Omega and Sharpe ratio

Eric Benhamou, Beatrice Guez, Nicolas Paris

To cite this version:

Eric Benhamou, Beatrice Guez, Nicolas Paris. Omega and Sharpe ratio. 2020. �hal-02886481�

HAL Id: hal-02886481

https://hal.science/hal-02886481
Preprint submitted on 1 Jul 2020

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Omega and Sharpe ratio
Eric Benhamou∗1,2 , Beatrice Guez1 , and Nicolas Paris1
1
A.I Square Connect
2
Lamsade PSL

October 16, 2019

Abstract return level. In an enlightening research, Keating and

Shadwick (2002b) and Keating and Shadwick (2002a)
Omega ratio, defined as the probability-weighted ra- introduce the Omega ω ratio, and claimed that this
tio of gains over losses at a given level of expected universal performance measure, designed to redress
return, has been advocated as a better performance the information impoverishment of traditional mean-
indicator compared to Sharpe and Sortino ratio as it variance analysis would address these concern. They
depends on the full return distribution and hence en- emphasize that the Omega metric has the great ad-
capsulates all information about risk and return. We vantage over traditional measures to encapsulate all
compute Omega ratio for the normal distribution and information about risk and return as it depends on
show that under some distribution symmetry assump- the full return distribution, as well as to avoid looking
tions, the Omega ratio is oversold as it does not pro- at a specific level as this measure is provided for all
vide any additional information compared to Sharpe level, hence entitling each investor to look at his/her
ratio. Indeed, for returns that have elliptic distribu- risk appetite level. Strictly speaking, the omega ra-
tions, we prove that the optimal portfolio according to tio is defined as the probability-weighted ratio of gains
Omega ratio is the same as the optimal portfolio ac- over losses at a given level of expected return. In fi-
cording to Sharpe ratio. As elliptic distributions are a nancial words, this ratio determines the quality of the
weak form of symmetric distributions that generalized investment bet relative to the return threshold. As nice
Gaussian distributions and encompass many fat tail as it may sound, we argue that this is oversold, as for
distributions, this reduces tremendously the potential a large class of returns distributions, that is distribu-
interest for the Omega ratio. tions that are elliptic, we prove in this paper that the
optimal portfolio according to the Omega ratio is also
keywords: Omega ratio, Sharpe ratio, normal distri-
the optimal portfolio according to the Sharpe ratio. In
bution, elliptical distribution.
other words, optimizing weights for a given portfolio
of assets in order to get the optimal Omega ratio is
1 Introduction equivalent to optimizing the portfolio weights to find
the optimal Sharpe ratio. As elliptic distributions are
Omega ratio has been introduced on the pledge that a weak form of symmetric distributions that general-
the Sharpe ratio and many performance measurement ized Gaussian distributions and encompass many fat
and risk ratios rely on two excessive simplifications. tail distributions, this kills the potential interest of the
The first one states that a limited numbers of statisti- Omega ratio.
cal characteristics can fully describe returns. Typically
mean and variance, or first and second statistical mo-
ments for Sharpe or information ratio, mean and down- 2 Related Work
side standard deviation for Sortino ratio. The second
one states that this performance ratios should have a Omega ratio have been studied in many papers pro-
∗ [email protected]/dauphine.eu. ducing a vast literature on it. We will review here the
The authors would like to thank Francois Bertrand and Stephane main papers. Winton Research (2003) is an empirical
Mysona for fruitful conversations about the Omega ratio. research work that looked at the historical performance

