Sharpe Ratio:

Estimation, Confidence Intervals, and Hypothesis Testing

Two Sigma Technical Report 2018-001

Matteo Riondato
Labs, Two Sigma Investments, LP
matteo@twosigma.com
June 14, 2018

Abstract
We survey and discuss methods proposed in the literature for 1. estimating the Sharpe
ratio; 2. computing confidence intervals around a point estimation of the Sharpe ratio; and 3.
performing hypothesis testing on a single Sharpe ratio and on the difference between two Sharpe
ratios.

1 Introduction
The Sharpe ratio [Sha65; Sha94] is a widely used measure of the performance of an investment
strategy (i.e., of a portfolio). Informally, the Sharpe ratio is the risk-adjusted expected excess return
of a portfolio w.r.t. a benchmark strategy. Formally, let p be a portfolio and let Rp be the return of
p over a time interval t (e.g., Rp could be the daily return, so t = one day).1 We assume that Rp is
a random variable with non-zero variance, i.e., Var[Rp ] > 0. Let also b be a benchmark investment
strategy, and let Rb denote its return over t. The benchmark b may be riskless, hence Rb may be a
fixed constant, or b may involve risk, in which case Rb is a random variable with non-zero variance.
In either case, we assume that Var[Rp − Rb ] exists and is non-zero.

Definition 1 (Sharpe ratio). The Sharpe ratio of p w.r.t. b is the ratio between the expected excess
return of p w.r.t. b over the standard deviation of this same excess return:

E[Rp − Rb ]
ζp,b = p . (1)
Var[Rp − Rb ]

Unless otherwise specified, we consider the portfolio p and the benchmark b to be fixed, hence
we almost always drop the subscripts from the notation of the Sharpe ratio, using ζ to denote ζp,b .
(1) (1) (n) (n)
Given a finite set D of observed returns from n time intervals, D = {(Rp , Rb ), . . . , (Rp , Rb )},
it is not possible, in general, to obtain the exact Sharpe ratio ζ from D. For all practical purposes,
a high-quality estimation ζ̂ of ζ, i.e., an estimation enjoying specific desirable statistical properties,
can be used in place of the exact value ζ. The goal is then to compute the best possible estimation ζ̂
1 We consider the case for absolute returns, i.e., the unit of measure for the values Rp is “dollars.” Another
possibility is to use logarithmic relative returns (“log-returns”), where Rp is the logarithm of the relative return
w.r.t. the previous period p − 1. The logarithm is necessary to preserve additivity. Most of what we discuss in this
work also applies to the case of log-returns. Whether to use absolute returns or log-relative returns must be a very
deliberate and informed decision. Scenarios in which using log-relative return would probably be preferable are when
the absolute returns are heteroskedastic.

1
 —Two Sigma Technical Report 2018-001— Matteo Riondato

of ζ from D. Once such a point estimation ζ̂ is available, it can be used to derive confidence intervals
containing the true value ζ, or to test hypotheses about ζ or about the difference ζp1 ,b − ζp2 ,b of the
Sharpe ratios of two different portfolios (w.r.t. the same benchmark).

Purpose of this document. Our goal is to survey existing methods presented in the econometrics
literature for estimating the Sharpe ratio, computing confidence intervals around a point estimation,
and performing hypothesis testing involving the Sharpe ratio. We also aim to provide a few humble
clarifications to some of the confusion in the existing literature.
The Sharpe ratio has received wide attention in the finance and economics literature, and it
is heavily relied upon by practitioners. Not all the attention it has received has been in the form
of praise, and many researchers have developed other measures or variants of the Sharpe ratio
to measure the performance of investment strategies (see, e.g., [Sha94; Sch07; Isr05; Bac09, and
references therein]). We do not comment on which measure should be used and in which cases
one measure is better than another. Instead, we focus on the Sharpe ratio and how to estimate it
correctly from a statistical point of view. A similar study would be warranted if we were to consider
other performance measures.

Outline. This document is organized as follows. In Sect. 2 we present methods for point estimation
of the Sharpe ratio. Sect. 3 contains procedures to obtain confidence intervals around a point
estimation. We discuss hypothesis tests for a single Sharpe ratio and for the difference of two Sharpe
ratios in Sect. 4. We conclude in Sect. 5, presenting some open questions and research directions.
In Appendix A we discuss the double Sharpe ratio, an alternative measure of performance for
investment strategies.

2 Estimation
In this section we present estimators for ζ and discuss their properties, including bias, variance,
efficiency, and asymptotic distribution. We also touch on time aggregation (e.g., annualization),
given its important role in practice.

