Performance Measurement Under Uncertainty
Every performance ratio in quantitative finance is a single number. The Sharpe ratio, the Sortino ratio, the Calmar ratio, the Omega ratio, the Rachev ratio: each takes a stream of returns and compresses it into one scalar. This post asks what that compression discards and what it preserves.
The answer requires measure theory. A return stream is a realisation of a random variable on a probability space. A performance ratio is a functional on the space of return distributions. Different ratios correspond to different functionals, and the relationships between them (when they agree, when they diverge, what pathologies each one hides) become visible only when you write them in a common mathematical language.
Section 0 establishes the epistemic status of the probability model itself. Section 1 builds the measure-theoretic machinery: the probability space, moments as Lebesgue integrals, partial moments, tail functionals, and the Johnson-Lindenstrauss dimensionality argument that motivates the full taxonomy. Section 2 defines the general performance functional and derives the Omega ratio as a canonical representative. Section 3 instantiates each classical ratio as a special case, proves that the distributional ratios agree within a location-scale family and derives how each ratio scales under aggregation. Section 4 tests these results on backtest data. Section 5 closes with a geometric perspective on the space of performance functionals.
0. Philosophical Prior
The market is a deterministic system. A large number of agents, each with private information, capital constraints, and objectives, submit, cancel, and execute orders according to deterministic decision rules. The resulting price at any moment is the output of this high-dimensional deterministic dynamical system, projected onto a one-dimensional observable: the mid-price.
The apparent randomness of returns is a consequence of this projection. When a portfolio manager computes the daily return of a strategy, the single number she records is the image of an enormously complex deterministic state under a function that discards almost all of the information in that state. The discarded information (the full order book, every agent's private signal, every latent intention to trade) is what makes the retained information look stochastic.
This observation has a precise mathematical structure. Let denote the full state of the market at time , where is enormous (on the order of the number of agents times the dimension of each agent's internal state). The return over one period is the output of a deterministic function applied to two successive states:
We do not observe . We observe only the sequence . Because is a radical dimensionality reduction (from to ), the sequence inherits statistical regularities that are well-modelled by a stochastic process, even though the underlying system is deterministic.
We therefore construct a probability space and model returns as random variables on it. This is a modelling choice. The probability measure encodes our ignorance of the full state.
Every performance ratio inherits this epistemic status. The Sharpe ratio of a strategy is a property of the model , not of the market. Two traders who use different probability models for the same strategy will compute different Sharpe ratios. Both are "correct" relative to their respective models. Which model is closer to the data-generating process at the resolution of interest is an empirical question.
The projection discards information. How much? The Johnson-Lindenstrauss lemma (Section 1.4) gives a rigorous lower bound on how much geometric structure remains after dimensionality reduction. We state it there. Compressing a distribution to a single scalar is far more aggressive than any random projection that preserves pairwise distances. A single ratio therefore cannot capture the full shape of a return distribution. Each ratio captures one aspect of the shape, and the rest of this post makes the aspects precise.
1. Measure-Theoretic Foundations
This section constructs the mathematical objects that every performance ratio is built from. The treatment is self-contained: the only prerequisite is familiarity with Lebesgue integration at the level of a first graduate course in real analysis.
1.1. Probability space
We work on a probability space , where is the sample space of all possible market histories, is a -algebra of events, and is a probability measure. The choice of is the modelling decision discussed in Section 0: it encodes our beliefs about which market histories are likely, given incomplete knowledge of the full state.
A return random variable is a measurable function , where is the Borel -algebra on . The cumulative distribution function (CDF) of is
and the survival function is
The CDF is right-continuous, non-decreasing, with and . It completely characterises the distribution of under . The pushforward measure on is the law of , and integration against is equivalent to integration against via the change-of-variables formula:
for any measurable for which the integral exists. This identity is the bridge between the abstract probability space and the observed distribution of returns. Every formula in this post reduces to an integral against .
1.2. Moments as Lebesgue integrals
The moments of are the building blocks of the simplest performance ratios. We define them as Lebesgue integrals to make the existence conditions explicit.
The -th raw moment of is
provided the integral exists (i.e., ). The first raw moment is the mean return. The -th central moment is
The familiar summary statistics of a return distribution are low-order central moments. The variance is , measuring dispersion around the mean. The standard deviation has the same units as . The skewness is the normalised third central moment,
which is positive when the right tail is heavier than the left and negative when the left tail dominates. The excess kurtosis is
where the subtraction of 3 normalises so that a Gaussian distribution has . Positive excess kurtosis indicates heavier tails than Gaussian; negative indicates lighter tails.
The existence of the -th moment requires , that is . Empirical return distributions for most asset classes exhibit power-law tails, with a tail index between 2 and 5 for most data sets studied.[1] For the fourth moment (and hence excess kurtosis) does not exist. Any performance ratio that depends on the fourth moment of returns is potentially ill-defined for fat-tailed distributions. The point recurs in Section 3.
1.3. Partial moments and tail functionals
The Sharpe ratio uses the full variance , which penalises upside and downside dispersion equally. Most practitioners care more about downside risk than upside volatility. Partial moments and tail functionals formalise this asymmetry.
For a threshold and order , the lower partial moment of is
and the upper partial moment is
The lower partial moment of order 0 is , the probability of falling below the threshold. Order 1 gives the expected shortfall below . Order 2, , is the downside semivariance, and its square root is the downside deviation that appears in the Sortino ratio. The threshold is typically set to zero (for absolute returns) or to the risk-free rate (for excess returns).
The tail functionals Value-at-Risk and Conditional Value-at-Risk provide complementary characterisations of downside risk.
For , the Value-at-Risk at confidence level is the left quantile of :
The Conditional Value-at-Risk (also called Expected Shortfall) at level is
answers the question: what is the worst return that occurs with probability at least ? For , it is the 5th percentile of the return distribution. answers a different question: given that we are in the worst -fraction of outcomes, what is the expected return? It is always at least as severe as , and unlike VaR, it is a coherent risk measure in the sense of Artzner, Delbaen, Eber and Heath: it satisfies subadditivity, monotonicity, positive homogeneity, and translation invariance.[12]
The partial moments, VaR, and CVaR are the building blocks from which every ratio in the taxonomy of Section 3 is assembled. The Sharpe ratio uses only and . The Sortino ratio replaces with . The Calmar ratio uses the maximum drawdown, a path functional (we define it in Section 3). The Omega ratio uses the entire CDF. The Rachev ratio uses CVaR in both tails. The progression from simple to complex reflects an increasing commitment to using more of the information in .
1.4. Johnson-Lindenstrauss and the dimensionality argument
The philosophical prior of Section 0 framed performance measurement as dimensionality reduction: from the full market state , through the return vector , to a single scalar ratio . The Johnson-Lindenstrauss lemma is the sharpest known result on how much geometric structure random projections can preserve.