1
of CTAs hedge funds and that put light on Omega ra- at-Risk at different tolerance levels and has analytical
tio. Passow (2004) looked at the analytical property closed-form expressions for commonly used distribu-
and tractabilty of the Omega ratio for Johnson distri- tion like Normal, Log-normal, Student-t and Gener-
butions. Kaffel and Prigent (2010) investigated perfor- alized Pareto distributions. They showed that under
mance measurement for financial structured products, certain condition, a subset of GlueVaR risk measures
thanks to the so called Sharpe-Omega ratio that is an fulfils the property of tail-subadditivity Sharma et al.
extension of the Omega ratio. Their originality was (2016) worked on the threshold to be used in portfolio
to compute downside risk measure using put option optimization with Omega ratio. In order to maximize
volatility instead of historical volatility. This allows the Omega ratio for the overall portfolio, one needs to
them to take account of the asymmetry of the return consider a threshold point to compute the Omega ratio
probability distribution. They determined that the op- as optimizing at all thresholds is not realistic. They de-
timal combination of risk free, stock and call/put in- cided to use the conditional value-at-risk at an α confi-
struments with respect to this performance measure, dence level CVaRα of the benchmark market. They ar-
is not necessarily increasing and concave as opposed to gue that this α-value reflects the attitude of an investor
traditional optimal Sharpe ratio portfolio for the same towards losses. Like in Kapsos et al. (2014), this for-
instruments. Similarly, Gilli et al. (2011) studied port- mulation can be cast as a linear program for mixed and
folios using the Omega function, looking at their empir- box uncertainty sets and a second order cone program
ical performance, especially the effects of allowing short under ellipsoidal sets, and hence becomes tractable.
positions. Their originality was to consider short po- They showed that the optimal portfolios resulting from
sition which is traditionally ignored. They found that the Omega-CVaRα model exhibit a superior perfor-
overall, short positions can improve risk-return charac- mance compared to the classical CVaRα model in the
teristics of a portfolio but mitigated this findings with sense of higher expected returns, Sharpe ratios, modi-
the constraints involved in short positions that often fied Sharpe ratios, and lesser losses in terms of VaRα
carries additional constraints in terms of transactions and CVaRα values. Guo et al. (2016) worked on the
costs and liquidity. property for one asset to have a higher Omega ratio
Bertrand and luc Prigent (2011) analyzed the per- than a second one. They showed that second-order
formance of two main portfolio insurance methods, the stochastic dominance (SSD) and/or second-order risk
OBPI and CPPI strategies, using downside risk mea- seeking stochastic dominance (SRSD) alone for any two
sures, thanks to Omega measure. They showed that prospects is not sufficient to imply that the Omega
CPPI strategies perform better than OBPI. Kapsos ratio of one asset is always greater than that of the
et al. (2014) looked at the maximum Omega ratio. other one. Indeed, they proved that the second-order
They established that it can be computed as a linear stochastic dominance only implies higher Omega ratios
program optimization problem. While the Omega ra- only for thresholds that are between the mean of the
tio is theoretically a nonconvex function, Kapsos et al. smaller-return asset and the mean of the higher-return
(2014) l showed that this can be reformulated as a con- asset. When considering first-order stochastic domi-
vex optimization problem that can be solved thanks to nance, the restriction on the thresholds does not apply
a linear program. This convex reformulation for Omega and first-order stochastic dominance implies preference
ratio maximization can be seen as a direct analogy of the corresponding Omega ratios for any threshold.
of the mean-variance framework for the Sharpe ratio Krezolek and Trzpiot (2017) introduced an extension
maximization and paved the way for our work that of Omega ratio called GlueVaR risk measure and illus-
will show the strong connection between Omega and trated this on metals market investments. GlueVaR
Sharpe ratio. van Dyk et al. (2014) provided a nice risk measures combine Value-at-Risk and Tail Value-
empirical research on the difference of ranking between at-Risk at different tolerance levels and have analytical
Sharpe and Omega ratio. They compared the ranking closed-form expressions for the most frequently used
of 184 international long/short (equity) hedge funds, distribution functions in many applications, i.e. Nor-
over the period January 2000 to December 2011 using mal, Log-normal, Student-t and Generalized Pareto
their monthly returns. They concluded that Omega distributions. Metel et al. (2017) is an illuminating
ratio does indeed provide useful additional informa- paper as it is the first to notice the correspondence be-
tion to investors compared to the one only provided tween Sharpe and Omega ratio under jointly elliptic
by Sharpe ratio alone. Belles-Sampera et al. (2014) distributions of returns. Compared to our work, their
generalized Omega ratio in a so called GlueVaR risk proof is more convoluted and does not emphasize the
measure. It combines Value-at-Risk and Tail Value- important fact that elliptic distributions satisfy some

2
symmetry properties that validates the proof. Rambo 4 Elliptical distributions
and Vuuren (2017) ranked fund returns and compared
results obtained with those obtained from the Sharpe It is well known that a normal distribution is fully char-
ratio over two periods: 2001 to 2007 and 2008 to 2013. acterized by its first and second moments. It is less well
They found that Omega ratio provides far superior known or at least less emphasized that the normal dis-
rankings. Guo et al. (2018) investigated whether there tribution also has a very nice property in terms of sym-
is any Sharpe or Omega ratio rule that prove that metry with respect to its first and second moments. If
one asset outperforms another one. They found that one plots iso-density curves in two or three dimension
Sharpe ratio rule is not able to detect preference under for the multi variate normal, one would obtain ellip-
general strong dominance cases. In contrast to mean- soid as illustrated by 1 and 2. This symmetry accord-
variance rule implied by Sharp ratio that does not work ing to axes leads to the so called elliptic distributions
under first order dominance, Omega ratio can help de- called like this because they are elliptically contoured
tecting better performance under first order stochastic distributions. Elliptic distribution were introduced by
dominance. Caporin et al. (2018) is a nice work on the Kelker (1970) and further studied in Cambanis et al.
overall developments on Omega ratio over the last two (1981) and by Fang et al. (1990).
decades. They emphasized two flaws of Omega ratio.
First, Omega ratio does not comply with Second-order
Stochastic Dominance as already noted by Guo et al.
(2016). Second, trade-off between return and risk cor-
responding to the Omega measure, is highly influenced
by the mean return. They illustrated their work on
long-only asset and hedge fund databases to confirm
the issues with Omega ratio. Bernard et al. (2019)
proved that in a continuous-time setting, the problem
of maximizing Omega ratio is ill-posed and leads to
excessive risk taking. They investigated if additional
constraints could offset the Omega ratio risk problem
but concluded that this was not obvious and caution
should be taken when using Omega ratio for making
asset allocation decisions.