Notation. In our general setting, each excess return Rp −Rb is a random variable, which we denote
with Y . Independently from the distribution p of Y , the expectation of Y is denoted with µ = E[Y ]
and the standard deviation of Y with σ = Var[Y ]. We use µ̂ to denote the sample mean of a set
D = {Y (1) , . . . , Y (n) } of n sample excess returns, and σ̂ to denote the sample standard deviation
computed from D. Formally:
v
n n
1 X (i) u 1 X
u

2
µ̂ = Y and σ̂ = t Y (i) − µ̂ .
n i=1 n − 1 i=1

Definition 2. The basic estimator ζ̂basic for ζ is defined as the ratio between the sample mean µ̂
and the sample standard deviation σ̂:
µ̂
ζ̂basic = . (2)
σ̂
As observed by Miller and Gehr [MG78], the basic estimator is closely related to the Student’s
t-statistic for testing whether the expectation of a random variable is zero:
µ̂ √
t= √ = nζ̂basic . (3)
σ̂/ n

Hence results that hold for this t-statistic also hold for the Sharpe ratio, up to the scaling factor, as
we discuss in the following.

2
Sharpe Ratio —Two Sigma Technical Report 2018-001—

2.1 Distribution under normal i.i.d. excess returns

We now look at a first example of the close relationship between the t-statistic from (3) and ζ̂basic .
First, we temporarily make the following key assumption, which simplifies the analysis but is
extremely unrealistic, especially in finance settings [LM02]. We remove this assumption later.

Assumption 1 (Independence and normality of the excess returns). The excess returns Y (i) , 1 ≤
i ≤ n, are i.i.d. samples from a normal distribution N (µ, σ 2 ).

Under Assumption 1, we have the following fact about the distribution of ζ̂basic .

Fact 1 (Sect. 17.3.2 [SS10]). Assume that Assumption 1 holds. Then the distribution of ζ̂basic is
√ rescaled non-central t-distribution with n − 1 degrees of freedom. The non-centrality parameter is
a
nζ.

2.2 Bias and variance

We now discuss the first and second moments of the basic estimator, i.e., its expectation and variance.

Bias. The bias of ζ̂basic is a first symptom of the limitations of this estimator. Under some
assumption on the distribution of the excess returns, it is possible to quantify (or approximate) the
bias of ζ̂basic . Specifically:

• under Assumption 1, the exact expectation of ζ̂basic is [Pav15, Sect. 1.3]:

n − 1 Γ n−2
r

E[ζ̂basic ] = ζ 2

(> ζ) ,
2 Γ n−12

where Γ(x) = (x − 1)! is the gamma function. An unbiased estimator ζ̂unbiased for ζ can be
obtained by dividing ζ̂basic by the bias factor on the r.h.s. of the above equation.
The bias factor is greater than 1 and tends towards 1 from above quite rapidly as n grows.
For example, it is ≈ 1.08 for n = 12, ≈ 1.02 for n = 40, and ≈ 1.01 for n = 75.

• under the assumption that the excess returns are i.i.d. (but not necessarily normal), the ex-
pectation of ζ̂basic can be rewritten as [Bao09, Sect. 1.2]:2

3 49 1 3 3 15 3 5
   
E[ζ̂basic ] = ζ + ζ+ ζ − γ1 + + γ2 ζ − + 2 γ3 − γ4 ζ
4n 32n2 2n 8n2 8n 32n2 8n 16n2
5 105 2 15
− 2 γ12 ζ + γ2 ζ − γ1 γ2 + o(n−2 ) (4)
4n 128n 2 16n2
where
E[(Y − µ)i+2 ]
γi = , i = 1, 2, 3, 4,
σ i+2
i.e., γ1 is the skewness, γ2 is the kurtosis, and γ3 and γ4 are the third and fourth standardized
moments of the distribution of the excess returns, according to the so-called Pearson’s definition
(e.g., a normal distribution has γ1 = 0 and γ2 = 3).
Stopping at the first order term, the expectation can be written as [Bao09, Remark 2]:
3 1 3
E[ζ̂basic ] = ζ + ζ− γ1 − γ2 ζ + o(n−1 ) . (5)
4n 2n 8n
2 Christie [Chr05] and Opdyke [Opd07] also give expressions for the expectation of ζ̂
basic under the same assump-
tions, but Bao [Bao09] observes that such expressions are not correct, because they consider µ̂ to be a deterministic
value.