For any and any set of points in , there exists a linear map with
such that for all ,
A random projection onto a -dimensional subspace (with entries drawn i.i.d. from ) satisfies this condition with high probability.[9]
Proof. Let be a random matrix with entries drawn i.i.d. from , and define . Fix any pair and let . The projected squared norm is
where i.i.d. Each inner sum is a linear combination of independent Gaussians with variance , so . Therefore
and . We need to be small. The moment-generating function of a variable is for , so by independence,
Applying the Chernoff bound to the upper tail: for any ,
Setting (which lies in for ) and simplifying the exponent yields
The exponent of the base is . By the alternating series remainder for the Mercator series, , and for we have , so . Raising to the power :
The same argument for the lower tail (optimising in the Chernoff bound) gives . Combining both tails via a union bound:
There are pairs in . A union bound over all pairs gives failure probability at most . Setting this less than and solving for :
The theorem says that points in arbitrarily high dimension can be embedded into dimensions while preserving all pairwise distances to within a factor of . The target dimension depends only logarithmically on and is independent of the ambient dimension .
Now consider the performance measurement pipeline. The market state lives in with enormous. A return vector over periods is the output of a map . A performance ratio then compresses further: .
JL guarantees that a random linear projection to dimensions preserves pairwise distances. The non-asymptotic bound from the Chernoff proof above is explicit:
For trading days, , and failure probability : . A single performance ratio has , far below this threshold.
The market-to-returns map is a structured, nonlinear function determined by the microstructure of order matching and price formation. JL does not directly apply to this map. JL states that dimensions suffice for random linear projections. Larsen and Nelson (2017)[13] proved the bound optimal: some -point sets need dimensions under any map that preserves their pairwise distances to within . Nothing guarantees that the structured map does as well as a random projection.
A single scalar ratio is a far more aggressive compression than any distance-preserving projection allows. Two return distributions that look identical under the Sharpe ratio can differ in their tail behaviour, skewness, or drawdown profile. Hence the full functional taxonomy (Sharpe, Sortino, Calmar, Omega, Rachev, and the general performance functional of Section 2) is necessary. Each ratio is a different one-dimensional projection of the return distribution, and together they recover more of the shape of than any single ratio can.
2. General Performance Functional
Every ratio in the taxonomy shares a common structure: a numerator that measures location (how much return?) divided by a denominator that measures dispersion (how much risk?). This section formalises that structure, states the axioms that the numerator and denominator should satisfy, and then proves that the Omega ratio is the canonical representative of the entire family, in a precise sense: knowledge of for all is equivalent to knowledge of the full distribution .
2.1. Definition and axioms
Let denote the space of probability distributions on with finite second moment. Every return random variable with has its law in .
A performance functional is a map that admits a decomposition
where is a location functional and is a dispersion functional.
We impose axioms on each component separately. The location functional should behave like a generalised mean: it should respond to shifts and respect stochastic dominance.
A location functional is admissible if it satisfies:
- Translation-equivariance. For any constant ,
- Monotonicity. If almost surely, then .
Translation-equivariance says that adding a deterministic return to every outcome shifts the location by exactly . The mean satisfies this, as does any quantile. Monotonicity says that a pointwise-dominant return stream is never assigned a lower location. Together, these two axioms rule out pathological location measures (such as the mode, which can jump discontinuously under small perturbations of ) while remaining permissive enough to admit , , medians, and partial-moment-based location measures.
The dispersion functional should measure the scale of the distribution without responding to shifts.
A dispersion functional is admissible if it satisfies:
- Positive homogeneity. For any ,
- Shift-invariance. For any constant ,
Positive homogeneity says that scaling returns by a positive factor scales dispersion by the same factor. The standard deviation satisfies this (since ), as does and (after centring). Shift-invariance says that adding a constant to all returns does not change the spread. Together, these axioms mean that captures the shape and scale of the distribution and ignores its level.
The pair specifies a performance functional . Different choices of and yield different ratios. Before enumerating the taxonomy in Section 3, we note the connection to the axiomatic framework of Cherny and Madan (2009).[3] They define an acceptability index as a functional satisfying quasi-concavity (diversification does not decrease performance), monotonicity (higher returns yield higher performance), scale-invariance ( for ) and a continuity condition (the Fatou property). Our axiom set is deliberately weaker: we require the decomposition into location and dispersion, and impose axioms on each component, but we do not require quasi-concavity or scale-invariance of the composite at this stage. Several widely used ratios (the Calmar ratio among them) fail quasi-concavity or scale-invariance, and the framework includes them. Section 3 identifies which classical ratios satisfy the stronger Cherny-Madan conditions and which do not.
2.2. Omega master functional
The Omega ratio, introduced by Keating and Shadwick (2002),[4] is the master functional of the taxonomy. Every other ratio extracts partial information from the distribution . The Omega ratio, viewed as a function of its threshold parameter, encodes the entire distribution. Its numerator and denominator are the expected payoffs of a call and a put struck at , so Omega is a call-to-put price ratio under .[7] At it is the gain-loss ratio, whose market-wide maximum Biagini and Pinar show is either infinite or not attained in most continuous-time models.[8]
For a threshold with , the Omega ratio is
The equivalence of the three expressions follows from integration by parts. The numerator equals because
where the exchange of integral and expectation is justified by Fubini's theorem (the integrand is non-negative). An analogous computation gives the denominator.
The function for uniquely determines . Conversely, every ratio in the taxonomy except the Calmar ratio is a functional of (for the Information ratio, the Omega function of ).
Proof. We prove both directions.
determines . Define the denominator and numerator functions
By the fundamental theorem of calculus, . To relate and , observe that identically (check both cases and ). Taking expectations:
Substituting into :
Rearranging for wherever :
Differentiating via the quotient rule with and recovers the CDF:
This expresses entirely in terms of , , and . The mean is itself determined by : since and , the unique zero of is . Therefore uniquely determines .
Recovery of each ratio from . Each distributional ratio is a functional of , which determines. Several have direct expressions. Differentiating [eq:omega-rewritten] at the mean gives . The Sharpe ratio is . The downside semivariance in the Sortino ratio is . The Rachev ratio uses the numerator and denominator functions at two quantiles. The Calmar ratio is the only ratio that cannot be recovered from alone, because it depends on path ordering. The explicit formulas and full derivations are given in the recovery lemmas of Section 3.
2.3. Functional taxonomy table
The decomposition gives a uniform description of the classical ratios. The following table collects the location and dispersion functionals for each ratio in the taxonomy, using the notation established in Section 1.
| Ratio | Location | Dispersion |
|---|---|---|
| Sharpe | ||
| Sortino | ||
| Information | ||
| Calmar | (path functional) | |
| Omega | ||
| Rachev |
Here is the risk-free rate, is the benchmark mean return, is the benchmark return, is the annualised mean return, is the maximum drawdown, and denotes the conditional tail expectation in the right () or left () tail at level .
