Elsevier

Journal of Empirical Finance

Volume 46, March 2018, Pages 11-33
Journal of Empirical Finance

“On the (Ab)use of Omega?”

https://doi.org/10.1016/j.jempfin.2017.11.007Get rights and content

Highlights

  • We question on the relevance of Omega ratio.

  • We show that the Omega ratio is inconsistent with the second-order stochastic dominance.

  • The trade-off between return and risk is essentially influenced by the mean return.

  • Omega-based portfolios can be associated with classical optimization paradigms.

  • Robustness checks on long-only asset and hedge fund databases confirm the results.

Abstract

Several recent finance articles use the Omega measure (Keating and Shadwick, 2002), defined as a ratio of potential gains out of possible losses, for gauging the performance of funds or active strategies, in substitution of the traditional Sharpe ratio, with the arguments that return distributions are not Gaussian and volatility is not always the relevant risk metric. Other authors also use Omega for optimizing (non-linear) portfolios with important downside risk. However, we question in this article the relevance of such approaches. First, we show through a basic illustration that the Omega ratio is inconsistent with the Second-order Stochastic Dominance criterion. Furthermore, we observe that the trade-off between return and risk corresponding to the Omega measure, may be essentially influenced by the mean return. Next, we illustrate in static and dynamic frameworks that Omega-based optimal portfolios can be closely associated with classical optimization paradigms depending on the chosen threshold used in Omega. Finally, we present robustness checks on long-only asset and hedge fund databases, that confirm our results.

Introduction

A relatively wide stream of the financial economics literature focuses on performance measurement with two main motivations: the introduction of measures capturing stylized facts of financial returns such as asymmetry or non-Gaussian densities, among many others; the use of performance measures in empirical studies dealing with managed portfolio evaluation or asset allocation. The interest in this research field started with the seminal contribution of Sharpe in 1966. An increasing number of studies appeared in the following decades (see the survey by Caporin et al., 2014). One of the most studied topics is the performance evaluation of active management, which probably represents the main element promoting a renewed interest in performance measurement. Among the most recent contributions, we can quote Cherny and Madan (2009), Capocci (2009), Darolles et al. (2009), Jha et al. (2009), Jiang and Zhu (2009), Stavetski (2009), Zakamouline and Koekebakker (2009), Darolles and Gouriéroux (2010), Glawischnig and Sommersguter-Reichmannn (2010), Billio et al. (2012), Cremers et al. (2013), Kapsos et al. (2014), Weng (2014), Billio et al. (2014) and Billio et al. (2015).

Performance evaluation has relevant implications from both a theoretical point of view, as it allows us to understand agent choices, and an empirical one, since practitioners are interested in ranking assets or managed portfolios according to a specific non-subjective criterion. As an example, financial advisors often rank mutual funds according to a specific performance measure. Moreover, when rankings produced by advisors are recognized as a reference by the investors, changes in the rankings might influence inflows and outflows (see Hendricks et al. (1993), Blake and Morey (2000), Powell et al. (2002), Del Guercio and Tkac (2008), Jagannathan et al. (2010)).

A large number of performance measures have already been proposed and, as a consequence, related ranks can sensibly vary across different financial advisors, portfolio managers and investment institutions. The identification of the most appropriate performance measure depends on several elements. Among them, we cite the investors preferences and the properties or features of the analyzed asset/portfolio returns. Furthermore, the choice of the optimal performance measure, across a number of alternatives, also depends on the purpose of the analysis which might be, for instance, one of the following: an investment decision, the evaluation of managers abilities, the identification of management strategies and of their impact, either in terms of deviations from the benchmark or in terms of risk/return. Despite some known limitations, the Sharpe (1966) ratio is still considered as the reference performance measure. However, if we derive this ratio within a Markowitz framework, it shares the same drawbacks as the Mean–Variance model, where the representative investor is characterized by a quadratic utility function and/or the portfolio returns are assumed to be Gaussian.1 Clearly, it is well established that financial returns are not Gaussian, also due to investment strategies based on derivatives with time-varying exposures and leverage effects. Furthermore, numerous theoretical articles and empirical studies show that it is unlikely that investors do not care about higher-order moments (e.g. Scott and Horwath (1980), Golec and Tamarkin (1998), Harvey and Siddique (2000), Jondeau and Rockinger (2006), Jurczenko and Maillet (2006), and references herein).