3
 —Two Sigma Technical Report 2018-001— Matteo Riondato

The expressions on the r.h.s. of (4) and (5) are not proportional to ζ. Hence developing
an (approximately) unbiased estimator becomes more problematic than under Assumption 1,
given that ζ, γ1 , γ2 , γ3 , and γ4 are unknown. In practice, it is possible to replace the unknown
values with estimates in, e.g., (5), and obtain an approximately unbiased estimator:
3 1 3
ζ̂apprunbiased = ζ̂basic − ζ̂basic + γ̂1 − ζ̂basic γ̂2 . (6)
4n 2n 8n
Estimates of γ1 and γ2 can be obtained using Fisher’s k statistics [Wei16].

Variance. A closed formula for the variance of ζ̂basic can be obtained under Assumption 1 [Pav15,
Sect. 1.3].
Fact 2. Under Assumption 1, we have

(1 + nζ 2 )(n − 1)
Var[ζ̂basic ] = − E[ζ̂basic ]2 .
n(n − 3)

When the excess returns are i.i.d. but not necessarily normal, it is possible to obtain an approx-
imation of the variance [Bao09, Sect. 1.2].
Fact 3. Under the assumption that the excess returns are i.i.d. samples, we have

1 ζ2 1 19 2 1 5 1 3
       
Var[ζ̂basic ] = 1+ + 2 ζ + 2 − γ1 ζ + + γ2 ζ 2
+
n 2 n 8 n 2n2 4n 8n2
5 3 7 3 2 39 2 2 15
 
+ 2 γ3 ζ − 2 γζ 2 + γ12 − ζ + γ ζ − 2 γ1 γ2 ζ + o(n−2 ) .
4n 8n 4n2 2n 32n2 2 4n

When the assumption of normality and independence does not hold, one can estimate the variance
of the estimator using the bootstrap techniques appropriate for non i.i.d. data [LW08; PR92]. In
Sect. 3 we give more details about applications of the bootstrap in the context of confidence intervals
estimation for the Sharpe ratio. We discuss the asymptotic variance of the basic estimator in
Sect. 2.4.

2.3 Best scale invariant estimator and other estimators

Unhapipat, Chen, and Pal [UCP16] propose a different estimator for ζ. Specifically, they consider
the parametric class of scale-invariant estimators
 
µ̂
C = ζ̂d = d , d ∈ R
σ̂

and identify the value d∗ such that ζ̂d∗ has the minimum Mean Squared Error (MSE) among all the
members of C. The optimal d∗ is

n−3 Γ n−2


ζ2
d =p
∗ 2

,
2(n − 1) ζ 2 + 1/n Γ n−1
2

which involves the unknown exact Sharpe ratio ζ. Unhapipat, Chen, and Pal [UCP16] thus propose
to use
n − 3 Γ n−2


d= p 2

(≥ d∗ ) .
2(n − 1) Γ n−1
2

The resulting estimator is called the best scale invariant estimator,3 denoted with ζ̂bsie .
3 The adjective “best” is slightly abused here, as it would be justified only when using d∗ .

4
Sharpe Ratio —Two Sigma Technical Report 2018-001—

Other estimators. Challet [Cha15] presents a completely different approach to Sharpe ratio
estimation, as his estimator does not use the sample mean and sample standard deviation. Instead,
he argues that the total duration of the drawdowns and the drawups of a price time series is an
efficient estimator of the Sharpe ratio. His derivation relies on the assumption that the returns are
i.i.d.. Challet’s estimator is much more efficient than ζ̂basic when the distribution of the returns
has fat tails, but not otherwise. This estimator was introduced very recently, and no results are
currently known on its properties.
Only for the sake of completeness, we mention the estimator proposed by Skrepnek and Sahai
[SS13]. This estimator resembles that of Unhapipat, Chen, and Pal [UCP16], with an additional
factor in the numerator, (claiming to be) taking into account an estimation of the coefficient of
variation. Evaluating the correctness and the performances of this estimator is a strenuous task,
due to lack of rigorous proofs in the paper. Similar issues also make it hard to evaluate the estimators
presented in another work by the same authors [SS11].

2.4 Asymptotic distribution

We now present results on the asymptotic normality of the basic estimator of the Sharpe ratio, i.e.,
on the fact that its distribution tends toward a normal distribution as the sample size n grows.
The most important aspect of this convergence is the asymptotic variance.4 We start from a result
requiring a most restrictive assumption on the distribution of the excess returns and then show how
it can be relaxed using the Generalized Method of Moments (GMM) [Han82]. We present the results
for ζ̂basic , but they also hold for ζ̂unbiased and ζ̂bsie , under the same assumptions [UCP16].

