Backtesting

Campbell R. Harvey
Duke University, Durham, NC 27708 USA
National Bureau of Economic Research, Cambridge, MA 02138 USA

Yan Liu
Texas A&M University, College Station, TX 77843 USA

Current version: October 4, 2014

Abstract
When evaluating a trading strategy, it is routine to discount the Sharpe Ratio from a historical backtest. The reason is simple: there is inevitable data
mining by both the researcher and by other researchers in the past. Our paper
provides a statistical framework that systematically accounts for these multiple
tests. We propose a method to determine the appropriate haircut for any given
reported Sharpe Ratio. We also provide a profit hurdle that any strategy needs
to achieve in order to be deemed significant.

Keywords: Sharpe Ratio, Multiple tests, Backtest, Haircut, Data mining,

Overfitting, Data snooping, VaR, Value at Risk, Out-of-sample tests, Trading
strategies, Minimum profit hurdle.

First posted to SSRN, October 25, 2013. Send correspondence to: Campbell R. Harvey,
Fuqua School of Business, Duke University, Durham, NC 27708. Phone: +1 919.660.7768, E-mail:
[Link]@[Link] . We appreciate the comments of Scott Linn, Marcos Lopez de Prado, Bernhard Scherer, Christian Walder, Nico Weinert and the seminar participants at the Inquire Europe
UK, Man-AHL, APG, and CPPIB seminars.

Electronic copy available at: [Link]

Introduction

A common practice in evaluating backtests of trading strategies is to discount the

reported Sharpe Ratios by 50%. There are good economic and statistical reasons for
reducing the Sharpe Ratios. The discount is a result of data mining. This mining may
manifest itself by academic researchers searching for asset pricing factors to explain
the behavior of equity returns or by researchers at firms that specialize in quantitative
equity strategies trying to develop profitable systematic strategies.
The 50% haircut is only a rule of thumb. The goal of our paper is to develop an
analytical way to determine the magnitude of the haircut.
Our framework relies on the statistical concept of multiple testing. Suppose you
have some new data, Y, and you propose that variable X explains Y. Your statistical
analysis finds a significant relation between Y and X with a t-ratio of 2.0 which has
a probability value of 0.05. We refer to this as an independent test. Now consider
the same researcher trying to explain Y with variables X1 , X2 , . . . , X100 . In this case,
you cannot use the same criteria for significance. You expect by chance that some of
these variables will produce t-ratios of 2.0 or higher. What is an appropriate cut-off
for statistical significance?
In Harvey and Liu (HL, 2014), we present three approaches to multiple testing.
We answer the question in the above example. The t-ratio is generally higher as the
number of tests (or X variables) increases.
Consider a summary of our method. Any given strategy produces a Sharpe Ratio.
We transform the Sharpe Ratio into a t-ratio. Suppose that t-ratio is 3.0. While a
t-ratio of 3.0 is highly significant in an independent test, it may not be if we take
multiple tests into account. We proceed to calculate a p-value that appropriately
reflects the multiple testing. To do this, we need to make an assumption on the
number of previous tests. For example, Harvey, Liu and Zhu (HLZ, 2014) document
that at least 315 factors have been tested in the quest to explain the cross-sectional
patterns in equity returns. Suppose the adjusted p-value is 0.05. We then calculate
an adjusted t-ratio which, in this case, is 2.0. With this new t-ratio, we determine a
new Sharpe ratio. The percentage difference between the original Sharpe ratio and
the new Sharpe ratio is the haircut.
The Sharpe Ratio that obtains as a result of the multiple testing has the following
interpretation. It is the Sharpe Ratio that would have resulted from an independent
test, that is, a single measured correlation of Y and X.
We argue that it is a serious mistake to use the rule of thumb 50% haircut. Our
results show that the multiple testing haircut is nonlinear. The highest Sharpe Ratios
are only moderately penalized while the marginal Sharpe Ratios are heavily penalized.

1
Electronic copy available at: [Link]

This makes economic sense. The marginal Sharpe Ratio strategies should be thrown
out. The strategies with very high Sharpe Ratios are probably true discoveries. In
these cases, a 50% haircut is too punitive.
Our method does have a number of caveats some of which apply to any use of
the Sharpe Ratio. First, high observed Sharpe Ratios could be the results of nonnormal returns, for instance an option-like strategy with high ex ante negative skew.
In this case, Sharpe Ratios should not be used. Dealing with these non-normalities is
the subject of future research. Second, Sharpe Ratios do not necessarily control for
risk. That is, the volatility of the strategy may not reflect the true risk. However, our
method also applies to Information Ratios which use residuals from factor models.
Third, it is necessary in the multiple testing framework to take a stand on what
qualifies as the appropriate significance level, e.g. is it 0.10 or 0.05? Fourth, a
choice needs to made on the multiple testing framework. We present results for three
frameworks as well as the average of the methods. Finally, some judgment is needed
specifying the number of tests.
Given choices (3)-(5), it is important to determine the robustness of the haircuts
to changes in these inputs. We provide a program at:
[Link]
that allows the user to vary the key parameters to investigate the impact on the
haircuts. We also provide a program that determines the minimal level of profitability
for a trading strategy to be considered significant.

2
2.1

Method
Independent Tests and Sharpe Ratio

Let rt denote the realized return for an investment strategy between time t 1 and t.
The investment strategy involves zero initial investment so that rt measures the net
gain/loss. Such a strategy can be a long-short strategy, i.e., rt = RtL RtS where RtL
and RtS are the gross investment returns for the long and short position, respectively.
It can also be a traditional stock and bond strategy for which investors borrow and
invest in a risky equity portfolio.
To evaluate if an investment strategy can generate true profits and maintain
those profits in the future, we form a statistical test to see if the expected excess returns are different from zero. Since investors can always switch their positions in the
long-short strategy, we focus on a two-sided alternative hypothesis. In other words, in
so far as the long-short strategy can generate a mean return that is significantly different from zero, we think of it as a profitable strategy. To test this hypothesis, we first
construct key sample statistics. Given a sample of historical returns (r1 , r2 , . . . , rT ),
2

let
denote the mean and
the standard deviation. A t-statistic is constructed to
test the null hypothesis that the average return is zero:
t-ratio =

/ T

(1)

Under the assumption that returns are i.i.d. normal,1 the t-statistic follows a tdistribution with T 1 degrees of freedom under the null hypothesis. We can follow
standard hypothesis testing procedures to assess the statistical significance of the
investment strategy.
The Sharpe Ratio one of the most commonly used summary statistics in finance
is linked to the t-statistic in a simple manner. Given
and
, the Sharpe Ratio
c
(SR) is defined as

c =
,
(2)
SR

which, based on Equation (1), is simply t-ratio/ T .2 Therefore, for a fixed T , a higher
Sharpe Ratio implies a higher t-statistic, which in turn implies a higher significance
level (lower p-value) for the investment strategy. This equivalence between the Sharpe
Ratio and the t-statistic, among many other reasons, justifies the use of Sharpe Ratio
as an appropriate measure of the attractiveness of an investment strategy given our
assumption.