As a consequence, both an incorrect assumption of Gaussianity and an under-estimation of the importance of higher-order moments for the investors, may lead to an underestimation of the portfolio total risk (see Fung and Hsieh (2001), Lo (2001), Mitchell and Pulvino (2001), Kat and Brooks (2002), Agarwal and Naik (2004)) and, consecutively, to biased investment rankings and financial decisions (a downward biased risk evaluation induces an upward biased Sharpe ratio). In this case, the Sharpe ratio does not completely reflect the attitude towards risk for all categories of investors. Moreover, the standard deviation equally weights positive and negative excess returns. In addition, it has been shown that volatility can be subject to manipulations (see Ingersoll et al., 2007). Using this quantity to evaluate the risk of assets with low liquidity can also be another issue (see Getmansky et al., 2004). Ultimately, the pertinence of using this performance measure relies heavily on the accuracy and stability of the first and second moment estimations ( Merton, 1981; Engle and Bollerslev, 1986). Finally, the so-called “μσ” Paradox illustrates that the Sharpe ratio is not consistent, in the sense of the Second-order Stochastic Dominance criterion 2 as introduced by Hadar and Russell (1969), when return distributions are not Gaussian (see Weston and Copeland (1998), Hodges (1998)).

Based on these critics, academics and practitioners proposed alternative criteria that can be adapted to more complex settings than the Gaussian one. Among the various proposed measures to remedy these aforementioned shortcomings, the Gain–Loss Ratio was introduced by Bernardo and Ledoit (2000) as a special case of the Grinblatt and Titman (1989) Positive Period Weighted Measure of performance; it was then generalized later on (with a variable threshold instead of the nil reference — see below) first by Keating and Shadwick (2002) and, secondly, by Farinelli and Tibiletti (2008), Caporin et al. (2014), Bellini and Bernardino (2015) using variants of expected shortfall and expectiles. These performance measures were originally derived to fill the gap between model-based pricing and no-arbitrage pricing, and thus belong to the so-called “Good-Deal-Bound” literature (see Cochrane and Saa-Requejo (2001), Cochrane (2009), Biagini and Pinar (2013)). This literature mainly derives price bounds by precluding Stochastic Discount Factors which would create too attractive assets, with respect to their “Good-Deal” measure in the terminology of Cochrane (2009), and thereby inclining the no-arbitrage price bounds. Cochrane (2009) derives Good-Deal price bounds for option prices by restricting the Sharpe ratio. Černỳ (2003) generalizes this approach and made it well-applicable for skewed assets, whilst Cherny and Madan (2009) study the good properties that should exhibit any good performance measure. In this context, it appears that the Gain–Loss Ratio approach provides several significant advantages. A bounded ratio implies the absence of arbitrage opportunities. Another advantage, which might be the main reason for its success, results from its asset pricing model foundation. The Gain–Loss Ratio quantifies the attractiveness of an investment concerning a benchmark investor, e.g. a representative investor of a specified asset pricing model. Alongside the determination of price bounds on incomplete markets, the Gain–Loss Ratio can also be used to compare asset pricing models or to measure the performance of funds in an economically meaningful way, i.e. concerning a benchmark Stochastic Discount Factor. In addition to the stated theoretical foundation and the preclusion of arbitrage opportunities, the Gain–Loss Ratio has many other desirable properties. It is not only very intuitive, but it also fulfills many requirements of a good performance measure, e.g. all conditions of an “acceptability index” as put forward by Cherny and Madan (2009).3

Nevertheless, some authors recently report a first notable drawback: in many standard models, e.g. the Black–Scholes model, the best Gain–Loss Ratio becomes infinity. Because of this shortcoming, Biagini and Pinar (2013) conclude that the “Gain–Loss Ratio is a poor performance measure...” (sic). One might argue that the Gain–Loss Ratio is only constructed for discrete probability space corresponding to real asset returns, but even in this case, the strong dependence on the specific discretization can be seen as a severe drawback. Voelzke (2015) recently attempted to solve this issue by proposing the so-called Substantial Gain–Loss Ratio, that allows us to work in continuous probability spaces without losing the positive properties of the Gain–Loss Ratio.