Assuming normal i.i.d. excess returns. The following result is a straightforward application
of the Central Limit Theorem and of the delta method [Was03, Sect. 9.9].
Fact 4 ([Opd07; SS10]5 ). Under Assumption 1,6 the basic estimator ζ̂basic is asymptotically normal
in n, i.e.,
√ ζ2
 
d
→ N 0, 1 +
n(ζ̂basic − ζ) − . (7)
2

Assuming a stationary distribution of excess returns. Mertens [Mer02] observes that wrongly
assuming normality of the excess returns leads to estimates of the asymptotic variance that may be
up to 70% off from their true values. The assumption of normality can be removed and the asymp-
totic variance can be computed under just the independence assumption. We focus here on the even
more general case of that the process generating the excess returns is stationary and ergodic. In
this case one can use the GMM [Han82] to prove the asymptotic normality of ζ̂basic and compute
the asymptotic variance.
Fact 5 ([Lo02; LW08]). Assume that the stochastic process from which the excess returns are sampled
is stationary and ergodic.7 Let m2 = E[Y 2 ], and for any time interval t let
 
Yt − µ
θt =
Yt2 − m2
where Yt is the excess return in the interval t. Let also
" m
#
2
(m2 −µ2 )3/2
q= µ .
− 2(m2 −µ 2 )3/2

4 One could argue that the rate of convergence is at least as important an aspect. No results are currently available

on the rate of convergence for the basic estimator.

5 Lo [Lo02] gives the same result with the additional claim that it holds for the case of non-necessarily-normal

i.i.d. excess returns but Bao [Bao09, Footnote 5] and Mertens [Mer02] observe that the claim is not justified.
6 More precisely, under the assumption that the excess returns are i.i.d. samples from a distribution with the same

constraints on the first four central moments as the normal [Mer02].

7 Some additional technical assumptions are needed but are quite standard [Lo02; LW08].

5
 —Two Sigma Technical Report 2018-001— Matteo Riondato

Then
n n
√ d 1 XX
→ N (0, q | Ψq) where Ψ = lim
n(ζ̂basic − ζ) − E[θs θt| ] .8 (8)
n→∞ n
s=1 t=1

An estimation of the asymptotic variance in this case can be computed by first obtaining an
Heteroskedastic and Autocorrelation Consistent (HAC) estimation of Ψ using the kernel methods
by Andrews [And91], and then an estimation of d from the sample excess returns.
For completeness, we mention the work by Bao and Ullah [BU06], who discuss the asymptotic
distribution of ζ̂basic when the excess returns are normal but not necessarily i.i.d.. There is significant
evidence suggesting that real excess returns are not normally distributed, so we do not discuss the
results by Bao and Ullah [BU06] any further. Pav [Pav15, Sect. 1.7.1] presents a discussion of a
similar case, but assuming fixed autocorrelation. He also discusses adjustements to ζ̂basic under a
limited case of heteroskedasticity [Pav15, Sect. 1.7.2].

2.5 Time aggregation (e.g., annualization)

We discuss now an important aspect of the Sharpe ratio that, despite not being a specific charac-
teristic of the basic estimator nor strictly related to estimation, must be taken into consideration in
practice: the Sharpe ratio is not independent from the time period t for which the excess returns Y (i)
are considered [Sha94; Lo02]. In other words, a Sharpe ratio that considers returns over monthly
periods cannot be compared directly to one that considers returns over years. The same holds,
naturally, also for the basic estimator. It is possible to “transform” a Sharpe ratio using returns
measured at higher frequency (e.g., monthly) into a Sharpe ratio using returns measured at lower
frequency (e.g., yearly).9
Suppose that we want to transform a Sharpe ratio measured w.r.t. shorter time periods of length
t (e.g., t = one month) into a Sharpe ratio measured w.r.t. longer time periods of length T (e.g., T =
one year). Assume that t divides T evenly and let q = T /t. Given a sequence of q excess returns
w.r.t. q consecutive time periods of length t, the excess return for the same period of length T is,
ignoring the effects of compounding,
q
(i)
X
YT = Yt .
i=1

Under the assumption that the excess returns are i.i.d., the variance of YT is proportional to q.
Hence the Sharpe ratio ζT w.r.t. periods of length T is

E[YT ] qE[Yt ] √
ζT = p =p = qζt , (9)
Var[YT ] qVar[Yt ]

where ζt is the Sharpe ratio w.r.t. periods of length t [Lo02].