2.2

Sharpe Ratio Adjustment under Multiple Tests

Despite its widespread use, the Sharpe Ratio for a particular investment strategy can
be misleading.3 This is due to the extensive data mining by the finance profession.
Since academics, financial practitioners and individual investors all have a keen interest in finding lucrative investment strategies from the limited historical data, it is
not surprising for them to discover a few strategies that appear to be very profitable. This data mining issue is well recognized by both the finance and the science
literature. In finance, many well-established empirical abnormalities (e.g, certain
technical trading rules, calendar effects, etc.) are overturned once data mining biases
are taken into account.4 Profits from trading strategies that use cross-sectional eq1

Without the normality assumption, the t-statistic becomes asymptotically normally distributed
based on the Central Limit Theorem.
2
Lower frequency Sharpe Ratios can be calculated straightforwardly assuming higher frequency
returns are independent. For instance, if
and
denotethe mean
and volatility of monthly returns,
respectively, then the annual Sharpe Ratio equals 12
/ 12
= 12
/
.
3
It can also be misleading if returns are not i.i.d. (for example, non-normality and/or autocorrelation) or if the volatility does not reflect the risk.
4
See Sullivan, Timmermann and White (1999, 2001) and White (2000).

uity characteristics involve substantial statistical biases.5 The return predictability

of many previously documented variables is shown to be spurious once appropriate
statistical tests are performed.6 In medical research, it is well-known that discoveries
tend to be exaggerated.7 This phenomenon is termed the winners curse in medical
science: the scientist who makes the discovery in a small study is cursed by finding
an inflated effect.
Given the widespread use of the Sharpe Ratio, we provide a probability based
multiple testing framework to adjust the conventional ratio for data mining. To
illustrate the basic idea, we give a simple example in which all tests are assumed
to be independent. This example is closely related to the literature on data mining
biases. However, we are able to generalize important quantities in this example using
a multiple testing framework. This generalization is key to our approach as it allows
us to study the more realistic case when different strategy returns are correlated.
To begin with, we calculate the p-value for the independent test:
pI = P r(|r| > t-ratio)

c T ),
= P r(|r| > SR

(3)

where r denotes a random variable that follows a t-distribution with T 1 degrees of

freedom. This p-value might make sense if researchers are strongly motivated by an
economic theory and directly construct empirical proxies to test the implications of
the theory. It does not make sense if researchers have explored hundreds or even thousands of strategies and only choose to present the most profitable one. In the latter
case, the p-value for the independent test may greatly overstate the true statistical
significance.
To quantitatively evaluate this overstatement, we assume that researchers have
tried N strategies and choose to present the most profitable (largest Sharpe Ratio)
one. Additionally, we assume (for now) that the test statistics for these N strategies
are independent. Under these simplifying assumptions and under the null hypothesis
that none of these strategies can generate non-zero returns, the multiple testing p5

See Leamer (1978), Lo and MacKinlay (1990), Fama (1991), Schwert (2003). A recent paper by
McLean and Pontiff (2013) shows a significant degradation of performance of identified anomalies
after publication.
6
See Welch and Goyal (2004).
7
See Button et al. (2013).

value, pM , for observing a maximal t-statistic that is at least as large as the observed
t-ratio is
pM = P r(max{|ri |, i = 1, . . . , N } > t-ratio)
N
Y
= 1
P r(|ri | t-ratio)
i=1

= 1 (1 pI )N .

(4)

When N = 1 (independent test) and pI = 0.05, pM = 0.05, so there is no multiple

testing adjustment. If N = 10 and we observe a strategy with pI = 0.05, pM = 0.401,
implying a probability of about 40% in finding an investment strategy that generates
a t-statistic that is at least as large as the observed t-ratio, much larger than the
5% probability for independent test. Multiple testing greatly reduces the statistical
significance of independent test. Hence, pM is the adjusted p-value after data mining
is taken into account. It reflects the likelihood of finding a strategy that is at least as
profitable as the observed strategy after searching through N independent strategies.
By equating the p-value of an independent test to pM , we obtain the defining
[
equation for the multiple testing adjusted (haircut) Sharpe Ratio HSR:
[
pM = P r(|r| > HSR

T ).

(5)

c For instance, assuming

[ will be smaller than SR.
Since pM is larger than pI , HSR
there are twenty years of monthly returns (T = 240), an annual Sharpe Ratio of
0.75 yields a p-value of 0.0008 for an independent test. When N = 200, pM = 0.15,
implying an adjusted annual Sharpe Ratio of 0.32 through Equation (5). Hence,
multiple testing with 200 tests reduces the original Sharpe Ratio by approximately
60%.
This simple example illustrates the gist of our approach. When there is multiple
testing, the usual p-value pI for independent test no longer reflects the statistical
significance of the strategy. The multiple testing adjusted p-value pM , on the other
hand, is the more appropriate measure. When the test statistics are dependent,
however, the approach in the example is no longer applicable as pM generally depends
on the joint distribution of the N test statistics. For this more realistic case, we build
on the work of HLZ to provide a multiple testing framework to find the appropriate
p-value adjustment.

Multiple Testing Framework

When more than one hypothesis is tested, false rejections of the null hypotheses
are more likely to occur, i.e., we incorrectly discover a profitable trading strategy.
Multiple testing methods are designed to limit such occurrences. Multiple testing
methods can be broadly divided into two categories: one controls the family-wise
error rate and the other controls the false-discovery rate.8 Following HLZ, we present
three multiple testing procedures.

3.1

Type I Error

We first introduce two definitions of Type I error in a multiple testing framework.

Assume that M hypotheses are tested and their p-values are (p1 , p2 , . . . , pM ). Among
these M hypotheses, R are rejected. These R rejected hypotheses correspond to R
discoveries, including both true discoveries and false discoveries. Let Nr denote the
total number of false discoveries, i.e., strategies incorrectly classified as profitable.
Then the family-wise error rate (FWER) calculates the probability of making at
least one false discovery:
FWER = P r(Nr 1).
Instead of studying the total number of false rejections, i.e., profitable strategies
that turn out to be unprofitable, an alternative definition the false discovery rate
focuses on the proportion of false rejections. Let the false discovery proportion
(FDP) be the proportion of false rejections:

r if R > 0,
R
FDP =

0 if R = 0.
Then the false discovery rate (FDR) is defined as:
FDR = E[F DP ].