In recent years, different authors, however, widely used the extended version of the Gain–LossRatio, called the Omega performance measure as introduced by Keating and Shadwick (2002), when ordering investment portfolios according to a rational criterion.

The first set of authors consider Omega in the evaluation of active management strategies in contrast to the well-known Sharpe ratio, supporting their choice by the non-Gaussianity of returns and by the inappropriateness of volatility as a risk measure when strategies are non-linear (e.g. Eling and Schuhmacher (2007), Annaert et al. (2009), Hamidi et al. (2009), Bertrand and Prigent (2011), Ornelas et al. (2012), Zieling et al. (2014), Hamidi et al. (2014)).

A second group of articles focuses on Omega as an objective function for portfolio optimization in order to introduce downside risk in the estimation of optimal portfolio weights (e.g. Mausser et al. (2006), Farinelli et al. (2008), Farinelli et al. (2009), Kane et al. (2009), Gilli and Schumann (2010), Hentati et al. (2010), Gilli et al. (2011)). The maximization program in an Omega paradigm is indeed, per nature, a non-convex non-linear one. As formulated in Mausser et al. (2006), the investor’s problem in finding the portfolio that optimizes Omega can be written, under some mild assumptions, as a linear one, after some variable changes, inspired by the works of Charnes and Cooper (1962) in the field of production optimization.4

Nevertheless, our main point in this article is that despite the fact that several authors in the literature claimed that the Gain–LossRatio (and its generalization: the Omega measure) “is consistent with second-order stochastic dominance criterion” (sic; e.g. Darsinos and Satchell, 2004, p.80; Biagini and Pinar, 2013, p.229), it happens to be only valid for special cases; more precisely, this general statement is, unfortunately, untrue for the original general definitions of both the measures and criterion (see also Voelzke, 2015, for a generalized kernel pricing result).

Similar to Weston and Copeland (1998), Hodges (1998), who illustrate the inconsistency of the Sharpe ratio when asset return densities do not belong to the same location-scale family (see Meyer, 1987, and Eling and Schuhmacher, 2011), we show in the present article that the Omega measure has some severe drawbacks. Of course, consistence with the stochastic dominance criteria, which is entailed by traditional rational agent behavior, is not a shared characteristic of all portfolio choice theories (for instance, both Rank Dependent Expected Utility and the Cumulative Prospect Theories 5 are not compliant), but this can be justified when optimal choices can be explained (based on a subjective transformation of objective probabilities and/or values for instance), and when they better reflect preferences of investors in empirical studies. This property of consistence with the stochastic dominance should, thus, not be overrated. However, in the case of Gain–Loss and Omega ratios, it might conduct investors to misleading and irrational rankings of risky assets, even, surprisingly, in an extremely simplified framework such as the Gaussian one. The main ideas that we present hereafter are: (1) the Omega measure is biased in some cases due to the importance of the mean return in its computation, and, (2) since gains can be compensated by losses in the computation of the Omega measure, two densities with (almost) the same means can share identical Omega measures, even when the risk related to one asset is obviously higher than the other one. In other words, the Omega ratio may, in some cases, overvalue performance and simply disregard risk, which is both counter-intuitive and far more severe than just a theoretical inconsistence with stochastic dominance criteria.

As shown below, these problems mainly come from the fact that Omega indeed does not comply, in general, with the Second-order Stochastic Dominance criterion (SSD, in short), as it is falsely stated in the literature. In the limit case, this implies indifference across assets with the same expected return but with different risk levels. In other words, we show that (1) on a theoretical basis, the Omega ratio is not compliant with the Second-order Stochastic Dominance criterion; (2) on an empirical basis, using Omega ratios can lead to obvious misleading financial choices and that (3) Omega and Sharpe approaches, whilst built on different concepts and measures, give very similar results when applied to real market data (long-only and hedge funds), which is in line with the empirical results by Eling and Schuhmacher (2007), showing very little differences between rankings based on various performance measures.

