
#
+∞ Y +i∞ Y X X X
41 × dµt dµ̂t exp µ̂t ut + v + xit wi − µt exp −γ(1 − β)T v − γ ut (14)
,
42 0 t=1 −i∞ t=1 t=1 i=1 t=1
43
On

44 where the constraints are imposed by means of the Lagrange multipliers λ, µ, µ̂. The replica trick is based on the
45 idea that the n-th power of the partition function appearing in (13), can be written as the partition function for n
46 independent replicas Y1 ,...,Yn of the system: all the replicas correspond to the hsame history of returns i{xit }, and
Pn
47 their joint probability distribution function is Pγ (Y1 , . . . , Yn ) = 1/Zγn [{xit }] exp − γ a=1 Eβ [Ya ; {xit }] . It is not
ly

48
difficult to write down the expression for Z n and average it over the distribution of samples xit . One introduces the
49 overlap matrix
50
51 N
1 X a b
52 Qab = w w , a, b = 1, . . . n , (15)
53 N i=1 i i
54
55 as well as its conjugate Q̂ab (the Lagrange multiplier imposing (15)), where a and b are replica indexes. This matrix
56 characterizes how the portfolios in different replicas differ: they provide some indication of how the measure Pγ (Y)
57
58
59
60

E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf

Page 5 of 10 Quantitative Finance Page 17 of 22
5
1
2 is spread. After (several) Gaussian integrations, one obtains
3 Z Z Z (
+∞ Y
n +∞ Y +i∞ Y X X X
4 Zγn [{xit }] ∼ dv a dQab dQ̂ab exp N Qab Q̂ab − N Q̂ab − γ(1 − β)T va
5 −∞ a=1 −∞ a,b −i∞ a,b a,b a,b a
6 )
7
T N nN
−T n log γ + T log Ẑγ {v a }, {Qab} − Tr log Q − Tr log Q̂ − log 2 , (16)
8 2 2 2
9
10 where
11  
Z n
+∞ Y n
X Xn
12
1
Ẑγ {v a }, {Qab } ≡ dy a exp − (Q−1 )ab (y a − v a )(y b − v b ) + γ y a θ(−y a ) . (17)
13 −∞ 2
a=1 a,b=1 a=1
14
15 We now write T = tN and work at fixed t while N → ∞.
16
Fo
The most natural solution is obtained by realizing that all the replicas are identical. Given the linear character of
17 the problem, the symmetric solution should be the correct one. The replica-symmetric solution corresponds to the
18 ansatz
19 ( (
20 q if a = b q̂1 if a = b
rP

1
21 Qab = ; Q̂ab = , (18)
q0 if a 6= b q̂0 if a 6= b
22
23 and v a = v for any a. As we discuss in detail in appendix A, one can show that the optimal cost function, computed
peer-00568435, version 1 - 23 Feb 2011

24 from eq. (??), is the minimum of

ee

25
 Z +∞ 
26 1 q0 t −s2
p
27 ε(v, q0 , ∆) = + ∆ t(1 − β)v − + √ dse g(v + s 2q0 ) , (19)
2∆ 2 2 π −∞
28
rR

29 where ∆ ≡ limγ→∞ γ∆q and the function g(·) is defined as

30 
31 
0 x≥0,
32 g(x) = x2 −1 ≤ x < 0 , (20)
33 
−2x − 1 x < −1 .
ev

34
35 Note that this function and its derivative are continuous. Moreover, v and q0 in (19) are solutions of the saddle point
36 equations
iew

37 Z
38 1 2 p
1−β+ √ dse−s g 0 (v + s 2q0 ) = 0 , (21)
39 2 π
40 Z p
t 2
41 −1 + √ dse−s s g 0 (v + s 2q0 ) = 0 . (22)
2πq0
42
43
On

We require that the minimum of (19) occur at a finite value of ∆. In order to understand this point, we recall the
44 meaning of ∆ (see also (18)):
45
N N
46 1 X (1)
2 1 X (1) (2)
47 ∆/γ ∼ ∆q = (q1 − q0 ) = wi − w wi ∼ w2 − w2 , (23)
N i=1 N i=1 i
ly

48
49
where the superscripts (1) and (2) represent two generic replicas of the system. We then find that ∆ is proportional
50
to the fluctuations in the distribution of the w’s. An infinite value of ∆ would then correspond to a portfolio which is
51 infinitely short on some particular positions and, because of the global budget constraint (1), infinitely long on some
52 other ones.
53 Given (19), the existence of a solution at finite ∆ translates into the following condition:
54
55 Z +∞ p
q0 t 2
56 t(1 − β)v − + √ dse−s g(v + s 2q0 ) ≥ 0 , (24)
2 2 π −∞
57
58 which defines, along with eqs. (21) and (22), our phase diagram.
59
60