8 We would like to clarify some of the confusion in the econometrics literature about the correctness of this result.
√
Lo [Lo02] uses a GMM approach to compute the asymptotic variance of n(θ̂ − θ), where θ = (µ, σ 2 )| and
2 |
θ̂ = (µ̂, σ̂ ) . He then uses the delta method to obtain the result in (8) (in a slightly different form). Ledoit and
Wolf [LW08] present essentially the same approach of Lo [Lo02], with a slightly different definition of the moment
conditions (using m2 instead of σ 2 ), which is what we follow.
Christie [Chr05] attacks the problem using a different GMM approach that does not require a successive application
of the delta method. When the results are stated in the main body of the paper, the stated assumptions are that
the process is stationary and ergodic. In the appendix, when the results are proved, it is mentioned that normality
is not assumed, but it is not clearly stated whether independence is assumed. Bao [Bao09, Footnote 6] comments
that it is “questionable” that Christie’s “asymptotic variance expression holds under the very general setup of non-
i.i.d. returns”, but he does not point to specific mistakes, while Opdyke [Opd07] embraces Christie’s result and claims
that it holds for non-i.i.d. excess returns. A close look at Christie’s derivation suggests that it may only hold for the
i.i.d. case. Specifically, his definition of the variance-covariance matrix of the moment conditions [Chr05, Eq. 8] may
only hold for the i.i.d. case. The definitions of the same matrix in [LW08, Sect. 3.1] and in [Lo02, Eq. A7] instead
seem to hold for the general case..
9 The opposite direction, i.e., transforming from lower to higher frequency, is mathematically possible, but not

sensible from a statistical point of view.

6
Sharpe Ratio —Two Sigma Technical Report 2018-001—

√
A basic estimation for ζT is ζ̂T = q ζ̂t . This is just the basic estimator and has the properties
discussed in the previous paragraphs. For example, it is asymptotically normal, although with an
appropriately scaled higher variance:
√ ζt2
  
d
T (ζ̂T − ζT ) −
→ N 0, q 1 + .
2

The simple scaling of ζt to obtain ζT cannot be used when the returns are not i.i.d., as the variance
(i)
of YT is no longer the sum of the variances of the Yt [Lo02]. Instead, the scaling factor becomes
a more complex expression of the autocorrelations of the excess returns. For ease of presentation,
we only present the case of stationary excess returns. Lo [Lo02] discusses the general case for
non-i.i.d. excess returns. Let ρk be the k-th order autocorrelation of Yt :10

Cov[Yt , Yt−k ]
ρk = .
Var[Yt ]

Then the relationship between ζt and ζT is

q
ζT = q Pq−1 ζt . (10)
q + 2 k=1 (q − k)ρk

When estimating ζT with ζ̂T using ζ̂t , the autocorrelations ρk must be replaced by appropriate
estimations ρ̂k . The accuracy of these estimations may have an impact on the performances of ζ̂T ,
including its asymptotic variance.

3 Confidence intervals
In this section we discuss how to compute confidence intervals for the Sharpe ratio. We start from
methods that use the exact distribution of the basic estimator, and then we move to approaches
relying on its asymptotic variance (see Sect. 2.4) and to bootstrap-based methods.

3.1 Confidence intervals with normal i.i.d. returns

From Fact 1, we have that under Assumption 1, we can obtain a confidence interval on ζ by invert-
ing the cumulative distribution of the non-central t-distribution [Pav15]. The resulting confidence
interval would have exactly the desired nominal coverage, but its computation may be too expensive,
due to the necessity of inverting the CDF.

3.2 Confidence intervals with the asymptotic variance

Approximate confidence intervals, i.e., confidence intervals whose actual coverage is not guaranteed
to be the nominal coverage, can be obtained by exploiting the fact that the basic estimator is
asymptotically normal (Sect. 2.4) [UCP16; Lo02; Mer02].

The plugin approach. The derivation of the approximate confidence intervals may follow a well-
known recipe [Was03, Thm. 6.16], known as the plugin approach. Let ζ̂ be a point estimate of ζ and
ˆ be the estimated standard error of ζ̂. Given α ∈ (0, 1), let zα/2 be the 1 − α/2-th quantile of
let se
the standard normal distribution, and define

ˆ ζ̂ + zα/2 se)
C = (ζ̂ − zα/2 se, ˆ .
10 On k ≥ 2. In this
the basis of the efficient market hypothesis, one can assume the daily returns have nρk = 0 for all p
case, the annualization can be (approximately) obtained by just multiplying the daily Sharpe ratio by q/1 + 2ρ1 .