3 Contribution and paper out- Figure 1: Contour plot for normal with positive corre-
line lation

In this paper, we first present elliptical distributions

and some of their key properties in section 4. We 4.1 spherical distributions
emphasize that these distributions encompasses many
A first intuitive idea is to define spherical distributions.
standard distributions and are a generalization of nor-
Let us denote by U ∈ Rn×n an orthogonal matrix,
mal distribution in terms of symmetry assumptions. 5
that is UUT = UT U = In . The matrix U defines an
we provide the exact computation of the Omega ra-
orthogonal linear transformation LU : Rn×n 7→ Rn×n
tio for the normal distribution and show that optimiz-
defined by X → UX. Please note that this is not
ing the Omega ratio for a portfolio is indeed similar
necessarily a rotation as the matrix associated with the
to optimizing the Sharpe ratio. We extend this result
transformation is not necessarily a rotation matrix. It
by noticing that the equivalence between Omega and
does not necessarily comply with det(U) = 1.
Sharpe ratio relies on properties of symmetry of the
normal distribution and can be extended to elliptical Definition 4.1. A random vector X = (X1 , . . . , Xn )
distributions. We show that this result can be proved has a spherical distribution if it is invariant according
very rapidly thanks to the specific nature of the Omega to any orthogonal linear transformation, that is for any
ratio. We conclude that Omega ratio is oversold and U ∈ Rn×n such that UUT = UT U = In , we have
is in most cases not providing additional information UX ∼ X where X ∼ Y means that X and Y are equal
compared to stochastic dominance. in law or in distribution.

3
 Remark 4.2. The random variable R is often referred
to as the radius of the spherical distribution and can be
interpreted geometrically in dimension two or three as
the mean radius of the ellipsoid, also seen as the radius
the major axis of the ellipse.
Remark 4.3. Spherical distributions means that there
are distributions that comply with some symmetrical
properties (invariance along any rotation but also along
any orthogonal linear transformation). This should
not be confused with distributions on the sphere like
the Fisher–Bingham or Kent distribution, the Von
Mises–Fisher distribution or even the Bingham dis-
tribution that are sometimes incorrectly called spher-
ical distributions! They are not at all spherical dis-
tributions but rather distributions on the sphere which
means that they are probability distribution such that
Figure 2: Contour plot for normal with weak correla- the probability assigned to the unit sphere is 1 and 0
tion elsewhere. In general, a spherically distributed random
vector does not have its density support restricted to the
Remark 4.1. In particular, the distribution of a sphere. It does not, either necessarily possess a density.
spherical distributed random variable, X, is invariant However, if it does, the marginal densities of dimension
under rotations as rotations matrices are special cases smaller than n − 1 are continuous and the marginal
of orthogonal matrices. densities of dimension smaller than n − 2 are differen-
tiable (except possibly at the origin in both cases). Uni-
It can be shown easily (see Fang et al. (1990)) that X variate marginal densities for n greater than 2 are non-
is a spherical distribution is equivalent to the existence decreasing on (−∞, 0) and non-increasing on (0, ∞).
of a function φ(.) such that for any t ∈ Rn , we have
Remark 4.4. A typical example of a spherical dis-
T 2 2
ΦX (t) = φ(t t) = φ(t1 + . . . + tn ) (1) tribution is the multi variate normal distribution with
covariance matrix proportional to the identity. The
where ΦX (t) denotes the characteristic function multivariate t-distribution is a typical example of a fat
T
ΦX (t) = E[eit X ]. The function φ is referred to as tailed spherical distribution. Let Z be a multi variate
the generator of the distribution and it is quite com- spherical normal distribution Z ∼ Nn (0, In ) and R be
2
mon to write X ∼ Sn (φ) where Sn denotes the spheri- a chi squared distribution R ∼ χk with k degrees of
cal distribution in R . A nice theorem that translates freedom independent of Z .The random vector
n

into an enlightening geometrical interpretation is that √ Z

there exists a representation form for any spherical dis- Y = k (4)
R
tribution. This is the following theorem known as the
spherical representation theorem: has a multivariate t-distribution with k degrees of free-
dom. Thanks to this writing, it is easy to see that its
Theorem 4.1. A random vector X = (X1 , . . . , Xn ) belongs to the family of n-dimensioned spherical distri-
has a spherical distribution if and only if there exists butions.
S a random variable uniformly distributed on the unit
 Obviously, for multi variate Gaussian, imposing that
sphere S n−1 = s ∈ Rn : ssT = 1 and R ≥ 0 a ran-
the covariance matrix is proportional to the identity
dom variable independent of S such that
matrix is too restrictive. This naturally leads to the
X ∼ RS. (2) extension of spherical distributions that are elliptical
distributions.
If in addition, the scalar random variable R has a finite
second moment E[R2 ] < ∞, then the first two moments 4.2 Elliptical distributions
of X exist and are given by:
Definition 4.2. A (n × 1) random vector X is said
E[X] = 0, Cov(X) = 1/n E[R2 ]In . (3) to have an elliptical distribution with parameters µ a