E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf

Page 18 of 22 Quantitative Finance Page 6 of 10
6
1
2 0.6 105
β = 0.95
3
β = 0.90
4 0.5
β = 0.80
5 104 √
1/ x
6 0.4 non-feasible
7
103

∆
8 0.3
N
T

9
10 0.2
11 feasible 102
12 0.1
13
1
14 0 10
0 0.2 0.4 0.6 0.8 1 -8 -7 -6 -5 -4 -3 -2
15 10 10 10 » 10
„ «∗ –10 10 10
N N
β −
16 T T
Fo

17
18 FIG. 2: Left panel: The phase diagram of the feasibility problem for expected shortfall. Right panel: The order parameter ∆
19 diverges with an exponent 1/2 as the transition line is approached. A curve of slope −1/2 is also shown for comparison.
20
rP

21
22 IV. THE FEASIBLE AND UNFEASIBLE REGIONS
23
peer-00568435, version 1 - 23 Feb 2011

24 We can now chart the feasibility map of the expected shortfall problem. Following the notation of [6], we will use
ee

25 as control parameters N/T ≡ 1/t and β. The limiting case β → 1 can be worked out analytically and one can show
26 that the critical value t∗ is given by
27 h i
1 1 3 −(4π(1−β)2 )
−1
28 = − O (1 − β) e . (25)
rR

29 t∗ 2
30 This limit corresponds to the over-pessimistic case of maximal loss, in which the single worst loss contribute to the
31 risk measure. The optimization problem is the following:
32   X 
33 min max − wi xit . (26)
ev

w t
34 i
35
36 A simple “geometric” argument by Kondor et al. [6] leads to the critical value 1/t∗ = 0.5 in this extreme case. The
idea is the following. According to eq. (26), one has to look for the minimum of a polytope made by a large number
iew

37
38 of planes, whose normal vectors (the xit ) are drawn from a symmetric distribution. The simplex is convex, but with
39 some probability it can be unbounded from below and then the optimization problem is ill defined. Increasing T
means that the probability of this event decreases, because there are more planes and thus it is more likely that for
40
large values of the wi the max over t has a positive slope in the i-th direction. The exact law for this probability
41
can be obtained by induction on N and T [6] and, as we said, it jumps in the thermodynamic limit from 1 to 0 at
42
N/T = 0.5. The example of the max-loss risk measure is also helpful because it allows to stress two aspects of the
43
On

problem: 1) even for finite N and T there is a finite chance that the risk measure is unbounded from below in some
44
samples, and 2) the phase transition occurs in the thermodynamic limit when N/T is strictly smaller than 1, i.e.
45
much before the covariance matrix develops zero modes. The very nature of the problem is that the risk measure
46
there is simply not bounded from below. As for the ES risk measure, the threshold value N/T = 0.5 can be thought
47 of as a good approximation of the actual value for many cases of practical interest (i.e. β & 0.9), since the corrections
ly

48 to this limit case are exponentially small (eq. (25)).

49 For finite values of β we have solved numerically eqs. (21), (22) and (24) using the following procedure. We first
50 solve the two equations (21) and (22), which always admit a solution for (v, q0 ). We then plot the l.h.s. of eq. (24) as
51 a function of 1/t for a fixed value of β. This function is positive at small 1/t and becomes negative beyond a threshold
52 1/t∗ . By keeping track of 1/t∗ (numerically obtaining via linear interpolations) for each value of β we build up the
53 phase diagram (Fig. 2, left). This diagram is in agreement with the numerical results obtained in ref. 6. We show in
54 the right panel of Fig. 2 the divergence of the order parameter ∆ versus 1/t − 1/t∗ . The critical exponent is found to
55 be 1/2:
56
 −1/2
57 1 1
58 ∆∼ − ∗ , (27)
t t (β)
59
60

E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf

Page 7 of 10 Quantitative Finance Page 19 of 22
7
1
2 1 1
3
4 0.8 0.8
5
6

probability
probability

7 0.6 0.6
8 N = 32 N = 32
9 0.4 N = 64 0.4 N = 64
10 N = 128 N = 128
11 N = 256 N = 256
0.2 0.2
12 N = 512 N = 512
13 N = 1024 N = 1024
14 0 0
0.4 0.45 0.5 0.55 0.6 -1.5 -1 -0.5
» „ 0«∗ – 0.5 1 1.5
15 N N N
− · N 0.5
16 T T T
Fo