Moving down the rows the location functional shifts from a simple mean excess return (Sharpe, Sortino, Information) to a partial moment (Omega) to a conditional tail expectation (Rachev). The dispersion functional moves from the full standard deviation (Sharpe) to downside-only measures (Sortino) to partial moments (Omega) to tail measures (Rachev). The Calmar ratio is the outlier: its dispersion functional, the maximum drawdown, depends on the path. The next proposition states this precisely.
The Calmar ratio is the only ratio in the taxonomy that is not a distributional functional. Specifically, the maximum drawdown depends on the temporal ordering of returns: there exist return sequences and (where is a permutation of ) that have identical empirical distributions but different maximum drawdowns.
Proof. We construct an explicit counterexample. Consider the return sequence
The cumulative wealth process, starting from , is
The running maximum is , so . The drawdown at each time is , and the maximum drawdown is
Now consider the permuted sequence
This is a permutation of the same multiset of returns, so the empirical distribution is identical for and . The cumulative wealth process for is
The running maximum is , and the maximum drawdown is
We have but . Therefore MDD is not a functional of the return distribution alone.
For the remaining five ratios (Sharpe, Sortino, Information, Omega, Rachev), each is defined entirely in terms of expectations, partial moments, quantiles, or conditional tail expectations, all of which are functionals of and are invariant under permutations of the return sequence. The Calmar ratio is therefore the unique path-dependent ratio in the taxonomy.
Maximum drawdown captures information about the serial structure of returns (clustering of losses, momentum and mean-reversion effects) that no distributional functional sees. Hence Calmar cannot be computed from alone and cannot be recovered from the Omega function . It penalises the risk that practitioners find most painful: sustained compounded losses from peak equity.
3. Ratio Taxonomy
With the general performance functional and the Omega master functional in hand, we instantiate each classical ratio as a special case, prove the recovery lemmas promised in Section 2, and then prove two results that relate the ratios to one another: their agreement within a location-scale family (Section 3.7) and their scaling under aggregation (Section 3.8).
3.1. Sharpe ratio
The Sharpe ratio, introduced by William Sharpe in 1966,[5] is the oldest and most widely used performance ratio. It measures the excess return per unit of total risk:
For a return random variable with mean and standard deviation , and a risk-free rate , the Sharpe ratio is
The location functional is and the dispersion functional is . Both are distributional functionals: they depend on alone and are invariant under permutation of the return sequence.
The Sharpe ratio can be recovered from the Omega function as
Proof. We start from the definition of at threshold :
The numerator and denominator decompose the mean excess return. We have the identity
This follows by writing and taking expectations. Denoting and , we obtain
From the first equation, . Substituting into the second:
Therefore
This expresses the Sharpe ratio entirely in terms of the Omega function evaluated at , the first lower partial moment at and the standard deviation. Both and are themselves recoverable from (since determines by Theorem (Omega sufficiency), and both functionals are defined in terms of ). Therefore is fully determined by .
Let belong to one location-scale family, for a fixed with mean 0 and variance 1. Elliptical families are the main case. Suppose each strategy can be combined with the risk-free asset at any weight. Then if and only if at every target volatility the scaled has higher expected utility than the scaled for every strictly increasing utility . Outside a common location-scale family this consistency can fail.
Proof. An elliptical distribution has the form , where has a spherically symmetric distribution (i.e., the density of depends only on ). The class includes the Gaussian, Student-, and Laplace families. The argument below uses only the common generator .
Common family. Let be strictly increasing and consider two strategies with returns and . Since both returns share the same generating variable , we can write
where we used , so . Consider the function . For fixed , we compute
Since is increasing, , so : the expected utility is strictly increasing in the Sharpe ratio for any fixed volatility level.
For any target volatility the investor can place a fraction in strategy and the rest in the risk-free asset. The return is with Sharpe ratio and volatility . Under this homogeneous scaling, the expected utility of the leveraged strategy at target volatility depends only on (since both strategies now have the same volatility and the same distributional shape, differing only in the location shift ). Since is increasing in , the strategy with higher Sharpe ratio yields higher expected utility at every target volatility. Equal Sharpe ratios give equal expected utility, which proves the converse.
Counterexample outside a common family. We construct an explicit counterexample using the skew-normal distribution. The skew-normal family has CDF
where is the standard normal CDF and is Owen's T-function. The parameter controls skewness: gives a Gaussian, gives positive skew, gives negative skew.
The construction requires two strategies where the higher-Sharpe strategy has heavy negative skew. A sufficiently risk-averse utility function then penalises its left tail enough to reverse the Sharpe ranking. Let and .
For : , the mean is , and the variance is . Hence and .
For : , , . Hence and .
So . The skewness of is (strongly negative), while (strongly positive).
For the exponential utility with , the expected utility of a skew-normal variable involves the moment-generating function , with . For the skew-normal, where .
For : , so .
For : , so . Since , the product is .
Therefore and . Hence . Strategy has lower Sharpe ratio but higher expected utility under this risk-averse exponential utility: the heavy negative skew of dominates its mean-variance advantage. The reversal persists at a common volatility. Scaling by gives with , so its expected utility is about .
Outside a common location-scale family there exist strategies and an increasing, concave utility function such that but , also after scaling both to a common volatility.
Proof. The skew-normal pair above is such a pair. The mechanism: outside a common location-scale family, the first two moments do not determine the distribution. A higher Sharpe ratio guarantees a better mean-variance profile and says nothing about higher moments (skewness, kurtosis, tail behaviour). An expected-utility maximiser whose utility function is sensitive to these higher moments (as any sufficiently risk-averse utility function must be, by the Taylor expansion ) can prefer a lower-Sharpe strategy that has more favourable skewness or kurtosis.
The third-order term in the utility expansion, , is positive for (which holds for DARA utilities). A prudent investor therefore favours the more right-skewed of two strategies with similar mean and variance. The skew-normal example above gives explicit parameters.
3.2. Sortino ratio
The Sortino ratio[2] replaces the standard deviation in the Sharpe denominator with the downside deviation. It penalises only below-threshold volatility:
For a return random variable with mean , risk-free rate , and downside deviation , the Sortino ratio is
The Sortino ratio is recoverable from the Omega function .
Proof. By the Omega sufficiency theorem (Theorem), the function uniquely determines . Given , we compute:
The mean: .