7
 —Two Sigma Technical Report 2018-001— Matteo Riondato

Then C is an approximate confidence interval for ζ, with approximate coverage 1 − α:

Pr(ζ ∈ C) → 1 − α,

where the probability is taken w.r.t. the data-generating process.

Unhapipat, Chen, and Pal [UCP16] present experimental evidence that, in general, the confidence
intervals obtained following this procedure fall short of the nominal coverage.

“Confidence interval – Hypothesis testing”–equivalency approach. Approximate confi-

dence intervals can also be computed by inverting the expression

(ζ̂ − ζ)
∈ [−zα/2 , zα/2 ],
se(ζ)

where se(ζ) is the standard error of ζ̂, where we highlighted the dependency on ζ. Inverting this
expression allows us to obtain inequalities for ζ. The square root of the estimated asymptotic
variance can be used in place of the actual standard error.

3.3 Confidence intervals with the bootstrap

No result is currently available on the rate of convergence of Sharpe ratio point estimators to normal-
ity, and the confidence intervals obtained using the asymptotic variance may have less than nominal
coverage. Improved confidence intervals can be obtained using the bootstrap [Was03, Ch. 8], espe-
cially the Studentized version [LW08, and references therein].11 Specifically, let ζ̄ be the original
point estimate, and let se¯ be the original estimation of the standard error. For any bootstrap re-
sample, let ζ̂ be the point estimate obtained from that resample, and let se b be the estimate of the
standard error obtained from the resample. Define, for each resample,

ζ̂ − ζ̄
T = .
se
b
An (approximate) 1 − α confidence interval for ζ can be obtained as

¯ ζ̄ − T1−α/2 se
C = ζ̄ − Tα/2 se, ¯ , 12

where Tβ is the 1 − β percentile of the values obtained from the bootstrap resamples.
Particular care must be taken when creating the bootstrap resamples when the original data is
not a collection of i.i.d. samples. For example, Ledoit and Wolf [LW08] suggest the use of the circular
block bootstrap [PR92]. The issue in using such a method in practice is that it involves fitting a
semi-parametric model (e.g., VAR, GARCH) to the observed data, which could be computationally
expensive and not theoretically motivated. Moreover, one may argue that the same model could be
assumed in the derivation of the asymptotic variance, resulting in better confidence intervals in this
case [SS10].
Unhapipat, Chen, and Pal [UCP16] report experimental results showing that, under the assump-
tion of i.i.d. samples, the confidence intervals obtained with the simple (non-Studentized) bootstrap
have coverage closer to the nominal than those obtained using the asymptotic normality.

4 Hypothesis testing
In this section we present statistical tests for the Sharpe ratio and for the difference of two Sharpe
ratios.
11 Vinod and Morey [VM99b] also present confidence intervals for the Sharpe ratio based on the bootstrap but

Ledoit and Wolf [LW08] observes that their application of the Studentized bootstrap is incorrect.
12 The “−” in the computation of the upper bound is not a typo: the value T
1−α/2 is negative because it is the α/2
percentile of the bootstrap distribution which is centered at zero.

8
Sharpe Ratio —Two Sigma Technical Report 2018-001—

4.1 Tests on a single Sharpe ratio

The task requires us to test

H0 : ζ = ζ 0 versus Ha : ζ 6= ζ0 .

It is easy to derive tests using the asymptotic normality of the point estimates [UCP16; Chr05;
LW08]. Let V (ζ0 ) be the asymptotic variance computed under H0 (see (7) and (8)), then H0 can
be rejected with significance level α if
√
n|ζ̂ − ζ0 | > zα/2 V (ζ0 ),
p

where zα/2 is the 1 − α/2-th -quantile of the standard normal distribution. Not surprisingly, given
the use of the asymptotic distribution, this test has low statistical power, i.e., it rejects a true null
hypothesis in more than a 1 − α fraction of the cases [LW08].
When using the bootstrap to compute a confidence interval C, it is possible to reject H0 with
nominal level α if ζ0 6∈ C.

4.2 Tests on the difference of two Sharpe ratios

A number of works [LW08; Chr05; SS10; VM99b; Opd07] study the problem of comparing the
Sharpe ratios of two investment strategies (w.r.t. the same benchmark). Specifically, given two
Sharpe ratios ζ1 and ζ2 , the goal is testing the following hypotheses:

H0 : ∆ = ζ1 − ζ2 = 0 versus Ha : ∆ 6= 0

The same approaches discussed for the estimation and testing of a single Sharpe ratio extend to
this case,13 including approaches using the GMM [Chr05; SS10], or the bootstrap [LW08; Opd07;
VM99b].
Methodological [LW08, Remarks 3.1 and 3.3] and empirical [AS13] evidence suggests that the
bootstrap-based approach by Ledoit and Wolf [LW08] is theoretically sound and outperforms other
methods. On the other hand, as mentioned in Sect. 3.3, a correct application of the Studentized cir-
cular block bootstrap requires fitting a semi-parametric model, a non-straightforward and potentially
computationally expensive operation.