Both FWER and FDR are generalizations of the Type I error probability in independent testing. Comparing the two definitions, procedures that control FDR allow
the number of false discoveries to grow proportionally with the total number of tests
8

For the literature on the family-wise error rate, see Holm (1979), Hochberg (1988) and Hommel
(1988). For the literature on the false-discovery rate, see Benjamini and Hochberg (1995), Benjamini
and Liu (1999), Benjamini and Yekutieli (2001), Storey (2003) and Sarkar and Guo (2009).

and are thus more lenient than procedures that control FWER. Essentially, FWER
is designed to prevent even one error. FDR controls the error rate.9

3.2

P-value Adjustment under FWER

We order the p-values in ascending orders, i.e., p(1) p(2) . . . p(M ) and let the
associated null hypotheses be H(1) , H(2) , . . . , H(M ) .
Bonferroni s method10 adjusts each p-value equally. It inflates the original p-value
by the number of tests M :
Bonferroni :

erroni
= min[M p(i) , 1], i = 1, . . . , M.
pBonf
(i)

For example, if we observe M = 10 strategies and one of them has a p-value of 0.05,
Bonferroni would say the more appropriate p-value is M p = 0.50 and hence the strategy is not significant at 5% in this multiple testing framework. For a more concrete
example that we will use throughout this section, suppose we observe M = 4 strategies and the ordered p-value sequence is (0.01, 0.015, 0.03, 0.04). All four strategies
would be deemed significant under independent tests. Bonferroni suggests that the
adjusted p-value sequence is (0.04, 0.06, 0.12, 0.16). Therefore, only the first strategy
is significant under Bonferroni.
Holms method11 relies on the sequence of p-values and adjusts each p-value by:
Holm :

pHolm
= min[max{(M j + 1)p(j) }, 1], i = 1, . . . , M.
(i)
ji

Starting from the smallest p-value, Holms method allows us to sequentially build
up the adjusted p-value sequence. Using the previous example, the Holm adjusted
p-value for the first strategy is pHolm
= 4p(1) = 0.04, which is identical to the level
(1)
prescribed by Bonferroni. Under 5% significance, this strategy is significant. The
second strategy yields pHolm
= max[4p(1) , 3p(2) ] = 3p(2) = 0.045, which is smaller
(2)
than the Bonferroni implied p-value. Given a cutoff of 5% and different from what
Bonferroni concludes, this strategy is significant. Similarly, the last two adjusted
p-values are calculated as pHolm
= max[4p(1) , 3p(2) , 2p(3) ] = 2p(3) = 0.06, pHolm
=
(3)
(4)
max[4p(1) , 3p(2) , 2p(3) , p(4) ] = 2p(3) = 0.06, making neither significant at 5% level.
Therefore, the first two strategies are found to be significant under Holm.
9

For more details on FWER and FDR, see HLZ.

For the statistical literature on Bonferronis method, see Schweder and Spjotvoll (1982) and
Hochberg and Benjamini (1990). For the applications of Bonferronis method in finance, see Shanken
(1990), Ferson and Harvey (1999), Boudoukh et al. (2007) and Patton and Timmermann (2010).
11
For the literature on Holms procedure and its extensions, see Holm (1979) and Hochberg
(1988). Holland, Basu and Sun (2010) emphasize the importance of Holms method in accounting
research.
10

Comparing the multiple testing adjusted p-values to a given significance level,

we can make a statistical inference for each of these hypotheses. If we made the
mistake of assuming independent tests, and given a 5% significance level, we would
discover four factors. In multiple testing, both Bonferronis and Holms adjustment
guarantee that the family-wise error rate (FWER) in making such inferences does
not exceed the pre-specified significance level. Comparing these two adjustments,
erroni
pHolm
pBonf
for any i.12 Therefore, Bonferronis method is tougher because it
(i)
(i)
inflates the original p-values more than Holms method. Consequently, the adjusted
Sharpe Ratios under Bonferroni will be smaller than those under Holm. Importantly,
both of these procedures are designed to eliminate all false discoveries no matter
how many tests for a given significance level. While this type of approach seems
appropriate for a space mission (parts failures), asset managers may be willing to
accept the fact that the number of false discoveries will increase with the number of
tests.

3.3

P-value Adjustment under FDR

Benjamini, Hochberg and Yekutieli (BHY )s procedure13 defines the adjusted p-values
sequentially:

BHY :

pBHY
(i)

p(M )

if i = M,

M c(M )
p(i) ] if i M 1,
min[pBHY
(i+1) ,
i

P
1
where c(M ) = M
j=1 j . In contrast to Holms method, BHY starts from the largest pvalue and defines the adjusted p-value sequence through pairwise comparisons. Again
using the previous example, we first calculate the normalizing constant as c(M ) =
P
4
1
j=1 j = 2.08. To assess the significance of the four strategies, we start from the
least significant one. BHY sets pBHY
at 0.04, the same as the original value of
(4)
42.08
p(4) . For the third strategy, BHY yields pBHY
= min[pBHY
=
p(3) ] = pBHY
(3)
(4) ,
(4)
3
BHY
0.04. Similarly, the first two adjusted p-values are sequentially calculated as p(2) =
42.08
42.08
min[pBHY
p(2) ] = pBHY
= 0.04 and pBHY
= min[pBHY
p(1) ] = pBHY
=
(3) ,
(3)
(1)
(2) ,
(2)
2
1
0.04. Therefore, the BHY adjusted p-value sequence is (0.04, 0.04, 0.04, 0.04), making
all four strategies significant at 5% level. Based on our example, BHY leads to two
more discoveries compared to Holm and Holm leads to one more discovery compared
to Bonferroni.
12

See Holm (1979) for the proof.