Finally, if the Omega measure (1) is not compliant with the main paradigm, (2) may sometimes lead to obvious irrational choices, (3) generally reaches the same conclusions than those based on the classical Sharpe ratio, we conclude that its use should always be cautious.

The article proceeds as follows. Section 2 introduces the Omega function in the general context of performance measurement and presents some of its properties. Section 3 demonstrates that the Omega measure is inconsistent with the Second-order Stochastic Dominance criterion. Then, we show that the trade-off of the Omega ratio – between return and risk – is mostly guided by the first moment of portfolio returns. In Section 4, we provide numerous illustrations, based on realistic simulations and several asset classes, showing incoherencies in fund analyses when considering various (static and dynamic) optimization settings. Section 5 presents some robustness checks based on different long-only asset and hedge fund databases. The last section concludes.6

Section snippets

Return density-based performance and the Omega measures

The Omega measure belongs to a general class of performance measures based on features of the analyzed return density. Following the taxonomy proposed by Caporin et al. (2014), these measures can be represented with the following general notations: PMp=Pp+r×Ppr1,where PMp is a performance measure of portfolio p, Pp+. and Pp. are two functions associated with the right and left parts of the support of the density of returns.

In most cases, measures belonging to this class can be re-defined as

Inconsistency of the Omega when ranking funds

Let us first grasp the intuition and consider the daily returns of fund A as illustrated in Fig. 1, where (Panel A) represents the Probability Density Functions (PDF in short) of returns on fund A and fund B in the Gaussian paradigm (for the sake of simplicity in this example). Both funds have exactly the same average daily return, but the return distribution of fund B has twice the volatility of that of fund A. Fig. 1 (Panel B) also presents the Cumulative Distribution Function (CDF), the

Omega as an optimization criterion?

We studied the inconsistency of the Omega ratio (Cf. Proposition 1) through simple illustrations based on realistic simulations comparing two investment portfolios. We have shown that using Omega as a ranking criterion may be misleading in some cases since it is inconsistent with the SSD criterion. We would like now to study the behavior of Omega in a static and dynamic optimization framework.

We also ran some robustness checks, comparing the performance of portfolios optimized according to the

Conclusion

We recently observed that an increasing number of significant finance articles refer to the Omega measure for evaluating the performance of funds or of active strategies (e.g. Eling and Schuhmacher (2007), Farinelli and Tibiletti (2008), Annaert et al. (2009), Bertrand and Prigent (2011), Zieling et al. (2014), Kapsos et al. (2014), Hamidi et al. (2014)), since both the return distributions may not be Gaussian and the volatility may not be the relevant risk measure in such a case. The Omega

Acknowledgments

We thank Philippe Bertrand, Alexis Direr, Georges Hübner, Patrick Kouontchou, Yannick Malevergne, Constantin Mellios, Patrice Poncet, Jean-Luc Prigent, Thierry Roncalli and Jan Voelzke for help and suggestions when preparing this article. We are also grateful to the participants in the IFC7 (Paris, March 2013), the XVth INFER Conference (Orléans, May 2013), the XXXth AFFI Conference (Lyon, May 2013), the XXXth JMA (Nice, June 2013) and the XVIth INFER Conference (Pescara, May 2014) for valuable

References (104)

  • FarinelliS. et al.

    Sharpe thinking in asset ranking with one-sided measures

    European J. Oper. Res.

    (2008)
  • GetmanskyM. et al.

    An econometric model of serial correlation and illiquidity in hedge fund returns

    J. Financ. Econ.

    (2004)
  • GilliM. et al.

    Distributed optimisation of a portfolio’s omega

    Parallel Comput.

    (2010)
  • GlawischnigM. et al.

    Assessing the performance of alternative investments using non-parametric efficiency measurement approaches: Is it convincing?

    J. Bank. Financ.

    (2010)
  • HadarJ. et al.

    Stochastic dominance and diversification

    J. Econom. Theory

    (1971)
  • HamidiB. et al.

    A dynamic autoregressive expectile for time-invariant portfolio protection strategies

    J. Econ. Dyn. Control

    (2014)
  • HommU. et al.

    Beyond the sharpe ratio: An application of the aumann-serrano index to performance measurement

    J Bank. Financ.