17
18 FIG. 3: Left: The probability of finding a finite solution as obtained from linear programming at increasing values of N and
19 with β = 0.8. Right: Scaling plot of the same data. The critical value is set equal to the analytical one, N/T = 0.4945 and the
20 critical exponent is 1/2, i.e. the one obtained by Kondor et al. [6] for the limit case β → 1. The data do not collapse perfectly,
rP

21 and better results can be obtained by slightly changing either the critical value or the exponent.
22
23
again in agreement with the scaling found in ref. 6. We have performed extensive numerical simulations in order
peer-00568435, version 1 - 23 Feb 2011

24
to check the validity of our analytical findings. For a given realization of the time series, we solve the optimization
ee

25
26 problem (8) by standard linear programming [10]. We impose a large negative cutoff for the w’s, that is wi > −W ,
27 and we say that a feasible solution exists if it stays finite for W → ∞. We then repeat the procedure for a certain
28 number of samples, and then average our final results (optimal cost, optimal v, and the variance of the w’s in the
rR

29 optimal portfolio) over those of them which produced a finite solution. In Fig. 3 we show how the probability of
finding a finite solution depends on the size of the problem. Here, the probability is simply defined in terms of the
30
frequency. We see that the convergence towards the expected 1-0 law is fairly slow, and a finite size scaling analysis
31
is shown in the right panel. Without loss of generality, we can summarize the finite-N numerical results by writing
32
the probability of finding a finite solution as
33
ev

34   
1 1 α(β)
35 p(N, T, β) = f − ∗ ·N , (28)
t t (β)
36
iew

37 where f (x) → 1 if x  1 and f (x) → 0 if x  1, and where α(1) = 1/2. It is interesting to note that these results do
38 not depend on some initial conditions of the algorithm we used to solve the problem: for a given sample, the algorithm
39 finds in a linear time the minimum of the polytope by looking at all its vertexes exhaustively. The statistics is taken
40 by repeating such a deterministic procedure on a large number of samples chosen at random.
41 In Fig. 4 (left panel) we plot, for a given value of β, the optimal cost found numerically for several values of the size
42 N compared to the analytical prediction at infinite N . One can show that the cost vanishes as ∆−1 ∼ (1/t − 1/t∗)1/2 .
43 The right panel of the same figure shows the behavior of the value of v which leads to the optimal cost versus N/T ,
On

44 for the same fixed value of β. Also in this case, the analytical (N → ∞ limit) is plotted for comparison. We note that
45 this quantity was suggested [5] to be a good approximation of the VaR of the optimal portfolio: We find here that
46 vopt diverges at the critical threshold and becomes negative at an even smaller value of N/T .
47
ly

48
49 V. CONCLUSIONS
50
51 We have shown that the problem of optimizing a portfolio under the expected shortfall measure of risk by using
52 empirical distributions of returns is not well defined when the ratio N/T of assets to data points is larger than
53 a certain critical value. This value depends on the threshold β of the risk measure in a continuous way and this
54 defines a phase diagram. The lower the value of β, the larger the length of the time series needed for the portfolio
55 optimization. The analytical approach we have discussed in this paper allows us to have a clear understanding of this
56 phase transition. The mathematical reason for the non-feasibility of the optimization problem is that, with a certain
57 probability p(N, T, β), the linear constraints in (9) define a simplex which is not bounded from below, thus leading
58 to a solution which is not finite (∆q → ∞ in our language), in the same way as it happens in the extreme case β → 1
59
60

E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf

Page 20 of 22 Quantitative Finance Page 8 of 10
8
1
2 1 0.4
N = 32
3 0.2
N = 64
4 0.8 N = 128 0
5
N = 256 -0.2
6
cost function

0.6 analytic
7 -0.4

vopt
8 -0.6
9 0.4
-0.8 N = 32
10 N = 64
11 -1
0.2 N = 128
12 -1.2 N = 256
13 -1.4 analytic
14 0
0.25 0.3 0.35 0.4 0.45 0.5 0.25 0.3 0.35 0.4 0.45 0.5
15 N N
16 T T
Fo

17
18 FIG. 4: Numerical results from linear programming and comparison with analytical predictions at large N . Left: The minimum
19 cost of the optimization problem vs N/T , at increasing values of N . The thick line is the analytical solution (19). Here β = 0.7,
20 (N/T )∗ ' 0.463. Right: The optimal value of v as found numerically for several values of N is compared to the analytical
rP