The second lower partial moment at :
Both integrals are functionals of , which is determined by . From the definition of , the denominator function is . Starting from the defining integral and integrating by parts (with , ):
where the boundary term vanishes since at the upper limit and as (because faster than grows, given ). Integrating by parts once more with (the boundary terms vanish because as ) gives
By [eq:D-recovery] the function is recoverable from for , and is the unique zero of .
Therefore is a functional of , and the Sortino ratio is fully determined by .
The Sortino and Sharpe ratios are related by
For symmetric distributions, when , and more generally and induce the same ordering when the downside multiplier is the same across all strategies being compared. When the downside multipliers differ (which occurs whenever the left-tail shape varies across strategies, even at the same skewness), the Sortino and Sharpe orderings can disagree.
Proof. The algebraic relationship follows immediately from the definitions:
The ratio is the downside multiplier. For symmetric distributions with , the downside semivariance equals half the total variance: . By symmetry of about , this integral equals , and the two integrals sum to , so . The downside multiplier is then , giving .
When comparing two strategies with the same downside multiplier (and the same threshold ), we have , and the ratio is a common positive rescaling of . Therefore if and only if : the orderings agree. Equal skewness is not sufficient for equal downside multipliers, because the multiplier depends on the full shape of the left tail (kurtosis and higher moments also contribute to ).
When skewness differs, the downside multipliers can differ, and disagreement is possible. Consider strategy with and negative skewness (heavy left tail, so is large relative to , giving a small downside multiplier). Consider strategy with and positive skewness (light left tail, so is small relative to , giving a large downside multiplier).
As an example, let have and (so the downside multiplier is ). Hence .
Let have and (so the downside multiplier is ). Hence .
We have but . The Sharpe and Sortino orderings disagree. Strategy has lower total risk-adjusted return but much less downside risk (positive skewness concentrates volatility on the upside), so the Sortino ratio, which penalises only downside volatility, ranks it higher.
3.3. Information ratio
The Information ratio measures the excess return of a strategy relative to a benchmark, scaled by the volatility of the excess return (the tracking error):
For return random variables (strategy) and (benchmark) with , the Information ratio is
The Information ratio is the Sharpe ratio of the excess return with zero risk-free rate. That is, if we define , then with .
Proof. Define the excess return . Then and , so . The Sharpe ratio of with is
This is an algebraic identity. The Information ratio inherits all properties of the Sharpe ratio when applied to excess returns.
With a deterministic benchmark ( almost surely) the Information ratio is the Sharpe ratio with . It is then scale-invariant and quasi-concave where it is non-negative, and like the Sharpe ratio it fails monotonicity. When is stochastic, the IR can violate quasi-concavity.
Proof. When is a constant, the excess return is , which is a translation of . The Information ratio becomes , which is the Sharpe ratio with . For scale invariance: for . For quasi-concavity: the mean is linear and the standard deviation is subadditive. Hence for a mixture with non-negative excess means . Monotonicity fails. With , let take the values 0 and 1 with probabilities 0.1 and 0.9, and let equal except that it pays 10 on an event of probability 0.1 where . Then but .
When is stochastic, quasi-concavity can fail. With a stochastic benchmark the tracking error of a mixture need not be bounded by the tracking errors of the components, because the benchmark can interact non-linearly with the portfolio weights.
Consider a non-linear benchmark dependence. Let be two strategy returns, and define the benchmark . For the equal-weight mixture :
When , this equals . When , this equals . So .
Numerical confirmation. Let and be independent (so ). Since , the benchmark mean is .
For each component: . Monte Carlo evaluation ( draws) gives tracking error , so .
For the equal-weight mixture : the excess return (as derived above). The numerator is the same: . But the tracking error drops to , because the operation clips the upper tail and reduces variance. This worsens the ratio: .
The quasi-concavity violation is . The mechanism: the operation reduces the denominator (tracking error) and preserves the negative numerator. The ratio falls further below zero.
3.4. Calmar ratio
The Calmar ratio measures annualised return per unit of maximum drawdown:
For a return process with annualised mean return and maximum drawdown , the Calmar ratio is
where
The path-dependence of MDD (and hence of Calmar) was proved in Proposition (Calmar path-dependence) in Section 2.3. The next result characterises how MDD scales with the time horizon.
For a standard Brownian motion on , the expected maximum drawdown is
With drift and volatility the growth is logarithmic, as . Magdon-Ismail and Atiya (2004)[6] give the full drift-dependent formula.
Proof. For a standard Brownian motion , the drawdown at time is , where is the running maximum. The maximum drawdown over is .
By Levy's identity, the drawdown process has the same law as (reflected Brownian motion). Therefore , and Brownian scaling gives .
Let be the exit time of from . By scaling, if and only if the exit time from exceeds 1. That exit time has the law of . Hence . The Laplace transform of the exit time is , and . Therefore
with the substitution . So . The drift case needs the full drawdown distribution and is taken from Magdon-Ismail and Atiya.
3.5. Omega ratio
The Omega ratio was defined in Definition (Omega ratio) and its sufficiency was proved in Theorem (Omega sufficiency). Two further properties follow.
if and only if .
Proof. From the definition, . The identity gives, after taking expectations,
Therefore
Since for all with (which holds whenever is above the essential infimum of ), the sign of equals the sign of .
If : , so .
If and : then , so .
It remains to check the boundary cases. If is at or below the essential infimum of , then (there is no probability mass below ), and is undefined (or formally ), so . If is at or above the essential supremum, then and . In the interior of the support the result follows as shown.
If first-order stochastically dominates (written ), then for all .
Proof. First-order stochastic dominance means for all , which is equivalent to for all .
The numerator of is
since pointwise and we integrate over the same domain.
The denominator of is
since pointwise and we integrate over the same domain.
Therefore, with and (the strict positivity conditions hold whenever is in the interior of both supports):
The inequality when and follows from (the first step uses with ; the second uses with ).
No other ratio in the taxonomy (Sharpe, Sortino, Information, Calmar, Rachev) satisfies FSD consistency unconditionally.
Proof. We provide an explicit counterexample for each ratio.
Sharpe. We need but . The mechanism is FSD-domination through support widening: the dominant distribution has a higher mean but disproportionately higher variance.
Let be uniform on and be uniform on . FSD holds: for and for . So . The Sharpe ratios (with ) are and . So while . The dominating distribution has 334 times the mean but 1000 times the standard deviation. Hence the Sharpe ratio decreases.
Information. With a deterministic benchmark the Information ratio is the Sharpe ratio with . The Sharpe counterexample applies with .
Sortino. The Sortino ratio is FSD-consistent whenever the dominated strategy has : FSD raises the mean and lowers , because is decreasing in . It fails when the means lie below the threshold. With , let equal with probability 0.9 and with probability 0.1, and let almost surely. Then , but while .
