Pseudo-Mathematics and Financial

Charlatanism: The Effects of

Backtest Overfitting on
Out-of-Sample Performance
David H. Bailey, Jonathan M. Borwein,
Marcos Lpez de Prado, and Qiji Jim Zhu
Another thing I must point out is that you cannot
prove a vague theory wrong. [] Also, if the process
of computing the consequences is indefinite, then
with a little skill any experimental result can be
made to look like the expected consequences.
Richard Feynman [1964]

Introduction
A backtest is a historical simulation of an algorithmic investment strategy. Among other things,
it computes the series of profits and losses that
such strategy would have generated had that algorithm been run over that time period. Popular
performance statistics, such as the Sharpe ratio
or the Information ratio, are used to quantify the
backtested strategys return on risk. Investors
typically study those backtest statistics and then
allocate capital to the best performing scheme.
Regarding the measured performance of a
backtested strategy, we have to distinguish between
two very different readings: in-sample (IS) and outof-sample (OOS). The IS performance is the one
simulated over the sample used in the design of
the strategy (also known as learning period or
David H. Bailey is retired from Lawrence Berkeley National
Laboratory. He is a Research Fellow at the University of California, Davis, Department of Computer Science. His email
address is [email protected].
Jonathan M. Borwein is Laureate Professor of Mathematics
at the University of Newcastle, Australia, and a Fellow of the
Royal Society of Canada, the Australian Academy of Science,
and the AAAS. His email address is jonathan.borwein@
newcastle.edu.au.
Marcos Lpez de Prado is Senior Managing Director at
Guggenheim Partners, New York, and Research Affiliate at
Lawrence Berkeley National Laboratory. His email address
is [email protected].
Qiji Jim Zhu is Professor of Mathematics at Western Michigan University. His email address is [email protected].
DOI: http://dx.doi.org/10.1090/noti1105

458

training set in the machine-learning literature).

The OOS performance is simulated over a sample
not used in the design of the strategy (a.k.a. testing
set). A backtest is realistic when the IS performance
is consistent with the OOS performance.
When an investor receives a promising backtest
from a researcher or portfolio manager, one of
her key problems is to assess how realistic that
simulation is. This is because, given any financial
series, it is relatively simple to overfit an investment
strategy so that it performs well IS.
Overfitting is a concept borrowed from machine learning and denotes the situation when a
model targets particular observations rather than
a general structure. For example, a researcher
could design a trading system based on some
parameters that target the removal of specific
recommendations that she knows led to losses
IS (a practice known as data snooping). After a
few iterations, the researcher will come up with
optimal parameters, which profit from features
that are present in that particular sample but may
well be rare in the population.
Recent computational advances allow investment managers to methodically search through
thousands or even millions of potential options for
a profitable investment strategy. In many instances,
that search involves a pseudo-mathematical argument which is spuriously validated through a
backtest. For example, consider a time series of
daily prices for a stock X. For every day in the
sample, we can compute one average price of
m
that stock using the previous m observations x
and another average price using the previous n
n , where m < n. A popular investobservations x
ment strategy called crossing moving averages
m > x
n . Indeed,
consists of owning X whenever x
since the sample size determines a limited number
of parameter combinations that m and n can adopt,

Notices of the AMS

Volume 61, Number 5

it is relatively easy to determine the pair (m, n)

that maximizes the backtests performance. There
are hundreds of such popular strategies, marketed
to unsuspecting lay investors as mathematically
sound and empirically tested.
In the context of econometric models several
procedures have been proposed to determine
overfit in White [27], Romano et al. [23], and
Harvey et al. [9]. These methods propose to adjust
the p-values of estimated regression coefficients
to account for the multiplicity of trials. These
approaches are valuable for dealing with trading
rules based on an econometric specification.
The machine-learning literature has devoted
significant effort to studying the problem of
overfitting. The proposed methods typically are
not applicable to investment problems for multiple
reasons. First, these methods often require explicit
point forecasts and confidence bands over a
defined event horizon in order to evaluate the
explanatory power or quality of the prediction
(e.g., E-mini S&P500 is forecasted to be around
1,600 with a one-standard deviation of 5 index
points at Fridays close). Very few investment
strategies yield such explicit forecasts; instead,
they provide qualitative recommendations (e.g.,
buy or strong buy) over an undefined period
until another such forecast is generated, with
random frequency. For instance, trading systems,
like the crossing of moving averages explained
earlier, generate buy and sell recommendations
with little or no indication as to forecasted values,
confidence in a particular recommendation, or
expected holding period.
Second, even if a particular investment strategy relies on such a forecasting equation, other
components of the investment strategy may have
been overfitted, including entry thresholds, risk
sizing, profit taking, stop-loss, cost of capital, and
so on. In other words, there are many ways to
overfit an investment strategy other than simply
tuning the forecasting equation. Third, regression
overfitting methods are parametric and involve a
number of assumptions regarding the underlying
data which may not be easily ascertainable. Fourth,
some methods do not control for the number of
trials attempted.
To illustrate this point, suppose that a researcher
is given a finite sample and told that she needs to
come up with a strategy with an SR (Sharpe Ratio,
a popular measure of performance in the presence
of risk) above 2.0, based on a forecasting equation
for which the AIC statistic (Akaike Information
Criterion, a standard of the regularization method)
rejects the null hypothesis of overfitting with a
95 percent confidence level (i.e., a false positive
rate of 5 percent). After only twenty trials, the
researcher is expected to find one specification