4
(n×1) constant vector and Σ a (n×n) constant matrix • Assume that X ∼ En (µ, Σ, φ) with rank(Σ) = k.
if X has the same distribution as µ + Λ> Y, where Y ∼ Then
Sn (φ) and Λ is a (k × n) matrix such that Λ> Λ = Σ Q(X) = (X − µ)> Σ− (X − µ)
with rank(Σ) = k. We shall write X ∼ ECn (µ, Σ, φ). has the same distribution as R2 in equation 5
Remark 4.5. Elliptical distributions can be seen as • An elliptic distribution has a density function of
an extension of the multi variate normal distribution the form
denoted by Np (µ, Σ).
Let X be a multinormal distribution X ∼ Nn (µ, Σ). f (x) = C · g (x − µ)> Σ−1 (x − µ)


Then X ∼ En (µ, Σ, φ) and φ(u) = exp (−u/2). As
the density surface for the multivariate normal dis- where C is the normalizing constant, x is an n-
1
tribution is given by f (x) = det(2πΣ)− 2 exp{− 12 (x − dimensional random vector with median vector mu
µ)> Σ−1 (x − µ)}, it is easy to see that this density is (which is also the mean vector if the latter exists),
constant on ellipses (see for instance 1). This explains and Σ is a positive definite matrix which is propor-
why these distributions have been called elliptical. tional to the covariance matrix if the latter exists.

Below are summarized the main properties of ellip- Remark 4.6. The second item of this theorem states
tical distributions: that Σ, φ, Λ are not unique, unless we impose a condi-
tion on the determinant, that is det(Σ) = 1.
Theorem 4.2. Elliptical random vectors X have the
following properties: Remark 4.7. The distribution function of R is a key
characteristic of the elliptical distribution and can lead
• Basic properties: any linear combination of ellip- to elliptical distributions that share most of their char-
tically distributed variables is elliptical and any of acteristic in common. Hence it is called the generat-
its marginal distributions is also elliptical. ing variate of the elliptical distribution family. Using
extreme value theory, it is useful to characterize fat
• Representation form: a scalar function φ(.) can
tailed elliptical distributions. If the generating vari-
determine an elliptical distribution En (µ, Σ, φ) for
ate R belongs to the maximum domain of attraction
every µ ∈ Rp and Σ ≥ 0 with rank(Σ) = k iff
of the Frechet distribution (Embrechts et al. (2003)),
φ(t> t) is a p-dimensional characteristic function.
this writes as F R = λ(x) × x−α for all x > 0, where
If there exist two representation forms of X such
α > 0 and λ is a slowly varying function (Resnick
that X ∼ En (µ, Σ, φ) and X ∼ En (µ∗ , Σ∗ , φ∗ ), then
(2014)). The parameter α is then called the tail in-
there exists a constant c > 0 such that
dex of the generating distribution function FR which
µ = µ∗ , Σ = cΣ∗ , φ∗ (.) = φ(c−1 .). corresponds also to the tail index of the regularly vary-
ing random vector X (Hult and Lindskog (2002)). This
• Characteristic function and representation: the shows in particular that multivariate elliptical distribu-
characteristic function of X, ΦX (t) is of the form tions allow for heavy tails while encompassing at the
same time a simple linear dependence structure from
ΦX (t) = e it> µ >
φ(t Σt) the multi variate normal. Hence, in addition to the
multi variate normal distribution, the multi variate t-
for a scalar function φ. X ∼ En (µ, Σ, φ) with distribution Fang et al. (1990), the symmetric gener-
rank(Σ) = k iff X has the same distribution as: alized hyperbolic distribution Barndorff-Nielsen et al.
(1982), the sub-Gaussian α-stable distribution (Rachev
µ + RΛ> U (k) (5) and Mittnik, 2000, p. 437) are elliptical distributions.