For the statistical literature on BHYs method, see Benjamini and Hochberg (1995), Benjamini
and Yekutieli (2001), Sarkar (2002) and Storey (2003). For the applications of methods that control
the false discovery rate in finance, see Barras, Scaillet and Wermers (2010), Bajgrowicz and Scaillet
(2012) and Kosowski, Timmermann, White and Wermers (2006).
13

Hypothesis tests based on the adjusted p-values guarantee that the false discovery
rate (FDR) does not exceed the pre-specified significance level. The constant c(M )
controls the generality of the test. In the original work by Benjamini and Hochberg
(1995), c(M ) is set equal to one and the test works when p-values are independent
or positively dependent. With our choice of c(M ), the test works under arbitrary
dependence structure for the test statistics.
The three multiple testing procedures provide adjusted p-values that control for
data mining. Based on these p-values, we transform the corresponding t-ratios into
Sharpe Ratios. In essence, our Sharpe Ratio adjustment method aims to answer
the following question: if the multiple testing adjusted p-value reflects the genuine
statistical significance for an investment strategy, what is the equivalent single test
Sharpe Ratio that one should assign to such a strategy as if there were no data
mining?
For both Holm and BHY, we need the empirical distribution of p-values for strategies that have been tried so far. We use the structural model estimate from HLZ.
The model is based on the performance data for more than 300 risk factors that have
been documented by the academic literature. However, a direct multiple testing adjustment based on these data is problematic for two reasons. First, we do not observe
all the strategies that have been tried. Indeed, thousands more could have been tried
and ignoring these would significantly affect our results on the haircut Sharpe Ratio.
Second, strategy returns are correlated. Correlation affects multiple testing in that
it effectively reduces the number of independent tests. Taking these two concerns
into account, HLZ proposes a new method to study multiple hypothesis testing in
financial economics and we follow this method.

3.4

Multiple Testing and Cross-validation

Recent important papers by Lopez de Prado and his coauthors also consider the
ex-post data mining issue for standard backtests.14 Due to data mining, they show
theoretically that only seven trials are needed to obtain a spurious two-year long
backtest that has an in-sample realized Sharpe Ratio over 1.0 while the expected
out of sample Sharpe Ratio is zero. The phenomenon is analogous to the regression
overfitting problem when models found to be superior in in-sample test often perform
poorly out-of-sample and is thus termed backtest overfitting. To quantify the degreee
of backtest overfitting, they propose the calculation of the probability of backtest
overfitting (PBO) that measures the relative performance of a particular backtest
among a basket of strategies using cross-validation techniques.
Their research shares a common theme with our study. We both attempt to
evaluate the performance of an investment strategy in relation to other available
14

See Bailey et al. (2013a,b) and Lopez de Prado (2013).

strategies. Their method computes the chance for a particular strategy to outperform
the median of the pool of alternative strategies. In contrast, our work adjusts the
statistical significance for each individual strategy so that the overall proportion of
spurious strategies is controlled.
Despite these similar themes, our research is different in many ways. First, the
objectives of analyses are different. Our research focuses on identifying the group of
strategies that generate non-zero returns while Lopez de Prado evaluates the relative
performance of a certain strategy that is fit in-sample. For example, consider a
case when where there are a group of factors that are all true. The one with the
smallest t-ratio, although dominated by other factors in terms of t-ratios, may still
be declared significant in our multiple testing framework. In contrast, it will rarely
be considered in the PBO framework as it is dominated by other more significant
strategies. Second, our method is based on a single test statistic that summarizes a
strategys performance over the entire sample whereas their method divides and joins
the entire sample in numerous ways, each way corresponding to an artificial holdout periods. Our method is therefore more in line with the statistics literature
on multiple testing while their work is more related to out-of-sample testing and
cross-validation. Third, the extended statistical framework in Harvey and Liu (2014)
needs only test statistics. In contrast, their work relies heavily on the time-series of
each individual strategy. While data intensive, in the Lopez de Prado approach, it
is not necessary to make assumptions regarding the data generating process for the
returns. As such, their approach is closer to the machine learning literature and ours
is closer to the econometrics literature. Finally, the PBO method assesses whether
a strategy selection process is prone to overfitting. It is not linked to any particular
performance statistics. We primarily focus on Sharpe Ratios as they are directly
linked to t-statistics and thus p-values, which are the required inputs for multiple
testing adjustment. Our framework can be easily generalized to incorporate other
performance statistics as long as they also have probabilistic interpretations.

3.5

In-sample Multiple Testing vs. Out-of-sample Validation

Our multiple testing adjustment is based on in-sample (IS) backtests. In practice,

out-of-sample (OOS) tests are routinely used to select among many strategies.
Despite its popularity, OOS testing has several limitations. First, an OOS test
may not be truly out-of-sample. A researcher tries a strategy. After running an
OOS test, she finds that the strategy fails. She then revises the strategy and tries
again, hoping it would work this time. This trial and error approach is not truly
OOS, but it is hard for outsiders to tell. Second, an OOS test, like any other test
in statistics, only works in a probabilistic sense. In other words, a success for an
OOS test can be due to luck for both the in-sample selection and the out-of-sample

10

testing. Third, given the researcher has experienced the data, there is no true OOS.15
This is especially the case when the trading strategy involves economic variables. No
matter how you construct the OOS test, it is not truly OOS because you know what
happened in the data.
Another important issue with the OOS method, which our multiple testing procedure can potentially help solve, is the tradeoff between Type I (false discoveries)
and Type II (missed discoveries) errors due to data splitting.16 In holding some data
out, researchers increase the chance of missing true discoveries for the shortened insample data. For instance, suppose we have 1,000 observations. Splitting the sample
in half and estimating 100 different strategies in-sample, i.e., based on 500 observations, suppose we identify 10 strategies that look promising (in-sample tests). We
then take these 10 strategies to the OOS tests and find that two strategies work.
Note that, in this process, we might have missed, say, three strategies after the first
step IS tests due to bad luck in the short IS period. These true discoveries are lost
because they never get to the second step OOS tests.
Instead of the 50-50 split, now suppose we use a 90-10 data split. Suppose we
identify 15 promising strategies. Among the strategies are two of the three true
discoveries that we missed when we had a shorter in-sample period. While this is
good, unfortunately, we have only 100 observations held out for the OOS exercise
and it will be difficult to separate the good from the bad. At its core, the OOS
exercise faces a tradeoff between in-sample and out-of-sample testing power. While a
longer in-sample period leads to a more powerful test and this reduces the chance of
committing a Type II error (i.e., missing true discoveries), the shorter out-of-sample
period provides too little information to truly discriminate among the factors that
are found significant in-sample.
So how does our research fit? First, one should be very cautious of OOS tests
because it is hard to construct a true OOS test. The alternative is to apply our
multiple testing framework to identify the true discoveries on the full data. This
would involve making a more stringent cutoff for test statistics.
Another, and in our opinion, more promising framework, is to merge the two
methods. Ideally, we want the strategies to pass both the OOS test on split data
and the multiple test on the entire data. The problem is how to deal with the true
discoveries that are missed if the in-sample data is too short. As a tentative solution,
we can first run the IS tests with a lenient cutoff (e.g., p-value = 0.2) and use the
OOS tests to see which strategy survives. At the same time, we can run multiple
testing for the full data. We then combine the IS/OOS test and the multiple test
by looking at the intersection of survivors. We leave the details of this approach to
future research.
15
16

See L
opez de Prado (2013) for a similar argument.
See Hansen and Timmermann (2012) for a discussion on sample splitting for univariate tests.