5 Conclusions and research directions

We presented and discussed existing methods for estimation, computation of confidence intervals,
and hypothesis testing of the Sharpe ratio. None of the approaches presented in the literature
are entirely satisfying. For example, methods deriving or using the asymptotic normality of the
estimator are known to perform worse (according to various metrics) than methods based on the
bootstrap. On the other hand, a correct application of bootstrap techniques taking into account the
time-series nature of the data requires impractical steps (fitting of a semi-parametric model), which
may hinder the usefulness of the methods. Ignoring the correlation structure of the data simplifies
the application of bootstrap-based methods, but doing so must be a deliberate choice.
Interesting directions for research include the derivation of more stringent confidence intervals
(e.g., studying the problem using martingales), and the development of more descriptive expressions
for the asymptotic variance when some information on the correlation structure of the data is
known (Schmid and Schmidt [SS10] present some results in this direction). Additionally, it would be
interesting to study the properties of Challet’s estimator [Cha15], given its unconventional approach.
13 More precisely, many of the approaches were originally motivated by the need of testing the difference of two

Sharpe ratios, but are easy to extend (and the authors often do that or at least mention this possibility) to the case
of a single Sharpe ratio.

9
 —Two Sigma Technical Report 2018-001— Matteo Riondato

References
[And91] Donald W. K. Andrews. “Heteroskedasticity and autocorrelation consistent covariance
matrix estimation”. In: Econometrica 59.3 (1991), pp. 817–858.
[AS13] Benjamin R. Auer and Frank Schuhmacher. “Performance hypothesis testing with the
Sharpe ratio: The case of hedge funds”. In: Finance Research Letters 10.4 (2013), pp. 196–
208.
[Bac09] Carl Bacon. How sharp is the Sharpe ratio? Risk-adjusted Performance Measures. Tech-
nical Report. StatPro, 2009.
[Bao09] Yong Bao. “Estimation Risk-Adjusted Sharpe Ratio and Fund Performance Ranking un-
der a General Return Distribution”. In: Journal of Financial Econometrics 7.2 (2009),
pp. 152–173.
[BU06] Yong Bao and Aman Ullah. “Moments of the estimated Sharpe ratio when the observa-
tions are not IID”. In: Finance Research Letters 3.1 (2006), pp. 49–56.
[Cha15] Damien Challet. “Moment-Free Sharpe Ratio Estimation from Total Drawdown Dura-
tions”. 2015. url: http://ssrn.com/abstract=2603682.
[Chr05] Steve Christie. Is the Sharpe ratio useful in asset allocation? MAFC Research Papers 31.
Applied Finance Center, Macquarie University, 2005.
[Han82] Lars Peter Hansen. “Large Sample Properties of Generalized Method of Moments Esti-
mators”. In: Econometrica 50.4 (1982), pp. 1029–1054.
[Isr05] Craig Israelsen. “A refinement to the Sharpe ratio and information ratio”. In: Journal of
Asset Management 5.6 (2005), pp. 423–427.
[LM02] Andrew W. Lo and A. Craig MacKinlay. A non-random walk down Wall Street. Princeton
University Press, 2002.
[Lo02] Andrew W. Lo. “The statistics of Sharpe ratios”. In: Financial Analysts Journal 58.4
(2002), pp. 36–52.
[LW08] Oliver Ledoit and Michael Wolf. “Robust performance hypothesis testing with the Sharpe
ratio”. In: Journal of Empirical Finance 15.5 (2008), pp. 850–859.
[Mer02] Elmar Mertens. “Comments on the correct variance of estimated Sharpe Ratios in Lo
(2002, FAJ) when returns are IID”. 2002. url: http : / / www . elmarmertens . com /
research/discussion/soprano01.pdf.
[MG78] Robert E. Miller and Adam K. Gehr. “Sample size bias and Sharpe’s performance mea-
sure: A note”. In: Journal of Financial and Quantitative Analysis 13.05 (1978), pp. 943–
946.
[Opd07] John Douglas Opdyke. “Comparing Sharpe ratios: so where are the p-values?” In: Journal
of Asset Management 8.5 (2007), pp. 308–336.
[Pav15] Steven E. Pav. “Notes on the Sharpe ratio”. 2015. url: http://cran.uni-muenster.
de/web/packages/SharpeR/vignettes/SharpeRatio.pdf.
[PR92] Dimitris N. Politis and Joseph P. Romano. “A circular block-resampling procedure for
stationary data”. In: Exploring the limits of bootstrap. Ed. by Raoul LePage and Lynne
Billard. Wiley, 1992, pp. 263–270.
[Sch07] Hendrik Scholz. “Refinements to the Sharpe ratio: Comparing alternatives for bear mar-
kets”. In: Journal of Asset Management 7.5 (2007), pp. 347–357.
[Sha65] William F. Sharpe. “Mutual fund performance”. In: Journal of Business 39.1 (1965),
pp. 119–138.
[Sha94] William F. Sharpe. “The Sharpe ratio”. In: The Journal of Portfolio Management 21.1
(1994), pp. 49–58.