21 solution.
22
23
discussed in [6]. From a more physical point of view, it is reasonable that the feasibility of the problem depend on
peer-00568435, version 1 - 23 Feb 2011

24
the number of data points we take from the time series with respect to the number of financial instruments of our
ee

25
portfolio. The probabilistic character of the time series is reflected in the probability p(N, T, β). Interestingly, this
26
probability becomes a threshold function at large N if N/T ≡ 1/t is finite, and its general form is given in (28).
27
These results have a practical relevance in portfolio optimization. The order parameter discussed in this paper is
28 tightly related to the relative estimation error [6]. The fact that this order parameter has been found to diverge
rR

29 means that in some regions of the parameter space the estimation error blows up which makes the task of portfolio
30 optimization completely meaningless. The divergence of estimation error is not limited to the case of expected
31 shortfall: as shown in [6], it happens in the case of variance and absolute deviation as well, but the noise sensitivity
32 of expected shortfall turns out to be even greater than that of these more conventional risk measures.
33
ev

There is nothing surprising about the fact that if there are no sufficient data, the estimation error is large and
34 we cannot make a good decision. What is surprising is the fact that there is a sharply defined threshold where the
35 estimation error actually diverges.
36 For a given portfolio size, it is important to know that a minimum amount of data points is required in order to
iew

37 perform an optimization based on empirical distributions. We also note that the divergence of the parameter ∆ at
38 the phase transition, which is directly related to the fluctuations of the optimal portfolio, may play a dramatic role
39 in practical cases. To stress this point, we can define a sort of “susceptibility” with respect to the data,
40
41 ∂hwj i
χtij = , (29)
42 ∂xit
43
On

44 and one can show that this quantity diverges at the critical point, since χij ∼ ∆. A small change (or uncertainty) in
45 xit becomes increasingly relevant as the transition is approached, and the portfolio optimization could then be very
46 unstable even in the feasible region of the phase diagram. We stress that the susceptibility we have introduced might
47 be considered as a measure of the effect of the noise on portfolio selection and is very reminiscent of the measure
ly

48 proposed in [11].
49 In order to present a clean, analytic picture, we have made several simplifying assumptions in this work. We have
50 omitted the constraint on the returns, liquidity constraints, correlations between the assets, non-stationary effects,
51 etc. Some of these can be systematically taken into account and we plan to return to these finer details in a subsequent
52 work.
53 Acknowledgments. We thank O. C. Martin, and M. Potters for useful discussions, and particularly J. P. Bouchaud
54 for a critical reading of the manuscript. S. C. is supported by EC through the network MTR 2002-00319, STIPCO,
55 I.K. by the National Office of Research and Technology under grant No. KCKHA005.
56
57
58
59
60

E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf

Page 9 of 10 Quantitative Finance Page 21 of 22
9
1
2 APPENDIX A: THE REPLICA SYMMETRIC SOLUTION
3
4 We show in this appendix how the minimum cost function corresponding to the replica-symmetric ansatz is obtained.
5 The ‘TrLogQ’ term in (16) is computed by realizing that the eigenvalues of such a symmetric matrix are (q1 + (n −
6 1)q0 ) (with multiplicity 1) and (q1 − q0 ) with multiplicity n − 1. Then,
7  
q1
8 Tr log Q = log detQ = log(q1 + (n − 1)q0 ) + (n − 1) log(q1 − q0 ) = n log ∆q + + O(n2 ) , (A1)
∆q
9
10 where ∆q ≡ q1 − q0 . The effective partition function in (17) depends on Q−1 , whose elements are:
11 (
12 −1 ab (∆q − q0 )/(∆q)2 + O(n) if a = b
(Q ) = (A2)
13 −q0 /(∆q)2 + O(n) if a =
6 b
14 2
ds
15 By introducing a Gaussian measure dPq0 (s) ≡ √2πq 0
e−s /2q0 , one can show that
16 (Z Z )
Fo

17 1 1 Y 1
P a 2 P a a s
P a
− 2∆q (x ) +γ (x +v)θ(−x −v) x
log Ẑ(v, q1 , q0 ) = log dxa e a a dPq0 (s)e ∆q a
18 n n a
19 Z
q0
20 = + dPq0 (s) log Bγ (s, v, ∆q) + O(n) (A3)
rP

21 2∆q
22 where we have defined
23 Z  
(x − s)2
peer-00568435, version 1 - 23 Feb 2011