11

Applications

4.1

Three Strategies

To illustrate how the Sharpe Ratio adjustment works, we begin with three investment
strategies that have appeared in the literature. All of these strategies are zero cost
hedging portfolios that simultaneously take long and short positions of the crosssection of the U.S. equities. The strategies are: the earnings-to-price ratio (E/P),
momentum (MOM) and the betting-against-beta factor (BAB, Frazzini and Pedersen
(2013)). These strategies cover three distinct types of investment styles (i.e., value
(E/P), trend following (MOM) and potential distortions induced by leverage (BAB))
and generate a range of Sharpe Ratios.17 None of these strategies reflect transaction
costs and as such the Sharpe Ratios are overstated and should be considered before
costs Sharpe Ratios.
Two important ingredients to the Sharpe Ratio adjustment are the initial value
of the Sharpe Ratio and the number of trials. To highlight the impact of these two
inputs, we focus on the simplest independent case as in Section 2. In this case,
the multiple testing p-value pM and the independent testing p-value pI are linked
through Equation (4). When pI is small, this relation is approximately the same as
in Bonferronis adjustment. Hence, the multiple testing adjustment we use for this
example can be thought of as a special case of Bonferronis adjustment.
Table 1 shows the summary statistics for these strategies. Among these strategies,
the strategy based on E/P is least profitable as measured by the Sharpe Ratio. It has
an average monthly return of 0.43% and a monthly standard
deviation of 3.47%. The

corresponding annual Sharpe Ratio is 0.43(= (0.43% 12)/3.47%). The p-value for
independent test is 0.003, comfortably exceeding a 5% benchmark. However, when
multiple testing is taken into account and assuming that there are ten trials, the
b which captures the
multiple testing p-value increases to 0.029. The haircut (hc),
percentage change in the Sharpe Ratio, is about 27%. When there are more trials,
the haircut is even larger.
Sharpe Ratio adjustment depends on the initial value of the Sharpe Ratio. Across
the three investment strategies, the Sharpe Ratio ranges from 0.43 (E/P ) to 0.78
(BAB). The haircut is not uniform across different initial Sharpe Ratio levels. For
instance, when the number of trials is 50, the haircut is almost 50% for the least prof17

For E/P , we construct an investment strategy that takes a long position in the top decile
(highest E/P ) and a short position in the bottom decile (lowest E/P ) of the cross-section of E/P
sorted portfolios. For M OM , we construct an investment strategy that takes a long position in the
top decile (past winners) and a short position in the bottom decile (past losers) of the cross-section
of portfolios sorted by past returns. Both the data for E/P and M OM are obtained from Ken
Frenchs on-line data library for the period from July 1963 to December 2012. For BAB, return
statistics are extracted from Table IV of Frazzini and Pedersen (2013).

12

Table 1: Multiple Testing Adjustment for Three Investment Strategies

Summary statistics for three investment strategies: E/P , M OM and BAB (betting-against-beta,
Frazzini and Pedersen (2013)). Mean and Std. report the monthly mean and standard deviation
c reports the annualized Sharpe Ratio; t-stat reports the t-statistic for
of returns, respectively; SR
p
c T /12);
the independent hypothesis test that the mean strategy return is zero (t-stat = SR
\ reports the
pI and pM report the p-value for independent and multiple test, respectively; HSR
c
Bonferroni adjusted Sharpe Ratio; hc reports the percentage haircut for the adjusted Sharpe Ratio
c = (SR
c HSR)/
c
\ SR).
(hc
Strategy

Mean(%)
(monthly)

Std.(%)
(monthly)

c
SR
(annual)

t-stat

pI
(individual)

pM
(multiple)

\
HSR
(annual)

c
hc
(haircut)

3.47
7.03
3.09

0.43
0.67
0.78

2.99
4.70
7.29

2.88 103
3.20 106
6.29 1013

2.85 102
3.20 105
6.29 1012

0.31
0.60
0.74

26.6%
10.9%
4.6%

3.47
7.03
3.09

0.43
0.67
0.78

2.99
4.70
7.29

2.88 103
3.20 106
6.29 1013

1.35 101
1.60 105
3.14 1011

0.21
0.54
0.72

50.0%
19.2%
7.9%

3.47
7.03
3.09

0.43
0.67
0.78

2.99
4.70
7.29

2.88 103
3.20 106
6.29 1013

2.51 101
1.60 105
6.29 1011

0.16
0.51
0.71

61.6%
23.0%
9.3%

Panel A: N = 10
E/P
M OM
BAB

0.43
1.36
0.70

Panel B: N = 50
E/P
M OM
BAB

0.43
1.36
0.70

Panel C: N = 100
E/P
M OM
BAB

0.43
1.36
0.70

itable E/P strategy but only 7.9% for the most profitable BAB strategy.18 We believe
this non-uniform feature of our Sharpe Ratio adjustment procedure is economically
sensible since it allows us to discount mediocre Sharpe Ratios harshly while keeping
the exceptional ones relatively intact.

4.2

Sharpe Ratio Adjustment for a New Strategy

Given the population of investment strategies that have been published, we now show
how to adjust the Sharpe Ratio of a new investment strategy. Consider a new strategy
c in T periods,19 or, equivalently, the p-value pI .
that generates a Sharpe Ratio of SR
Assuming that N other strategies have been tried, we draw N t-statistics from the
model in HLZ.20 These N + 1 p-values are then adjusted using the aforementioned
three multiple testing procedures. In particular, we obtain the adjusted p-value pM for
18

Mathematically, this happens because the p-value is very sensitive to the t-statistic when the
t-statistic is large. In our example, when N = 50 and for BAB, the p-value for a t-statistic of
7.29 (independent test) is one 50th of the p-value for a t-statistic of 6.64 (multiple testing adjusted
t-statistic), i.e., pM /pI 50.
p
19
c is an annualized Sharpe Ratio, t-stat = SR
c T /12; if SR
c is
Assuming T is in months, if SR

c
a monthly Sharpe Ratio, t-stat = SR T .
20
For both this section on haircut Sharpe ratios and the next section on return hurdles, we set
the average correlation among strategy returns at 0.2 the preferred estimate in HLZ. However,