    (2012)
  • HommU. et al.

    An operational interpretation and existence of the Aumann-Serrano index of riskiness

    Econom. Lett.

    (2012)
  • JhaR. et al.

    Measuring performance in a dynamic world: Conditional mean-variance fundamentals

    J. Bank. Financ.

    (2009)
  • KapsosM. et al.

    Worst-case robust omega ratio

    European J. Oper. Res.

    (2014)
  • QuigginJ.

    A theory of anticipated utility

    J. Econ. Behav. Organ.

    (1982)
  • RothschildM. et al.

    Increasing risk: I. A definition

    J. Econom. Theory

    (1970)
  • UnserM.

    Lower partial moments as measures of perceived risk: An experimental study

    J. Econ.Psychol.

    (2000)
  • VoelzkeJ.

    Weakening the gain-loss ratio measure to make it stronger

    Finance Res. Lett.

    (2015)
  • AgarwalV. et al.

    Risks and portfolio decisions involving hedge funds

    Rev. Financ. Stud.

    (2004)
  • AitchisonJ. et al.

    The Lognormal Distribution with Special References to Its Uses in Economics

    (1963)
  • Avouyi-DoviS. et al.

    Optimal asset allocation with omega Function. Technical Paper

  • BelliniF. et al.

    Risk management with expectiles

    Eur. J. Financ.

    (2015)
  • BernardoA.E. et al.

    Gain, loss and asset pricing

    J. Political Econ.

    (2000)
  • BiaginiS. et al.

    The best gain-loss ratio is a poor performance measure

    SIAM J. Financial Math.

    (2013)
  • BillioM. et al.

    A New Generalized Utility-based N-moments Measure of Performance. Working Paper

  • BlakeC.R. et al.

    Morningstar ratings and mutual fund performance

    J. Financ. Quant. Anal.

    (2000)
  • BriecW. et al.

    Input, output and graph technical efficiency measures on non-convex FDH models with various scaling laws: An integrated approach based upon implicit enumeration algorithms

    TOP

    (2006)
  • BriecW. et al.

    Non-convex technologies and cost functions: Definitions, duality and non-parametric tests of convexity

    J. Econ.

    (2004)
  • CapocciD.P.

    The persistence in hedge fund performance: Extended Analysis

    Int. J. Financ. Econ.

    (2009)
  • CaporinM. et al.

    A survey on the four families of performance measures

    J. Econ. Surv.

    (2014)
  • ČernỳA.

    Generalised sharpe ratios and asset pricing in incomplete markets

    Rev. Financ.

    (2003)
  • CharnesA. et al.

    Programming with linear fractional functionals

    Naval Res. Logist. Q.

    (1962)
  • ChernyA. et al.

    New measures for performance evaluation

    Rev. Financ. Stud.

    (2009)
  • CochraneJ.H.

    Asset Pricing

    (2009)
  • CochraneJ.H. et al.

    Beyond arbitrage: Good-Deal asset price bounds in incomplete markets

    J. Political Econ.

    (2001)
  • CremersM. et al.

    Should benchmark indices have alpha? Revisiting performance evaluation

    Crit. Financ. Rev.

    (2013)
  • DarsinosT. et al.

    Generalizing universal performance measures

    Risk

    (2004)
  • Del GuercioD. et al.

    Star power: The effect of morningstar ratings on mutual fund flow

    J. Financ. Quanti. Anal.

    (2008)
  • DeMiguelV. et al.

    Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy?

    Rev. Financ. Stud.

    (2009)
  • DenuitM. et al.

    Risk measurement with equivalent utility principles

    Stat. Decis.

    (2006)
  • EbertU.

    Measures of downside risk

    Econ. Bull.

    (2005)
  • ElingM. et al.

    Sufficient conditions for expected utility to imply drawdown-based performance rankings

    J. Bank. Financ.

    (2011)
  • EngleR. et al.

    Modelling the persistence of conditional variances

    Econometric Rev.

    (1986)
  • FishburnP.

    Mean-risk analysis with risk associated with below-target returns

    Amer.Econ. Rev.

    (1977)
  • Cited by (0)

    View full text