24 Bγ (s, v, ∆q) ≡ dx exp − + γ(x + v)θ(−x − v) . (A4)

2∆q
ee

25
26 The exponential in (16) now reads expN n[S(q0 , ∆q, q̂0 , ∆q̂) + O(n)], where
27 Z
28 S(q0 , ∆q, q̂0 , ∆q̂) = q0 ∆q̂ + q̂0 ∆q + ∆q∆q̂ − ∆q̂ − γt(1 − β)v − t log γ + t
dPq0 (s) log Bγ (s, v, ∆q)
rR

29   (A5)
30 t 1 q̂0 log 2
− log ∆q − log ∆q̂ + − .
31 2 2 ∆q̂ 2
32 The saddle point equations for q̂0 and ∆q̂ allow then to simplify this expression. The free energy (−γ)fγ =
33
ev

limn→0 ∂Zγn /∂n is given by

34
Z
35 1 1−t q0 − 1
36 −γfγ (v, q0 , ∆q) = − t log γ + log ∆q + − γt(1 − β)v + t dPq0 (s) log Bγ (s, v, ∆q) , (A6)
2 2 2∆q
iew

37
38 where the actual values of v, q0 and ∆q are fixed by the saddle point equations
39 ∂fγ ∂fγ ∂fγ
= = =0. (A7)
40 ∂v ∂q0 ∂∆q
41
A close inspection of these saddle point equations allows one to perform the low temperature γ → ∞ limit by assuming
42
that ∆q = ∆/γ while v and q0 do not depend on the temperature. In this limit one can show that
43
On


44 
s + v + ∆/2 s < −v − ∆
45 1
lim log Bγ (s, v, ∆/γ) = −(v + s)2 /2∆ −v − ∆ ≤ s < −v (A8)
46 γ→∞ γ 
0 s ≥ −v
47
ly

48 If we plug this expression into eq. (A6) and perform the large-γ limit we get the minimum cost:
49 Z −∆   Z 0
2 2
50 q0 − 1 dx − (x−v) ∆ t dx − (x−v)
E = lim fγ = − + t(1 − β)v − t √ e 2q0 x+ + √ e 2q0 x2 . (A9)
51 γ→∞ 2∆ −∞ 2πq0 2 2∆ −∆ 2πq0
52
We rescale x → x∆, v → v∆, and q0 → q0 ∆2 , and after some algebra we obtain eq. (19).
53
54
55
56
57 [1] see P. Artzner, F. Delbaen, J. M. Eber, and D. Heath, Mathematical Finance 9, 203–228 (1999), for an axiomatic definition
58 of coherence.
59
60

E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf

Page 22 of 22 Quantitative Finance Page 10 of 10
10
1
2 [2] C. Acerbi, and D. Tasche, Journal of Banking and Finance 26, 1487–1503 (2002).
[3] R. Frey, and A. J. McNeil, (2002).
3
[4] G. C. Pflug, in “Probabilistic Constrained Optimization: Methodology and Applications”, S. Uryasev (ed.), Kluwer Aca-
4 demic Publisher (2000).
5 [5] R. Rockafellar, and S. Uryasev, The Journal of Risk 2, 21–41 (2000).
6 [6] I. Kondor, Noise sensitivity of portfolio selection under various risk measures, Lectures given at the Summer School on
7 “Risk Measurement and Control”, Rome, June 9-17, 2005; I. Kondor, S. Pafka, and G. Nagy, Noise sensitivity of portfolio
8 selection under various risk measures, submitted to J. of Banking and Finance.
9 [7] see e.g. F. Guerra, F.L. Toninelli, J. Stat. Phys. 115, 531 (2004) and references therein.
10 [8] M. Mézard, G. Parisi, and M. .A. .Virasoro, “Spin Glass theroy and Beyond”, World Scientific Lecture Notes in Physics
11 Vol. 9, Singapore (1987).
[9] M. Talagrand, Spin Glasses: a Challenge for Mathematicians, Spinger-Verlag (2002).
12
[10] W. H. Press, S. H. Teukolsky, W. T. Wetterling, and B. P. Flannery, “Numerical Recipes in C”, Cambridge University
13 Press (Cambridge, UK, 1992).
14 [11] S. Pafka, and I. Kondor, Europ. Phys. J. B27, 277 (2002).
15 [12] see [2] for the subtleties related to a discrete distribution
16
Fo

17
18
19
20
rP

21
22
23
peer-00568435, version 1 - 23 Feb 2011

24
ee

25
26
27
28
rR

29
30
31
32
33
ev

34
35
36
iew

37
38
39
40
41
42
43
On

44
45
46
47
ly

48
49
50
51
52
53
54
55
56
57
58
59
60

E-mail: quant@tandf.co.uk URL://http.manuscriptcentral.com/tandf/rquf

View publication stats