where R ≥ 0 is a random scalar variable indepen- Remark 4.8. Because of their representation, ellipti-
dent of U (k) . U (k) is a random vector distributed cal distributions generalizes nicely Gaussian properties.
uniformly on the unit sphere surface in Rk and Λ This is because their characteristic function writes as
is a (k × n) matrix such that Λ> Λ = Σ. >
ΦX (t) = eit µ φ(t> Σt) (6)
• Assume that X ∼ En (µ, Σ, φ) and E(R2 ) < ∞.
Then its first two moments exist and are given by: which is a weakened form of the multi variate Gaussian
distribution that is given by
E(R2 ) >
E(X) = µ, Cov(X) = Σ = −2φ (0)Σ. >
ΦX (t) = eit µ exp(−1/2t> Σt) (7)
rank(Σ)

5
Remark 4.9. Because of their nice property of gen-
eralizing multi variate Gaussian distribution with po-
tentially fat tailed distributions, elliptical distributions
are meaningful for financial data modeling. The the-
ory of portfolio optimization developed by Markowitz
(1952) and continued by Tobin (1958), Sharpe (1963)
is the basis of modern portfolio risk management. It
relies on the Gaussian distribution hypothesis and its
quintessence is that the portfolio diversification effect
depends essentially on the covariance matrix, i.e. the
linear dependence structure of the portfolio compo-
nents. Elliptical distributions generalize nicely this Figure 3: Graphical interpretation of Omega ratio Ω(θ)
portfolio theory as they cope well with linear trans- defined as the ratio of the shaded red over the blue area
formations. In particular, if the returns on all as-
sets available for portfolio formation are jointly ellipti-
infinite for large negative value of θ decreases with θ
cally distributed, then all portfolios can be characterized
and tends to 0 for large positive value of θ. The perfect
completely by their location and scale – that is, any two
case is to have one curve of a portfolio above all the
portfolios with identical location and scale of portfolio
other one for any value of θ. However, this particular
return have identical distributions of portfolio return.
case of Pareto optimality (meaning the curve is above
Various features of portfolio analysis, including mutual
all other curves for any value of θ) is extremely rare
fund separation theorems and the Capital Asset Pricing
and in practice, curves cross each other making the
Model, can be easily extended for all elliptical distribu-
call to select one asset or portfolio among the other
tions, making this kind of distributions appealing.
ones harder.

5 Omega ratio
Definition 5.1. As presented in Keating and Shadwick
(2002b) and Keating and Shadwick (2002a), for an as-
set whose return r has a cumulative probability distri-
bution function F and θ is the target return threshold
defining what is considered a gain versus a loss, the
Omega ratio is defined as
R∞
[1 − F (r)] dr
Ω(θ) = θR θ (8)
−∞
F (r) dr

When θ is set to zero the gain-loss-ratio by Bernardo

and Ledoit arises as a special case Bernardo and Ledoit
(2000). The selling point of the omega ratio compared Figure 4: Ω(θ) ratio for four portfolios in practice
to Sharpe ratio and other traditional risk ratio is that
at first sight, it seems to depend on the entire return
distribution through the cumulative probability distri- Remark 5.1. It is worth mentioning that the Omega
bution function F as well as not rely on any particular ratio can beR ∞meaningless if for instance the numer-
moments in terms of value and even existence, making ator term θ [1 − F (r)] dr or the denominator term
Rθ
it intellectually very attractive. Graphically, for a θ F (r) dr are undefined. Surprisingly, this is ig-
−∞
value of 2.70 percent and the cumulative distribution nored in the literature about the Ω(θ) ratio. A typical
given by figure 3, the Ω(θ) ratio is defined as the ratio example of a distribution that has an undefined Ω(θ)
of the red area over the blue area. ratio is a power law F ∼ |r|α with 1 < α ≤ 2 .
Omega ratio has been created to be able to compare
different assets in terms of their risk profile. Figures Remark 5.2. The case of the power law for α = 2
4 and 5 provide in log scale the omega ratio for four is illuminating. A corresponding elliptical distribution
portfolio. The omega ratio is a real function that is is the Cauchy distribution whose density is given by

6
 As said in the introduction, Sharpe and Omega ra-
tio have been strongly opposed. We shall see this
is not completely correct. In the particular case of
a normal probability distribution function F , it is
fairly easy to compute the Ω(θ). Let us denote by
ψ the standard normal probability distribution func-
tion ψ(x) = √12π exp(−x2 /2) and Ψ(x), the standard
normal Rcumulative probability distribution function
x
Ψ(x) = −∞ ψ(u)du. We have