13

pI . To take the uncertainty in drawing N t-statistics into account, we repeat the above
procedure many times to generate a sample of pM s. The median of this sample is
taken as the final multiple testing adjusted p-value. This p-value is then transformed
back into a Sharpe Ratio the multiple testing adjusted Sharpe Ratio. Figure 2
shows the original vs. Haircut Sharpe Ratios and Figure 3 shows the corresponding
haircut.
First, as previously discussed, the haircuts depend on the levels of the Sharpe
Ratios. Across the three types of multiple testing adjustment and different numbers
of tests, the haircut is almost always above and sometimes much larger than 50% when
the annualized Sharpe Ratio is under 0.4. On the other hand, when the Sharpe Ratio
is greater than 1.0, the haircut is at most 25%. This shows the 50% rule of thumb
discount for the Sharpe Ratio is inappropriate: 50% is too lenient for relatively small
Sharpe Ratios (< 0.4) and too harsh for large ones (> 1.0). This nonlinear feature of
the Sharpe Ratio adjustment makes economic sense. Marginal strategies are heavily
penalized because they are likely false discoveries.
Second, the three adjustment methods imply different magnitudes of haircuts.
Given the theoretical objectives that these methods try to control (i.e., family-wise
error rate (FWER) vs false discovery rate (FDR)), we should divide the three adjustments into two groups: Bonferroni and Holm as one group and BHY as the other
group. Comparing Bonferroni and Holms method, we see that Holms method implies a smaller haircut than Bonferronis method. This is consistent with our previous
discussion on Holms adjustment being less aggressive than Bonferronis adjustment.
However, the difference is relatively small (compared to the difference between Bonferroni and BHY), especially when the number of tests is large. The haircuts under
BHY, on the other hand, are usually a lot smaller than those under Bonferroni and
Holm when the Sharpe Ratio is small (< 0.4). For large Sharpe Ratios (> 1.0),
however, the haircuts under BHY are consistent with those under Bonferroni and
Holm.
In the end, we would advocate the BHY method. The FWER seems appropriate
for applications where there is a severe consequence of a false discovery. In financial
applications, it seems reasonable to control for the rate of false discovery rather than
the absolute number.

the programs that we provide in the appendix allow the user to specify this number. With a user
specified level of correlation, we linearly interpolate among the five sets of parameter estimates in
HLZ to find a new set of parameter estimates that exactly achieves the assumed correlation level.

14

Haircut Sharpe Ratio

1.0
0.8
0.6
0.4
0.2
0

Haircut Sharpe Ratio

1.0
0.8
0.6
0.4
0.2
0

Haircut Sharpe Ratio

Figure 1: Original vs. Haircut Sharpe Ratios

1.0
0.8
0.6
0.4
0.2
0

Bonferroni
Holm
BHY

Number of tests = 10
No adjustment

50% adjustment
0

0.2

0.4

0.6

0.8

1.0

0.8

1.0

0.4
0.6
0.8
Original Sharpe Ratio (annualized)

1.0

Number of tests = 50

0.2

0.4

0.6

Number of tests = 100

0.2

15

Figure 2: Haircuts

Haircut

Number of tests = 10
100%
80%
60%
40%
20%
0
0.2

50% adjustment

Bonferroni
Holm
BHY

0.4

0.6

0.8

1.0

0.8

1.0

0.6
0.8
Original Sharpe Ratio (annualized)

1.0

Haircut

Number of tests = 50
100%
80%
60%
40%
20%
0
0.2

0.4

0.6

Haircut

Number of tests = 100

100%
80%
60%
40%
20%
0
0.2

0.4

16

4.3

Minimum Profitability for Proposed Trading Strategies

There is another way to pose the problem. Given an agreed upon level of significance,
such as 0.05, what is the minimum average monthly return that a proposed strategy
needs to exceed? Our framework is ideally suited to answer this question.
The answer to the question depends on a number of inputs. We need to measure
the volatility of the strategy. The number of observations is also a critical input.
Finally, we need to take a stand on the number of tests that have been conducted.
Table 2 presents an example. Here we consider four different sample sizes: 120,
240, 480 and 1,000 and three different levels of annualized volatility: 5%, 10% and
15%. We then assume the total number of tests is 300. To generate the table, we first
find threshold t-ratios based on the multiple testing adjustment methods provided in
the previous section and then transform these t-ratios into mean returns based on the
formula in Equation (1).
Table 2: Minimum Profitability Hurdles
Average monthly return hurdles under independent and
multiple tests. At 5% significance, the table shows the minimum average monthly return for a strategy to be significant
at 5% with 300 tests. All numbers are in percentage terms.
See Appendix for link to the program.

= 5%

Annualized volatility
= 10%
= 15%

Panel A: Observations = 120

Independent
Bonferroni
Holm
BHY

0.258
0.496
0.486
0.435

0.516
0.992
0.972
0.871

0.775
1.488
1.459
1.305

Panel B: Observations = 240

Independent
Bonferroni
Holm
BHY

0.183
0.351
0.344
0.307

0.365
0.702
0.688
0.616

0.548
1.052
1.031
0.923

Panel C: Observations = 480

Independent
Bonferroni
Holm
BHY

0.129
0.248
0.243
0.217

0.258
0.496
0.486
0.435

0.387
0.744
0.729
0.651

Panel D: Observations = 1000

Independent
Bonferroni
Holm
BHY

0.089
0.172
0.169
0.151

0.179
0.344
0.337
0.302

0.268
0.516
0.505
0.452

Table 2 shows the large differences between the return hurdles for independent
testing and multiple testing. For example, in Panel B (240 observations) and 10%,
17

the minimum required average monthly return for an independent test is 0.365% per
month or 4.4% annually. However, for BHY, the return hurdle is much higher, 0.616%
per month or 7.4% on an annual basis. Appendix A.2 details the program that we
use to generate these return hurdles and provides an Internet address to download
the program.