May 2014

that passes the AIC criterion. The researcher will

quickly be able to present a specification that not
only (falsely) passes the AIC test but also gives an
SR above 2.0. The problem is, AICs assessment did
not take into account the hundreds of other trials
that the researcher neglected to mention. For these
reasons, commonly used regression overfitting
methods are poorly equipped to deal with backtest
overfitting.
Although there are many academic studies that
claim to have identified profitable investment
strategies, their reported results are almost always
based on IS statistics. Only exceptionally do we
find an academic study that applies the hold-out
method or some other procedure to evaluate performance OOS. Harvey, Liu, and Zhu [10] argue that
there are hundreds of papers supposedly identifying hundreds of factors with explanatory power
over future stock returns. They echo Ioannidis
[13] in concluding that most claimed research
findings are likely false. Factor models are only
the tip of the iceberg.1 The reader is probably
familiar with many publications solely discussing
IS performance.
This situation is, quite frankly, depressing,
particularly because academic researchers are
expected to recognize the dangers and practice of
overfitting. One common criticism, of course, is
the credibility problem of holding-out when the
researcher had access to the full sample anyway.
Leinweber and Sisk [15] present a meritorious
exception. They proposed an investment strategy
in a conference and announced that six months
later they would publish the results with the
pure (yet to be observed) OOS data. They called
this approach model sequestration, which is an
extreme variation of hold-out.
Our Intentions
In this paper we shall show that it takes a relatively
small number of trials to identify an investment
strategy with a spuriously high backtested performance. We also compute the minimum backtest
length (MinBTL) that an investor should require
given the number of trials attempted. Although in
our examples we always choose the Sharpe ratio
to evaluate performance, our methodology can be
applied to any other performance measure.
We believe our framework to be helpful to
the academic and investment communities by
providing a benchmark methodology to assess the
reliability of a backtested performance. We would
1

We invite the reader to read specific instances of pseudomathematical financial advice at this website: http://www.
m-a-f-f-i-a.org/. Also, Edesses (2007) provides numerous examples.

Notices of the AMS

459

Figure 1. Overfitting a backtests results as the

number of trials grows.

Figure 1 provides a graphical representation of

Proposition 1. The blue (dotted) line shows the maximum of a particular set of N independent random
numbers, each following a Standard Normal distribution. The black (continuous) line is the expected
value of the maximum of that set of N random
numbers. The red (dashed) line is an upper bound
estimate of that maximum. The implication is that it
is relatively easy to wrongly select a strategy on the
basis of a maximum Sharpe ratio when displayed
IS.
feel sufficiently rewarded in our efforts if at least
this paper succeeded in drawing the attention
of the mathematical community regarding the
widespread proliferation of journal publications,
many of them claiming profitable investment
strategies on the sole basis of IS performance. This
is perhaps understandable in business circles, but
a higher standard is and should be expected from
an academic forum.
We would also like to raise the question of
whether mathematical scientists should continue
to tolerate the proliferation of investment products
that are misleadingly marketed as mathematically
sound. In the recent words of Sir Andrew Wiles,
One has to be aware now that mathematics
can be misused and that we have to protect
its good name. [29]
We encourage the reader to search the Internet for
terms such as stochastic oscillators, Fibonacci
ratios, cycles, Elliot wave, Golden ratio,
parabolic SAR, pivot point, momentum, and
others in the context of finance. Although such
terms clearly evoke precise mathematical concepts,
in fact in almost all cases their usage is scientifically
unsound.

460

Historically, scientists have led the way in exposing those who utilize pseudoscience to extract
a commercial benefit. As early as the eighteenth
century, physicists exposed the nonsense of astrologers. Yet mathematicians in the twenty-first
century have remained disappointingly silent with
regard to those in the investment community who,
knowingly or not, misuse mathematical techniques
such as probability theory, statistics, and stochastic calculus. Our silence is consent, making us
accomplices in these abuses.
The rest of our study is organized as follows:
The section Backtest Overfitting introduces the
problem in a more formal way. The section
Minimum Backtest Length (MinBTL) defines the
concept of Minimum Backtest Length (MinBTL).
The section Model Complexity argues how model
complexity leads to backtest overfitting. The section
Overfitting in Absence of Compensation Effects
analyzes overfitting in the absence of compensation
effects. The section Overfitting in Presence of
Compensation Effects studies overfitting in the
presence of compensation effects. The section
Is Backtest Overfitting a Fraud? exposes how
backtest overfitting can be used to commit fraud.
The section A Practical Application presents
a typical example of backtest overfitting. The
section Conclusions lists our conclusions. The
mathematical appendices supply proofs of the
propositions presented throughout the paper.