Proposition 5.2. If returns follow a normal proba-

bility distribution function N (µ, σ), the Ω(θ) ratio is
given by

Figure 5: Ω(θ) ratio for six portfolios in practice 1

Ω(θ) = 1 + ψ(S(θ))
(12)
S(θ) − Ψ(−S(θ))
1
f (r) = h 2 ,
i where µ is the location parame- Proof. Given in appendix 7.2
πσ 1+( r−µ
σ )

ter and σ the scale parameter. Its cumulative
density If we explicitly expand equation (12) for the normal
function is given by F (r) = π1 arctan r−µ σ + 21 . To
distribution, we get
keep things simple, let us take the case of µ = 0 and
σ = 1, which is the normalized Cauchy distribution. √
2π
The numerator of the omega ratio is then given as the Ω(θ) = 1 + exp(−S(θ)2 /2) R −S(θ)
limit for A → +∞ of S(θ) − −∞ exp(−u2 /2) du
Z A Z A (13)
1 1
[1 − F (r)] dr = − arctan(r)dr (9) Proposition 5.3. If returns follow a normal probabil-
θ θ 2 π
" 2 #A
ity distribution function N (µ, σ), the Ω(θ) ratio is an
r r arctan(r) − ln(1+r
2
)
increasing function of S(θ)
= −
2 π
θ Proof. Given in appendix 7.3
→ +∞ (10)
A→+∞ It follows that maximizing the Omega ratio
This shows that whenever, we speak about Omega ratio, at level θ is equivalent to maximize the Sharpe
we need to impose that the two terms (numerator and ratio for a risk free rate θ. Moreover, if we are
denominator) are well defined. This is the subject of looking at returns at a given level of volatility, it means
the following proposition that we impose that σ is a constant. In this particular
case, as the Sharpe only depends on the first moment:
Proposition 5.1. The Ω(θ) ratio terms are defined if S(θ) = µ−θ σ = µ−θσ
0
+ θ0σ−θ , maximizing the Sharpe
and only if the right and left tails of the cumulative does not depend on the corresponding risk free rate
distribution are dominated by |r|1α with α > 1. This level θ and can be done at the risk free rate. We there-
means in particular that the existence of the Ω(θ) ratio fore obtain the important result given by the following
implies that lim rF (r) = 0. proposition
r→−∞

Proof. Given in appendix 7.1 Proposition 5.4. Under a target volatility constraint,
and for a normal distribution, maximizing the Ω(θ) is
In contrast, the Sharpe ratio is for a given level θ is strictly the same as maximizing the Sharpe ratio!
defined as follows:
We will see in the following that the property is not
Definition 5.2. The Sharpe ratio (for the reference restricted to normal distribution but can easily be ex-
level θ) S(θ) is defined as the excess return over θ di- tended to elliptic distributions as we have just used
vided by the standard deviation symmetry properties of the distribution. More pre-
µ−θ cisely, we can compute explicitly the Ω(θ) ratio as fol-
S(θ) = (11) lows:
σ

7
Proposition 5.5. If the Ω(θ) ratio is well defined References
and if returns follow an elliptic  probability distribu-
2 
tion function proportional to g − 12 u−µ writing Barndorff-Nielsen, O., Kent, J., and Sorensen, M.

σ ,
v (1982). Normal variance-mean mixtures and z dis-
ψ(u) = Cg(− 21 u2 ) and Ψ(v) = −∞ Cg(− 12 u2 ) du, the
R
tributions. Int. Statist. Rev., 50:145–159.
corresponding re-normalized probability and cumulative
probability distribution functions, as well as the follow- Belles-Sampera, J., Guillén, M., and Santolino, M.
ing function G(u): (2014). Beyond value-at-risk: Gluevar distortion risk
Z u measures. Risk Analysis, 34(1):121–134.
G(u) = Cg(x) dx (14)
−∞ Bernard, C., Vanduffel, S., and Ye, J. (2019). Optimal
strategies under omega ratio. European Journal of
then the Ω(θ) ratio is given by Operational Research, 275:755–767.
1
Ω(θ) = 1 + G(− 21 S(θ)2 )
(15) Bernardo, A. and Ledoit, O. (2000). Gain, loss,
S(θ) − Ψ(−S(θ)) and asset pricing. Journal of Political Economy,
108(1):144–172.
Proof. Given in appendix 7.4
Bertrand, P. and luc Prigent, J. (2011). Omega perfor-
RemarkR 5.3. The re-normalizing constant C is de- mance measure and portfolio insurance. Journal of
∞
fined as −∞ Cg(− 12 u2 ) du = 1, or equivalently Banking and Finance, 35:1811–1823.
1 Cambanis, S., Huang, S., and Simons, G. (1981).
C = R∞ 1 2
(16)
g(− On the theory of elliptically contoured distributions.
−∞ 2 u ) du
Journal of Multivariate Analysis, 11(3):368–385.
Remark 5.4. In the case of the normal distribution,
the function g is simply the exponential function. Caporin, M., Costola, M., Jannin, G., and Maillet, B.
(2018). On the (ab)use of omega? Journal of Em-
Proposition 5.6. For elliptic distribution, the Ω(θ) pirical Finance, 46:11–33.
ratio is an increasing function of S(θ).
Embrechts, P., Mikosch, T., and Klüppelberg, C.
Proof. Given in appendix 7.5 (2003). Modelling Extremal Events: For Insurance
We come now to the central property and Finance. Springer-Verlag, Berlin, Heidelberg.