Conclusions

We provide a real time evaluation method for determining the significance of a candidate trading strategy. Our method explicitly takes into account that hundreds if
not thousands of strategies have been proposed and tested in the past. Given these
multiple tests, inference needs to be recalibrated.
Our method follows the following steps. First, we transform the Sharpe Ratio
into a t-ratio and determine its probability value. e.g., 0.05. Second, we determine
what the appropriate p-value should be explicitly recognizing the multiple tests that
preceded the discovery of this particular investment strategy. Third, based on this
new p-value, we transform the corresponding t-ratio back to a Sharpe Ratio. The
new measure which we call the Haircut Sharpe Ratio takes the multiple testing or
data mining into a account. Our method is readily applied to popular risk metrics,
like Value at Risk (VaR).21
Our method is ideally suited to determine minimum profitability hurdles for proposed strategies. We provide a computer program where the inputs are the desired
level of significance, the number of observations, the strategy volatility as well as the
assumed number of tests. The output is the minimum average monthly return that
the proposed strategy needs to exceed.
There are many caveats to our method. We do not observe the entire history of
tests and, as such, we need to use judgement on an important imput the number of
tests for our method. In addition, we use Sharpe Ratios as our starting point. Our
method is not applicable insofar as the Sharpe Ratio is not the appropriate measure
(e.g., non-linearities in trading strategy or the variance not being a complete measure
of risk).
Of course, true out-of-sample test of a particular strategy (not a holdout sample)
is a cleaner way to evaluate the viability of a strategy. For some strategies, models can
be tested on new (previously unpublished) data or even on different (uncorrelated)
markets. However, for the majority of trading strategies, true out of sample tests
21

Let V aR() of a return series to be the -th percentile of the return distribution. Assuming
that returns are approximately normally distributed, it can be shown that V aR is related to Sharpe
Ratio by V aR()
= SR z , where z is the z-score for the (1 )-th percentile of a standard

normal distribution and is the standard deviation of the return. Multiple testing adjusted Sharpe
Ratios can then be used to adjust VaRs.

18

are not available. Our method allows for decision to be made, in real time, on the
viability of a proposed strategy.

19

References
Bajgrowicz, Pierre and Oliver Scaillet, 2012, Technical trading revisited: False discoveries, persistence tests, and transaction costs, Journal of Financial Economics
106, 473-491.
Barras, Laurent, Oliver Scaillet and Russ Wermers, 2010, False discoveries in mutual
fund performance: Measuring luck in estimated alphas, Journal of Finance 65, 179216.
Benjamini, Yoav and Daniel Yekutieli, 2001, The control of the false discovery rate
in multiple testing under dependency, Annals of Statistics 29, 1165-1188.
Benjamini, Yoav and Wei Liu, 1999, A step-down multiple hypotheses testing procedure that controls the false discovery rate under independence, Journal of Statistical
Planning and Inference 82, 163-170.
Benjamini, Yoav and Yosef Hochberg, 1995, Controlling the false discovery rate: A
practical and powerful approach to multiple testing, Journal of the Royal Statistical
Socitey. Series B 57, 289-300.
Boudoukh, Jacob, Roni Michaely, Matthew Richardson and Michael R. Roberts,
2007, On the importance of measuring payout yield: implications for empirical asset
pricing, Journal of Finance 62, 877-915.
Button, Katherine, John Ioannidis, Brian Nosek, Jonathan Flint, Emma Robinson
and Marcus Munaf`o, 2013, Power failure: why small sample size undermines the
reliability of neuroscience, Nature Reviews Neuroscience 14, 365-376.
Bailey, David, Jonathan Borwein, Marcos Lopez de Prado and Qiji Jim Zhu, 2013a,
Pseudo-mathematics and financial charlatanism: The effects of backtest overfitting
on out-of-sample, Working Paper, Lawrence Berkeley National Laboratory.
Bailey, David, Jonathan Borwein, Marcos Lopez de Prado and Qiji Jim Zhu, 2013b,
The probability of back-test overfitting, Working Paper, Lawrence Berkeley National
Laboratory.
Fama, Eugene F., 1991, Efficient capital markets: II, Journal of Finance 46, 15751617.
Fama, Eugene F. and Kenneth R. French, 1992, The cross-section of expected stock
returns, Journal of Finance 47, 427-465.
Ferson, Wayne E. and Campbell R. Harvey, 1999, Conditioning variables and the
cross section of stock returns, Journal of Finance 54, 1325-1360.
Frazzini, Andrea and Lasse Heje Pedersen, 2013, Betting against beta, Working
Paper, AQR Capital Management.
20

Hansen, Peter Reinhard and Allan Timmermann, 2012, Choice of sample split in
out-of-sample forecast evaluation, Working Paper, Stanford University.
Harvey, Campbell R., Yan Liu and Heqing Zhu, 2014, . . . and the crosssection of expected returns, Working Paper, Duke University. Available at
[Link] id=2249314.
Harvey, Campbell R. and Yan Liu, 2014, Multiple Testing in Economics, Available
at [Link] id=2358214
Hochberg, Yosef, 1988, A sharper Bonferroni procedure for multiple tests of significance, Biometrika 75, 800-802.
Hochberg, Yosef and Benjamini, Y., 1990, More powerful procedures for multiple
significance testing, Statistics in Medicine 9, 811-818.
Hochberg, Yosef and Tamhane, Ajit, 1987, Multiple comparison procedures, John
Wiley & Sons.
Holland, Burt, Sudipta Basu and Fang Sun, 2010, Neglect of multiplicity when
testing families of related hypotheses, Working Paper, Temple University.
Holm, Sture, 1979, A simple sequentially rejective multiple test procedure, Scandinavian Journal of Statistics 6, 65-70.
Hommel, G., 1988, A stagewise rejective multiple test procedure based on a modified
Bonferroni test, Biometrika 75, 383-386.
Kosowski, Robert, Allan Timmermann, Russ Wermers and Hal White, 2006, Can
mutual fund stars really pick stocks? New evidence from a Bootstrap analysis,
Journal of Finance 61, 2551-2595.
Leamer, Edward E., 1978, Specification searches: Ad hoc inference with nonexperimental data, New York: John Wiley & Sons.
Lo, Andrew W., 2002, The statistics of Sharpe Ratios, Financial Analysts Journal
58, 36-52.
Lo, Andrew W. and Jiang Wang, 2006, Trading volume: Implications of an intertemporal capital asset pricing model, Journal of Finance 61, 2805-2840.
Lopez de Prado, Marcos, 2013, What to look for in a backtest, Working Paper,
Lawrence Berkeley National Laboratory.
McLean, R. David and Jeffrey Pontiff, 2013, Does academic research destroy stock
return predictability? Working Paper, University of Alberta.
Patton, Andrew J. and Allan Timmermann, 2010, Monotonicity in asset returns:
New tests with applications to the term structure, the CAPM, and portfolio sorts,
Journal of Financial Economics 98, 605-625.
21