Calmar. The Calmar ratio depends on path ordering and trivially fails distributional FSD consistency: two path-rearrangements of the same return sequence have identical distributions (hence identical FSD relationships to any other distribution) but different Calmar ratios, as shown in Proposition (Calmar path-dependence).
Rachev. FSD orders the quantile functions, . It therefore raises the right-tail mean and lowers the left-tail loss. When both are positive the Rachev ratio rises, so a violation needs a negative right-tail mean. Let and almost surely, so . At the right-tail mean of is and its left-tail mean is , so . For the constant , .
Hence the Omega ratio is the unique ratio in the taxonomy that respects first-order stochastic dominance at every threshold and for every sign of its inputs. Its numerator and denominator are integrals of the complementary and ordinary distribution functions, which FSD constrains directly.
3.6. Rachev ratio
The Rachev ratio of Biglova, Ortobelli, Rachev and Stoyanov[14] compares the expected gain in the right tail to the expected loss in the left tail:
For confidence levels , the Rachev ratio is
The Rachev ratio directly compares tail behaviour: the numerator is the expected return conditional on being in the top -fraction of outcomes, and the denominator is the magnitude of the expected return conditional on being in the bottom -fraction. Typical values are .
The Rachev ratio is recoverable from the Omega function .
Proof. By the Omega sufficiency theorem, determines . Given , the quantile function is determined, and the conditional expectations
are functionals of , hence of , hence of .
For a more direct connection, write and take continuous at . Then .
The left-tail CVaR (the expected loss, positive when the left tail lies below zero) is
Since (the denominator of at ), and is determined by which is determined by , the left-tail CVaR is a functional of .
An analogous argument shows that the right-tail CVaR is determined by (the numerator of at the -quantile):
where is the numerator of .
Therefore
which expresses the Rachev ratio entirely in terms of the numerator and denominator functions of evaluated at specific quantiles (themselves determined by ).
The Rachev ratio is the only ratio in the taxonomy (besides Omega) that captures tail asymmetry without requiring the existence of finite higher moments (skewness, kurtosis). Specifically, for distributions with infinite variance in both tails (e.g., symmetric stable distributions with index ), the Sharpe, Sortino, and Information ratios are undefined, while the Rachev ratio remains well-defined and finite whenever the tail conditional expectations exist.
Proof. The Sharpe ratio requires , which fails when the variance is infinite. The Sortino ratio requires , which also fails for distributions with infinite second moment (since , but this gives only an upper bound, and for symmetric heavy-tailed distributions like stable laws with , the second moment is infinite on both sides, so is infinite as well). The Information ratio inherits the same limitation since it is the Sharpe ratio of excess returns.
The Rachev ratio, by contrast, requires only the finiteness of and for appropriate quantiles. For a stable distribution with index , the mean exists (since ), and the conditional expectations and exist and are finite for any quantiles in the interior of the support (because conditioning on being above/below a fixed quantile truncates the distribution, and the resulting truncated moments are finite for ).
To verify: for a stable random variable with index and scale , the tail probability satisfies as . The conditional expectation, given , is
The integrand as (since ). This integral converges if and only if , which is our assumption. The same argument applies to the left tail.
The Calmar ratio requires only the mean return (for ) and the maximum drawdown (a path functional), so it can be defined for . The Omega ratio at a fixed threshold requires and , both of which are first moments of truncated distributions, hence finite for .
Therefore, among the moment-based ratios (Sharpe, Sortino, Information), none can handle infinite-variance distributions. Among the remaining ratios, both Omega and Rachev handle heavy tails, but Rachev is specifically designed to measure tail asymmetry (right tail gain vs. left tail loss), whereas Omega at a fixed threshold measures the ratio of total upside to total downside over the entire distribution. The Rachev ratio isolates the tails by conditioning on extreme quantiles.
3.7. Performance ratio divergence
When do two distributional ratios rank strategies the same way? Within one location-scale family they always do. Across families of different shape they can disagree, even at zero skewness.
Fix a threshold and a random variable with and . Let each strategy in a family have returns with , so that its Sharpe ratio is . Then on this family
where , , is the mean of over its top -fraction and is minus its mean over the bottom -fraction. Sortino is a strictly increasing function of wherever , Omega wherever and Rachev on . Sharpe, Sortino, Omega at and Rachev of the excess return therefore induce the same ordering on the family.
Proof. Write . Then and . The Sortino formula follows from [eq:sortino-def] and the Omega formula from [eq:omega-rewritten]. The quantiles of are , so the right-tail mean is , the left-tail loss is and cancels in the Rachev ratio.
Differentiating under the integral gives and .
Sortino. The derivative of is . Put and , so that . If then and . If then , so . Equality in both steps forces almost surely, that is almost surely, which excludes.
Omega. The derivative of is
For the integrand exceeds on the event , which has positive probability because . For the condition gives , and this is positive because means .
Rachev. The derivative of is . The top-tail mean of a non-constant mean-zero variable is positive and its bottom-tail mean is negative, so .
The theorem extends Theorem (Sharpe completeness). On a common location-scale family the distributional ratios contain no ranking information beyond the Sharpe ratio. The functions and depend on the shape of . Two strategies of different shape can therefore be ranked differently.
There exist two strategies with zero skewness whose Sharpe and Sortino orderings disagree.
Proof. Take , a Gaussian strategy with and a Laplace strategy with . Both are symmetric. Evaluating [eq:ls-ratios] by quadrature for each generator gives and . Sharpe ranks first and Sortino ranks first.
For small Sharpe ratios the comparison reduces to two shape constants. Expanding gives
To leading order a pair with is reordered by Sortino when and by Omega when . For the Gram-Charlier density
with the standard normal density, the skewness and the excess kurtosis, the half-line moments of give
The leading Sortino comparison depends on skewness and not on kurtosis. The leading Omega comparison depends on kurtosis and not on skewness. Negative skewness raises and lowers the Sortino ratio at a given Sharpe ratio. Excess kurtosis lowers and raises the Omega ratio at a given Sharpe ratio. These formulas describe shapes near the Gaussian, where the Gram-Charlier density is non-negative for small and .
The return distribution explorer draws a Gram-Charlier density for a chosen skewness and kurtosis and recomputes each ratio for it. Moving the skewness slider changes the Sortino ratio at fixed mean and variance; moving the kurtosis slider changes the Omega ratio.
3.8. Scaling exponents under coarse-graining
How does each ratio change when the same returns are observed at a coarser resolution? For i.i.d. returns the Sharpe ratio scales exactly as . The other ratios have no fixed exponent. Their local exponents start at and drift with , and the Calmar ratio obeys a pathwise bound with exponent 1.
Let be i.i.d. returns at resolution with mean , variance and Sharpe ratio . Let be the sum of consecutive returns, with threshold , and write . For a positive ratio define the local exponent .