Proposition 5.7. Under a target volatility constraint, Fang, K., Kotz, S., and Ng, K. (1990). Symmetric
and for an elliptic distribution, maximizing the Ω(θ) is multivariate and related distributions. Number 36
strictly the same as maximizing the Sharpe ratio! in Monographs on statistics and applied probability.
Chapman & Hall, London [u.a.].

6 conclusion Gilli, M., Schumann, E., di Tollo, G., and Cabej,

G. (2011). Constructing 130/30-portfolios with
In this paper, we have shown that under elliptic dis- the omega ratio. Journal of Asset Management,
tributions assumptions, with tail dominated by 1/rα 12(2):94–108.
for some α > 1, the Ω(θ) ratio is well defined. Under
these assumption, the Ω(θ) ratio can be computed in Guo, X., Jiang, X., and Wong, W.-K. (2016). A
closed form and is an increasing function of the Sharpe note on stochastic dominance and the omega ratio.
ratio for a risk free rate taken to θ. Assuming that SSRN:2827058.
we impose a target volatility constraint, meaning we Guo, X., Levy, H., Lu, R., and Wong, W.-K. (2018).
only care about portfolio for a given level of volatility, Could omega ratio perform better than sharpe ratio.
maximizing the Ω(θ) ratio is not any more dependent SSRN:3198033.
on the parameter θ and is obtained by maximizing the
true Sharpe ratio for the real risk free rate. This means Hult, H. and Lindskog, F. (2002). Multivariate ex-
that under target volatility constraint, first order dom- tremes, aggregation and dependence in elliptical dis-
inance (higher first moment) provides the best Sharpe tributions. Advances in Applied Probability, 34:587–
and the best Omega ratio. 608.

8
Kaffel, H. R. and Prigent, J.-L. (2010). Struc- van Dyk, F., Vuuren, G., and Heymans, A. (2014).
tured portfolio analysis under sharpe-omega ratio. Hedge fund performance evaluation using the sharpe
SSRN:1792807. and omega ratios. International Business & Eco-
nomics Research Journal, 13:485–512.
Kapsos, M., Zymler, S., Christofides, N., and Rustem,
B. (2014). Optimizing the omega ratio using linear Winton Research (2003). Assessing cta quality with
programming. Journal of Computational Finance, the omega performance measure. Winton Capital
17:49–57. Management Research.

Keating, C. and Shadwick, W. (2002a). An introduc-

tion to omega. The Finance Development Centre 7 Appendix
Limited.
7.1 Proof of Proposition 5.1
Keating, C. and Shadwick, W. (2002b). A universal
performance measure. Journal of Performance Mea- The case of the numerator and denominator are the
surement, 6. same. Hence, let us only prove the case for the de-
nominator. Using the comparison test for improper
Kelker, D. (1970). Distribution theory of spherical dis- integral and the fact that F is always positive, we have
tributions and a location-scale parameter generaliza- that θ F (r) dr is well defined for A → ∞ if and only
R
−A
tion. Sankhyā, Series A, 32:419–430. F is dominated by some function |r|1α for α > 1. To
conclude the proof, let us first notice that
Krezolek, D. and Trzpiot, G. (2017). The effectiveness
of the gluevar risk measure on the metals market - lim rF (r) = − lim rF (−r) (17)
the application of omega performance measure. Folia r→−∞ r→∞

Oeconomica. As F is dominated by 1 for α > 1 for large nega-

|r|α
tive value of r, we have easily that
Markowitz, H. (1952). Portfolio selection. Journal of
Finance, 7(1):77–91. 1
lim rF (−r) ≤ lim (18)
r→∞ r→∞ r α−1
Metel, M. R., A. Pirvu, T., and Wong, J. (2017). Risk → 0 (19)
management under omega measure. Risks, 5(2). r→∞

which shows that lim rF (r) = 0 .

Passow, A. (2004). Omega Portfolio Construction with r→−∞
Johnson Distributions. International Center for Fi-
nancial Asset Management and Engineering, FAME 7.2 Proof of Proposition 5.2
Research Paper Series.
If returns follow a normal probability distribution func-
Rambo, J. and Vuuren, G. (2017). An omega ratio tion N (µ, σ), the Omega ratio, Ω(θ), writes as:
analysis of global hedge fund returns. Journal of R∞
[1 − Ψ r−µ


θ σ ] dr
Applied Business Research (JABR), 33:575. Ω(θ) = Rθ r−µ

(20)
−∞
Ψ σ dr
R∞
Resnick, S. (2014). Extreme Values, Regular Variation θ−µ [1 − Ψ (x)] σ dx
and Point Processes. Springer. = σ
R θ−µ
−∞
σ
Ψ (x) σ dx
Sharma, A., Utz, S., and Mehra, A. (2016). Omega- R∞
−S(θ)
[1 − Ψ (x)] dx
cvar portfolio optimization and its worst case analy- = (21)
R −S(θ)
sis. OR Spectrum, 39. −∞
Ψ (x) dx