10
Sharpe Ratio —Two Sigma Technical Report 2018-001—

[SS10] Friedrich Schmid and Rafael Schmidt. “Statistical Inference for Sharpe Ratio”. In: In-
terest Rate Models, Asset Allocation and Quantitative Techniques for Central Banks and
Sovereign Wealth Funds. Ed. by Arjan B. Berkelaar, Joachim Coche, and Ken Nyholm.
Palgrave Macmillan UK, 2010, pp. 337–357.
[SS11] Grant H. Skrepnek and Ashok Sahai. “An Estimation Error Corrected Sharpe Ratio
Using Bootstrap Resampling”. In: Journal of Applied Finance and Banking 1.2 (2011),
pp. 189–206.
[SS13] Grant H. Skrepnek and Ashok Sahai. “Efficient Point Estimation of the Sharpe Ratio”.
In: Journal of Statistical and Econometric Methods 2.4 (2013), pp. 129–142.
[UCP16] Suntaree Unhapipat, Jun-Yu Chen, and Nabendu Pal. “Small Sample Inferences on the
Sharpe Ratio”. In: American Journal of Mathematical and Management Sciences 35.2
(2016), pp. 105–123.
[VM99a] Hrishikesh D. Vinod and Matthew R. Morey. “A double Sharpe ratio”. 1999. url: http:
//ssrn.com/abstract=168748.
[VM99b] Hrishikesh D. Vinod and Matthew R. Morey. “Confidence intervals and hypothesis testing
for the Sharpe and Treynor performance measures: A bootstrap approach”. In: Compu-
tational Finance. The MIT Press, 1999, pp. 25–39.
[Was03] Larry Wasserman. All of statistics: a concise course in statistical inference. Springer
Science & Business Media, 2003.
[Wei16] Eric W. Weisstein. “k-Statistic”. from MathWorld – A Wolfram Web Resource. 2016.
url: http://mathworld.wolfram.com/k-Statistic.html.

Appendix A The double Sharpe ratio

Vinod and Morey [VM99a] introduce the double Sharpe ratio, a modified version of the basic es-
timator that takes into account the standard deviation of ζ̂basic . The double Sharpe ratio is not
actually an estimator for the Sharpe ratio ζ but rather (an estimator for) an alternative measure
of the performance of an investment strategy. We present it here mainly because it attracted our
interest.
Definition 3 (Double Sharpe ratio). The double Sharpe ratio is defined as

ζ̂basic
ζ̂double = q . (11)
Var[ζ̂basic ]

The standard deviation of the basic estimator appearing in the definition of ζ̂double is unknown,
and in practice it is estimated using the bootstrap procedure [Was03, Ch. 8].
Vinod and Morey [VM99a] do not clarify whether the numerator in (11) should actually be ζ̂basic
or the mean of the estimations of the Sharpe ratio created during the bootstrap (or even the median).
The choice may have a practical impact: in the experimental results reported by Vinod and Morey
[VM99a], the latter quantity was always (slightly) larger than the former. The authors argue that
this fact can be explained by the non-normality and skewness properties of the bootstrap sampling
distribution of the basic estimator, although they do not explain why such a distribution should be
expected, and the argument is not entirely convincing. We hypothesize that this overestimation is
due to the convexity of the Sharpe ratio, and would follow from Jensen’s inequality. Our hypothesis
is inspired by an observation by Christie [Chr05, App. C].
Bao [Bao09] introduces a minor variant of the double Sharpe ratio, which takes the bias of ζ̂basic
into account in the numerator of (11).

11