-
Sharpe and Information. exactly, and likewise for the Information ratio when the excess returns are i.i.d. The exponent is .
-
Gaussian returns. With , and ,
the Rachev ratio taken on the excess return . Their local exponents, evaluated at , are
Each tends to as , exceeds when and falls below it when .
- Calmar, any return sequence. Let the horizon be periods of log returns, and compute the mean and the maximum drawdown on the coarse series of block sums. Then the coarse mean is times the fine mean and . If the mean is positive, the per-period Calmar ratio satisfies , and the annualised Calmar ratio satisfies .
Proof. Sharpe and Information. and, by independence, , so . The Information ratio is the Sharpe ratio of the excess return (Lemma), and the same computation applies.
Gaussian returns. with standard normal. The coarse-grained returns therefore form the Gaussian location-scale family with Sharpe ratio , and [eq:ls-ratios] applies with , and . The closed forms of and follow from and . Since , and , the chain rule gives the three local exponents. The expansions use , and . All three numerators are positive for and negative for .
Calmar. The block sums add up to the fine sum, so their mean over blocks is times the fine mean. The coarse cumulative log-wealth path is the fine path sampled at times . The maximum drawdown is a maximum of over pairs , and restricting to a subset of pairs cannot increase a maximum. The annualised mean is the same at every resolution.
The script scripts/verifications/performance_ratios_theory.py checks the Gaussian formulas symbolically and against Monte Carlo sums.
The Calmar bound is tight in the Brownian limit. For a continuous path and a fixed horizon, the maximum drawdown on a grid converges to the continuous-time maximum drawdown as the mesh shrinks. The per-period Calmar ratio therefore has exponent 1 to leading order and the annualised Calmar ratio has exponent 0. Proposition (MDD scaling) describes a different limit, in which the horizon grows at a fixed resolution.
For non-Gaussian i.i.d. returns with a finite third moment, the Edgeworth expansion adds a term of order to the density of the standardised sum. The Gaussian formulas then hold up to a correction that vanishes as grows. For returns in the domain of a stable law with index , the scale of grows as and the drift as . A positive drift therefore pushes the left-tail mean of the excess sum above zero at a finite , where the Rachev ratio of the excess return diverges and changes sign, as it does in the Gaussian case at .
The exponents in the theorem separate into three groups. Sharpe and Information share the exponent exactly. Sortino, Omega minus one and Rachev minus one start at and move away from it at rates , and (at ) per unit of . Calmar computed with a per-period mean has exponent at least 1.
4. Empirical Comparison on Polybius Backtests
The preceding sections built the theory: a common functional language for performance ratios, the conditions under which the distributional ratios agree, and explicit predictions for how each ratio scales under aggregation. This section tests those predictions on data.
The data source is the proprietary Polybius backtest engine. We do not describe the strategy, the market or the execution mechanism. What matters for the mathematical analysis is the return series itself: its distributional shape, its serial structure and the behaviour of each ratio computed from it. Every number below is printed by scripts/generate_performance_ratios_figures.py.
4.1. Data and methodology
The backtest produces 2,355 fills over a two-week window (February 1 through 14, 2025). The equity series records one value after each fill and yields 2,354 log returns . We use log returns throughout this section because their aggregation in Section 4.4 is a sum, as in Theorem (Scaling under coarse-graining).
The backtest reinvests its equity on every fill. Equity grows from 9,750 USD to USD, a factor of about that reflects full reinvestment in a simulation with no capacity limit. The empirical moments of the return series are
Two consecutive fills dominate the higher moments. Fill 1,558 loses 99.8% of equity (log return , or ) and fill 1,559 multiplies equity by 1,643 (log return , or ). With those two fills removed the skewness is 5.67 and the excess kurtosis 106. Sample moments this sensitive to two observations describe the sample. They do not estimate a population skewness or kurtosis to any useful precision.
4.2. Ratio computation
We compute all six ratios from the 2,354 fill returns with threshold zero. For each ratio we also construct a 95% bootstrap confidence interval from 10,000 nonparametric resamples (sampling fills with replacement, recomputing the ratio for each resample and taking the 2.5th and 97.5th percentiles).
| Ratio | Value | 95% Bootstrap CI | Notes |
|---|---|---|---|
| Sharpe | 0.040 | [0.001, 0.089] | Per fill |
| Sortino | 0.068 | [0.002, 0.345] | Target = 0 |
| Information | 0.034 | [−0.006, 0.077] | Benchmark = median fill return (0.0016) |
| Calmar | 0.0094 | [0.0003, 0.022] | Mean fill return / MDD, MDD = 99.97% |
| Omega () | 1.43 | [1.01, 2.00] | Threshold at zero |
| Rachev (5%, 5%) | 1.84 | [1.04, 3.28] | Right tail heavier than left |
Four observations follow.
Sharpe-Sortino multiplier. The Sortino ratio (0.068) exceeds the Sharpe ratio (0.040) by a factor of 1.68. By Proposition (Sortino-Sharpe relationship) this factor is the downside multiplier . It implies that 35% of the second moment about zero lies below zero. A Gaussian with the same Sharpe ratio has multiplier . The larger observed value fits the positive skewness: the left tail contributes less to the second moment than a Gaussian left tail would.
Rachev above unity. The Rachev ratio at the 5% level is 1.84. The mean of the best 5% of fills exceeds the mean loss of the worst 5%, in line with the positive skewness. The interval [1.04, 3.28] is wide because each tail mean rests on 118 fills, a few of which are extreme.
Calmar suppression. With a maximum drawdown of 99.97% the Calmar ratio (0.0094) equals the mean fill return to two figures. A single episode sets it (Section 4.3).
Ordering conflict. Compare the backtest with a Gaussian strategy whose Sharpe ratio is 0.045. Sharpe prefers the Gaussian strategy (0.045 > 0.040). Its Sortino ratio is , below the backtest's 0.068, and Sortino prefers the backtest. The two strategies have different shapes, the condition under which Proposition (Disagreement across shapes) allows a conflict. The bootstrap intervals overlap heavily and establish neither ordering.
4.3. Equity curve and drawdown anatomy
The equity curve reveals the temporal structure that scalar ratios compress away.

The maximum drawdown runs from the equity peak at fill 1,544 to the trough at fill 1,558, fourteen fills that end with the single fill losing 99.8% of equity. The next fill restores most of it and the previous peak is regained at fill 1,565. Earlier drawdowns reach 58% by fill 120. The Calmar ratio is set entirely by the one episode at fill 1,558. A strategy can have a positive Sharpe ratio and still a negligible Calmar ratio if one sufficiently adverse sequence occurs.
The rolling Sharpe and Sortino ratios (computed over a trailing window of 100 fills) show how the downside multiplier moves through the sample.