Sharpe, W. (1963). A simplified model for portfolio where S(θ) = µ−θ . Using an integration by parts, we
σ
analysis. Management Science, 9(2):277–293. have
Z a Z a
Tobin, J. (1958). Liquidity preference as behavior to- a
Ψ (x) dx = [xΨ(x)]−∞ − xψ (x) dx
wards risk. Review of Economic Studies, 25(2):65– −∞ −∞
86. = aΨ(a) + ψ (a) (22)

9
Using this result and the fact that ψ(u) = ψ(−u), 1 − where we have used that limv→−∞ vΨ(v) = 0 and
Ψ (x) = Ψ (−x), we have: limv→−∞ vψ(v) = 0.
Z −S(θ) Let us come back to our initial problem about com-
puting the Ω(θ) ratio. We have
Ψ (x) dx = ψ(S(θ)) − S(θ)Ψ(−S(θ)) (23)
−∞ R∞
[1 − Ψ r−µ

dr
Z ∞
θ σ ] σ
[1 − Ψ (x)] dx = ψ(S(θ)) − S(θ)Ψ(−S(θ)) + S(θ) Ω(θ) = Rθ r−µ dr

(28)
−S(θ) −∞
Ψ σ σ
R∞
θ−µ [1 − Ψ (x)] dx
Hence the result = σ
R θ−µ
1 −∞
σ
Ψ (x) dx
Ω(θ) = 1 + ψ(S(θ)) (24) R S(θ)
S(θ) − Ψ(−S(θ)) −∞
Ψ (x) dx
= R −S(θ) (29)
which concludes the proof −∞
Ψ (x) dx

µ−θ
where S(θ) = σ . Using the intermediate result (27),
7.3 Proof of Proposition 5.3
we have
As a function of S(θ), the Ω(θ) ratio ’s derivative with
respect to S(θ) is given by S(θ)Ψ(S(θ)) + G(− 12 S(θ)2 )
Ω(θ) = (30)
−S(θ)Ψ(−S(θ)) + G(− 21 S(θ)2 )
∂ ψ(S(θ)) 1
Ω(θ) = 2 (25) = 1 + G(− 1 S(θ)2 ) (31)
∂S(θ)

ψ(S(θ))
S(θ)2 S(θ) − Ψ(−S(θ)) 2
S(θ) − Ψ(−S(θ))

The strict positiveness of the derivatives concludes the which concludes the proof
proof that Ω(θ) ratio is an increasing function of S(θ).
7.5 Proof of Proposition 5.6
7.4 Proof of Proposition 5.5 Under the elliptical assumption for the distribution,
proposition (5.5) applies. As a function of S(θ), the
With notations of Proposition 5.5, we can remark that Ω(θ) ratio ’s derivative with respect to a is given by
0
ψ(u) = ψ(−u), 1 − Ψ(u) = Ψ(−u), Ψ (u) = ψ(u) and
that G(u) is defined (see equation (14)) such that ∂ ∂ 1
Ω(θ) = G(− 21 S(θ)2 )
(32)
∂S(θ) ∂S(θ) − Ψ(−S(θ))
∂ 1 S(θ)
G(− u2 ) = −uψ(u) (26) ∂
G(− 12 S(θ)2 ) G(− 12 S(θ)2 )
∂u 2 ∂S(θ)
−
+ ψ(−S(θ))
S(θ) S(θ)2
We can also remark that the proper definition of the = −  1 2
2
G(− 2 S(θ) )
Ω(θ) ratio implies (see proposition 5.1) that Ψ(v) is S(θ) − Ψ(−S(θ))
dominated by some function |v|1α for some α > 1 for 1
G(− 2 S(θ)2 )
large negative value of v. In particular, we have that =  2 (33)
S(θ) 2 ψ(S(θ)) − Ψ(−S(θ))
limv→−∞ vΨ(v) = 0. The fact that Ψ(v) is dominated S(θ)
by some function |v|1α for some α > 1 for large negative
1 where we have used again that
value of v implies also that g is dominated by α+1
|v| 2
for some α > 1 for large negativeRvalue of v, hence the ∂ 1
u G(− u2 ) = −uψ(u) (34)
function G defined by G(u) = −∞ Cg(x) dx is well ∂u 2
defined. We also have that limv→−∞ vψ(v) = 0.
and the parity of the function ψ: ψ(−u) = ψ(u). The
We can therefore compute the following integral by
strict positiveness of the derivatives concludes the proof
integration by parts:
that Ω(θ) ratio is an increasing function of S(θ).
Z a Z a
a
Ψ (x) dx = [xΨ(x)]−∞ − xψ(x) dx
−∞ −∞
1
= aΨ(a) + G(− a2 ) (27)
2

10