Sarkar, Sanat K. and Wenge Guo, 2009, On a generalized false discovery rate, The
Annals of Statistics 37, 1545-1565.
Schweder, T. and E. Spjotvoll, 1982, Plots of p-values to evaluate many tests simultaneously, Biometrika 69, 439-502.
Schwert, G. William, 2003, Anomalies and market efficiency, Handbook of the Economics of Finance, edited by G.M. Constantinides, M. Haris and R. Stulz, Elsevier
Science B.V., 939-974.
Shanken, Jay, 1990, Intertemporal asset pricing: An empirical investigation, Journal
of Econometrics 45, 99-120.
Storey, John D., 2003, The positive false discovery ratee: A Bayesian interpretation
and the q-value, The Annals of Statistics 31, 2013-2035.
Sullivan, Ryan, Allan Timmermann and Halbert White, 1999, Data-snooping, technical trading rule performance, and the Bootstrap, Journal of Finance 54, 16471691.
Sullivan, Ryan, Allan Timmermann and Halbert White, 2001, Dangers of data mining: The case of calender effects in stock returns, Journal of Econometrics 105,
249-286.
Welch, Ivo and Amit Goyal, 2008, A comprehensive look at the empirical performance of equity premium prediction, Review of Financial Studies 21, 1455-1508.
White, Halbert, 2000, A reality check for data snooping, Econometrica 68, 10971126.

22

Programs

We make the code and data for our calculations publicly available at:
[Link]

A.1

Haircut Sharpe Ratios

The Matlab function Haircut SR allows the user to specify key parameters to make
Sharpe Ratio adjustments and calculate the corresponding haircuts. It has eight
inputs that provide summary statistics for a return series of an investment strategy
and the number of tests that are allowed for. The first input is the sampling frequency
for the return series. Five options (daily, weekly, monthly, quarterly and annually) are
available.22 The second input is the number of observations in terms of the sampling
frequency provided in the first step. The third input is the Sharpe Ratio of the returns.
It can either be annualized or based on the sampling frequency provided in the first
step; it can also be autocorrelation corrected or not. Subsequently, the fourth input
asks if the Sharpe Ratio is annualized and the fifth input asks if the Sharpe Ratio
has been corrected for autocorrelation.23 The sixth input asks for the autocorrelation
of the returns if the Sharpe Ratio has not been corrected for autocorrelation.24 The
seventh input is the number of tests that are assumed. Lastly, the eighth input is the
assumed average level of correlation among strategy returns.
To give an example of how the program works, suppose that we have an investment strategy that generates an annualized Sharpe Ratio of 1.0 over 120 months.
The Sharpe Ratio is not autocorrelation corrected and the monthly autocorrelation
coefficient is 0.1. We allow for 100 tests in multiple testing and assume the average
22
We use number one, two, three, four and five to indicate daily, weekly, monthly, quarterly and
annually sampled returns, respectively.
23
For the fourth input, 1 denotes a Sharpe Ratio that is annualized and 0 denotes otherwise.
For the fifth input, 1 denotes a Sharpe Ratio that is not autocorrelation corrected and 0 denotes
otherwise.
24
We follow Lo (2002) to adjust Sharpe Ratios for autocorrelations.

23

level of correlation is 0.4 among strategy returns. With this information, the input
vector for the program is

Input vector =

D/W/M/Q/A=1,2,3,4,5
# of obs
Sharpe Ratio
SR annualized? (1=Yes)
AC correction needed? (1=Yes)
AC level
# of tests assumed
Average correlation assumed

3
120
1
1
1
0.1
100
0.4

Passing this input vector to Haircut SR, the function generates a sequence of outputs,
as shown in Figures A.1. The program summarizes the return characteristics by
showing an annualized, autocorrelation corrected Sharpe Ratio of 0.912 as well as
the other data provided by the user. The program output includes adjusted p-values,
haircut Sharpe Ratios and the haircuts involved for these adjustments under a variety
of adjustment methods. For instance, under BHY, the adjusted annualized Sharpe
Ratio is 0.444 and the associated haircut is 51.3%.

24

Figure A.1: Program Outputs

Inputs:
Frequency = Monthly;
Number of Observations = 120;
Initial Sharpe Ratio = 1.000;
Sharpe Ratio Annualized = Yes;
Autocorrelation = 0.100;
A/C Corrected Annualized Sharpe Ratio = 0.912
Assumed Number of Tests = 100;
Assumed Average Correlation = 0.400.
Outputs:
Bonferroni Adjustment:
Adjusted P-value = 0.465;
Haircut Sharpe Ratio = 0.232;
Percentage Haircut = 74.6%.
Holm Adjustment:
Adjusted P-value = 0.409;
Haircut Sharpe Ratio = 0.262;
Percentage Haircut = 71.3%.
BHY Adjustment:
Adjusted P-value = 0.169;
Haircut Sharpe Ratio = 0.438;
Percentage Haircut = 52.0%.
Average Adjustment:
Adjusted P-value = 0.348;
Haircut Sharpe Ratio = 0.298;
Percentage Haircut = 67.3%.

25

A.2

Profit Hurdles

The Matlab function Profit Hurdle allows the user to calculate the required mean
return for a strategy at a given level of significance. It has five inputs. The first input
is the user specified significance level. The second input is the number of monthly
observations for the strategy. The third input is the annualized return volatility of
the strategy. The fourth input is the number of tests that are assumed. Lastly, the
fifth input is the assumed average level of correlation among strategy returns. The
program does not allow for any autocorrelation in the strategy returns.
To give an example of how the program works, suppose we are interested in the
required return for a strategy that covers 20 years and has an annual volatility of
10%. In addition, we allow for 300 tests and specify the significance level to be 5%.
Finally, we assume that the average correlation among strategy returns is 0.4. With
these specifications, the input vector for the program is

Input vector =

Significance level
# of obs
Annualized return volatility
# of tests assumed
Average correlation assumed

0.05
240
0.1
300
0.4

Passing the input vector to Profit Hurdle, the function generates a sequence of
outputs, as shown in Figure A.2. The program summarizes the data provided by
the user. The program output includes return hurdles for a variety of adjustment
methods. For instance, the adjusted return hurdle under BHY is 0.621% per month
and the average multiple testing return hurdle is 0.670% per month.

26

Figure A.2: Program Outputs

Inputs:
Significance Level = 5.0%;
Number of Observations = 240;
Annualized Return Volatility = 10.0%;
Assumed Number of Tests = 300;
Assumed Average Correlation = 0.400.
Outputs:
Minimum Average Monthly Return:
Independent = 0.365%;
Bonferroni = 0.702%;
Holm = 0.686%;
BHY = 0.621%;
Average for Multiple Tests = 0.670%.

27