For windows ending between fills 100 and 900 the median rolling Sharpe ratio is 0.17 and the median rolling Sortino ratio 1.09. The multiplier of 5.3 says that losses in these windows are small and few. For windows ending between fills 1,300 and 1,800, which contain the large losses, the median multiplier falls to 1.7 and both ratios sit near zero. They turn negative for windows ending near fill 1,555 and between fills 1,795 and 1,875. Over the whole sample the rolling multiplier ranges from 1.5 to 12 (5th to 95th percentile), and the two rolling series have correlation 0.69. A rolling Sharpe ratio alone hides this variation in the left tail.
The MDD evolution figure tracks the running maximum drawdown alongside a running Calmar ratio.

The running MDD is non-decreasing by definition. It steps up to 58% by fill 120, stays there until fill 1,558 and then jumps to 99.97%. The staircase shape, with long plateaus and sudden jumps, is the path-dependence that distinguishes Calmar from the distributional ratios.
4.4. Scaling verification
The scaling theorem predicts how each ratio responds to temporal aggregation when returns are i.i.d. We compute each ratio at aggregation windows of 1, 5, 15, 30 and 60 fills. At window size we partition the fills into non-overlapping blocks, sum the log returns within each block and compute the ratio on the coarse series. We then fit for each ratio.
The i.i.d. hypothesis can be tested directly. Shuffling the fill order keeps the distribution of returns and destroys any serial structure. We repeat the fit on 2,000 random permutations of the series. The table reports the fitted exponent for the real fill order, the 95% range over shuffles and the Gaussian prediction from Theorem (Scaling under coarse-graining) at the sample Sharpe ratio .
| Ratio | Gaussian i.i.d. prediction | Fill order | Shuffled, 95% range |
|---|---|---|---|
| Sharpe | 0.500 | 0.717 | [0.465, 0.606] |
| Information | 0.500 | 0.760 | [0.508, 0.649] |
| Sortino | 0.554 | 1.000 | [0.472, 0.712] |
| Calmar | at least 1 | 1.195 | [0.985, 1.028] |
| Omega () | 0.161 | 0.581 | [0.148, 0.270] |
| Rachev (5%, 5%) | 0.063 | 0.256 | [−0.106, 0.076] |
The shuffled series behave as the theorem predicts. The median shuffled Sharpe exponent is 0.502, the Sortino exponent 0.525 lies above as the positive Sharpe ratio requires, and the Calmar exponent 1.002 sits at the pathwise bound. The real fill order lies above the shuffled range for every ratio, with permutation p-values of 0.002 or less. The i.i.d. hypothesis fails for this series.
The cause is negative serial dependence. The lag-1 autocorrelation of fill returns is , and the variance of a -fill block sum is 0.20 times at and 0.18 times at . The two extreme fills of Section 4.1 fall in the same block at every window size in the table. Their and cancel to on aggregation. Aggregation therefore removes most of the variance and raises every ratio faster than i.i.d. scaling allows. With the pair removed the lag-1 autocorrelation is , the variance ratio at is 0.58 and the Sharpe exponent falls to 0.585. The dependence is weaker without the pair and still present.

4.5. Distribution diagnostics
The ratios disagree only when the strategies compared have different shapes (Section 3.7), and the scaling predictions of Section 3.8 assume i.i.d. returns. This subsection examines the shape directly.

The QQ-plot shows a body narrower than the Gaussian and isolated observations far out in both tails, the largest at and . The empirical density is sharply peaked. The Gaussian overlay misses both the peak and the tails.
A maximum-likelihood Student- fit gives degrees of freedom. A Student- law with has infinite variance, and the fitted model therefore has no Sharpe ratio. The finite sample variance reflects the finite sample.
Tail index estimation. The Hill estimator on the largest 10% of observations in each tail gives (right tail) and (left tail). On the largest 5% the estimates are 1.10 and 1.80, and on the largest 2% they are 1.44 and 1.49. The estimates move with the tail fraction, as Hill estimates do on small samples, and all of them lie below 2. They agree with the Student- fit: the data are compatible with infinite variance. The right tail estimate sits near 1, the boundary below which even the mean, and with it Omega and Rachev, would not exist.
The return series is therefore far from the regime of the Gaussian formulas. Its tail estimates put it in the domain where Proposition (Rachev tail asymmetry) applies: the sample Sharpe, Sortino and Information ratios are computable, but their population counterparts may not exist, while Omega and Rachev need only a finite mean.
5. Geometric Coda
The preceding sections treated performance ratios as functionals on a fixed probability model. This final section considers the space of all strategies and equips it with a natural geometric structure. The ratios become scalar fields on this space, and the question "which ratio should I optimise?" becomes a question about directions on a Riemannian manifold.
5.1. Fisher-Rao metric on strategy space
Let be a parameter space indexing a family of trading strategies. Each parameter vector determines a return distribution with density with respect to Lebesgue measure. As the trader adjusts (position sizes, entry thresholds, trade durations), the return distribution changes smoothly.
The natural metric on is the Fisher information matrix. For parameters , define
The matrix is symmetric and positive semi-definite. It is positive definite when the score functions are linearly independent, which can fail even for an identifiable model (Section 5.2 gives an example). Where is positive definite, the pair is a Riemannian manifold, called the statistical manifold of the strategy family.[11]
The Fisher-Rao metric has a statistical interpretation. The geodesic distance between two parameter values and measures how statistically distinguishable the corresponding return distributions are. Two strategies that are close in Euclidean parameter space may be far apart on the statistical manifold (if a small parameter change produces a large distributional shift), and vice versa.
5.2. Gradient divergence in strategy space
Each performance ratio from Section 3 defines a scalar field on the statistical manifold:
The Riemannian gradient is the direction of steepest ascent of with respect to the Fisher-Rao metric. In coordinates,
where denotes the entries of the inverse metric tensor, which exists where is positive definite. The gradient answers the question: in which direction should I perturb my strategy to achieve the greatest improvement in per unit of statistical distinguishability?
Different ratios can point in different directions. The angle
between the gradients of the Sharpe ratio and the Sortino ratio quantifies the tension between the two objectives.
Take the threshold .
-
On a location-scale family with a fixed generator and coordinates , the gradients satisfy with , so at every point, whatever the skewness of .
-
In the skew-normal family with coordinates the Fisher matrix is singular on the Gaussian slice , where the Riemannian gradients and are undefined.
Proof. (1) Standardise to mean 0 and variance 1 and absorb the constants into . By Theorem (Agreement within a location-scale family), on the family with strictly increasing. The chain rule gives , and raising the index with preserves the proportionality. The two gradients differ by the positive factor , and .
(2) The skew-normal density is
with and the standard normal density and distribution function. At the scores are and . Hence , and the vector satisfies .
Away from the metric is positive definite and the angle can be computed. At , with the Fisher matrix and both gradients evaluated by quadrature (scripts/verifications/performance_ratios_theory.py):
| −4 | −2 | −1 | −0.5 | 0.5 | 1 | 2 | 4 | |
|---|---|---|---|---|---|---|---|---|
| Skewness | −0.78 | −0.45 | −0.14 | −0.02 | 0.02 | 0.14 | 0.45 | 0.78 |
| Angle | 3.6° | 5.4° | 4.4° | 1.1° | 17.8° | 29.7° | 52.5° | 78.4° |
The angle depends on the sign of the skewness as well as its size. At it is 29.7° for right skew and 4.4° for left skew, and on the left-skewed side it falls and then rises again. Near the metric is nearly singular and the computed angle stays between about 5° and 9° at . It does not tend to zero as the distribution approaches the Gaussian.
The field above uses a toy two-parameter strategy family, bet fraction against lookback, whose Gram-Charlier skewness and kurtosis vary with the parameters, and it draws Euclidean gradients. Part (1) of the proposition locates the source of any angle. Moving only location and scale never separates the Sharpe and Sortino directions. A separation needs a direction in parameter space that changes the shape of the return distribution, and its size depends on which shape parameter moves and in which direction.
5.3. Connection to KL divergence
The Fisher-Rao metric has a second characterisation that connects it to information theory and to the Kelly criterion.
The Kullback-Leibler divergence between two return distributions in the strategy family is
This quantity measures how distinguishable the two strategies are from their return streams. It is non-negative, equals zero if and only if (for identifiable models), but is not symmetric and does not satisfy the triangle inequality.
The Fisher information matrix is the Hessian of the KL divergence at the identity:
Proof. Write as a function of the second argument, with the first fixed at . Expanding the logarithm,
The first term is independent of , so all derivatives fall on the second term. Define . Then
At , the score identity gives (the expected score is zero under the true model), so the first derivative vanishes: . This confirms that is a critical point.
For the second derivative,
At , use the identity and the fact that (the density integrates to 1 for all ). Therefore
This is the Fisher information matrix.
The proposition gives the Fisher-Rao metric a second reading: it measures the local curvature of the KL divergence around each strategy. A direction in parameter space along which the KL divergence grows quickly is a direction along which the return distribution changes rapidly, and the Fisher-Rao metric assigns a large norm to that direction.
This connects to the Kelly criterion. Kelly-optimal betting, as discussed in an earlier post, maximises the expected logarithmic growth rate. At fair odds the maximal rate equals the mutual information between the bettor's signal and the market outcome.[10] The mutual information is itself a KL divergence (between the joint distribution and the product of marginals). The Fisher-Rao metric is therefore the infinitesimal geometry underlying both Kelly betting and the performance ratios: both are governed by the curvature of information-theoretic divergences on the space of probability distributions.
5.4. Closing remark
Section 0 began with the observation that performance ratios are functionals on a chosen probability model. The intervening sections made this precise: the model determines the moments, partial moments, and tail functionals that each ratio consumes, and the relationships between ratios (agreement, divergence, scaling, ordering) are determined entirely by the distributional properties of the model.
The space of strategies, equipped with the Fisher-Rao metric, is a curved manifold, and each performance ratio defines a scalar field on it. Four results organise the relations between these fields. The Omega Sufficiency Theorem, a result of Keating and Shadwick, establishes that the full distribution is recoverable from the Omega functional. The entire taxonomy therefore rests on a single master object. The other three locate the separation between ratios:
Section 3.7 identifies when the rankings of different ratios can separate: never within one location-scale family, and possibly between families of different shape, symmetric ones included.
Section 3.8 identifies at which resolution the separation matters: under i.i.d. aggregation Sharpe scales exactly as the square root of the window, Sortino, Omega and Rachev drift away from that exponent as the aggregated Sharpe ratio grows, and per-period Calmar scales at least linearly. Section 4.4 shows that serial dependence in real returns can override all of these predictions.
Section 5.2 identifies in which direction the separation matters: the Sharpe and Sortino gradients coincide along location and scale and can diverge only along directions that change the shape of the return distribution.
Together these results support one conclusion. The choice of which performance ratio to optimise is a choice about which projection of the high-dimensional market state you trust most. The Sharpe ratio trusts the second moment. The Sortino ratio trusts the lower partial second moment. The Calmar ratio trusts the path-dependent maximum drawdown. The Omega ratio trusts the entire distribution (but in a one-dimensional compression via the threshold parameter). The Rachev ratio trusts the tails. No projection is universally dominant. The results above state when different projections agree, when they can disagree and what that disagreement means geometrically for strategy optimisation.
References
- Cont, R. (2001). Empirical properties of asset returns: stylized facts and statistical issues. Quantitative Finance, 1(2), 223-236. DOI
- Sortino, F. A. and van der Meer, R. (1991). Downside risk. Journal of Portfolio Management, 17(4), 27-31. DOI
- Cherny, A. and Madan, D. (2009). New measures for performance evaluation. Review of Financial Studies, 22(7), 2571-2606. DOI
- Keating, C. and Shadwick, W. F. (2002). A universal performance measure. Journal of Performance Measurement, 6(3), 59-84.
- Sharpe, W. F. (1966). Mutual Fund Performance. Journal of Business, 39(1), 119-138. DOI
- Magdon-Ismail, M. and Atiya, A. F. (2004). Maximum drawdown. Risk, 17(10), 99-102.
- Kazemi, H., Schneeweis, T. and Gupta, B. (2004). Omega as a performance measure. Journal of Performance Measurement, 8(3), 16-25.
- Biagini, S. and Pinar, M. C. (2013). The best gain-loss ratio is a poor performance measure. SIAM Journal on Financial Mathematics, 4(1), 228-242. DOI
- Johnson, W. B. and Lindenstrauss, J. (1984). Extensions of Lipschitz mappings into a Hilbert space. Contemporary Mathematics, 26, 189-206. DOI
- Kelly, J. L. Jr. (1956). A new interpretation of information rate. Bell System Technical Journal, 35(4), 917-926. DOI
- Amari, S. (2016). Information Geometry and Its Applications. Springer. DOI
- Artzner, P., Delbaen, F., Eber, J.-M. and Heath, D. (1999). Coherent measures of risk. Mathematical Finance, 9(3), 203-228. DOI
- Larsen, K. G. and Nelson, J. (2017). Optimality of the Johnson-Lindenstrauss lemma. In Proceedings of the 58th IEEE Symposium on Foundations of Computer Science (FOCS).
- Biglova, A., Ortobelli, S., Rachev, S. T. and Stoyanov, S. (2004). Different approaches to risk estimation in portfolio theory. Journal of Portfolio Management, 31(1), 103-112. DOI