Backtest Overfitting
The design of an investment strategy usually
begins with a prior or belief that a certain pattern
may help forecast the future value of a financial
variable. For example, if a researcher recognizes a
lead-lag effect between various tenor bonds in a
yield curve, she could design a strategy that bets on
a reversion towards equilibrium values. This model
might take the form of a cointegration equation,
a vector-error correction model, or a system of
stochastic differential equations, just to name a
few. The number of possible model configurations
(or trials) is enormous, and naturally the researcher
would like to select the one that maximizes the
performance of the strategy. Practitioners often rely
on historical simulations (also called backtests) to
discover the optimal specification of an investment
strategy. The researcher will evaluate, among
other variables, what are the optimal sample sizes,
signal update frequency, entry and profit-taking
thresholds, risk sizing, stop losses, maximum
holding periods, etc.
The Sharpe ratio is a statistic that evaluates an
investment managers or strategys performance on
the basis of a sample of past returns. It is defined as
the ratio between average excess returns (in excess
of the rate of return paid by a risk-free asset, such as

Notices of the AMS

Volume 61, Number 5

a government note) and the standard deviation of

the same returns. Intuitively, this can be interpreted
as a return on risk (or as William Sharpe put it,
return on variability). But the standard deviation
of excess returns may be a misleading measure
of variability when returns follow asymmetric or
fat-tailed distributions or when returns are not
independent or identically distributed. Suppose
that a strategys excess returns (or risk premiums),
rt , are independent and identically distributed (IID)
following a Normal law:
(1)

rt N (, 2 ),

where N represents a Normal distribution with

mean and variance 2 . The annualized Sharpe
ratio (SR) can be computed as

(2)
SR =
q,

where q is the number of returns per year (see Lo

[17] for a detailed derivation of this expression).
Sharpe ratios are typically expressed in annual
terms in order to allow for the comparison of
strategies that trade with different frequency. The
great majority of financial models are built upon the
IID Normal assumption, which may explain why the
Sharpe ratio has become the most popular statistic
for evaluating an investments performance.
Since , are usually unknown, the true value
SR cannot be known for certain. Instead, we can
d = q, where

estimate the Sharpe ratio as SR

are the sample mean and sample standard

and
deviation. The inevitable consequence is that
SR calculations are likely to be the subject of
substantial estimation errors (see Bailey and Lpez
de Prado [2] for a confidence band and an extension
of the concept of Sharpe Ratio beyond the IID
Normal assumption).
From Lo [17] we know that the distribution of the
d converges
estimated annualized Sharpe ratio SR
asymptotically (as y ) to

2
1 + SR
a
2q
,
d - N SR,
(3)
SR
y
where y is the number of years used to estimate
d 2 As y increases without bound, the probaSR.
d approaches a Normal
bility distribution of SR
distribution with mean SR and variance
2

1+

SR 2
2q

. For a

Most performance statistics assume IID Normal returns

and so are normally distributed. In the case of the Sharpe
ratio, several authors have proved that its asymptotic distribution follows a Normal law even when the returns are
not IID Normal. The same result applies to the Information
Ratio. The only requirement is that the returns be ergodic.
We refer the interested reader to Bailey and Lpez de Prado
[2].

May 2014

Figure 2. Minimum Backtest Length needed to

avoid overfitting, as a function of the number of
trials.

Figure 2 shows the tradeoff between the number

of trials (N) and the minimum backtest length
(MinBTL) needed to prevent skill-less strategies to be
generated with a Sharpe ratio IS of 1. For instance,
if only five years of data are available, no more than
forty-five independent model configurations should
be tried. For that number of trials, the expected
maximum SR IS is 1, whereas the expected SR OOS
is 0. After trying only seven independent strategy
configurations, the expected maximum SR IS is 1 for
a two-year long backtest, while the expected SR OOS
is 0. The implication is that a backtest which does
not report the number of trials N used to identify
the selected configuration makes it impossible to
assess the risk of overfitting.
sufficiently large y, (3) provides an approximation
d
of the distribution of SR.
Even for a small number N of trials, it is
relatively easy to find a strategy with a high Sharpe
ratio IS but which also delivers a null Sharpe ratio
OOS. To illustrate this point, consider N strategies
with T = yq returns distributed according to a
Normal law with mean excess returns and with
standard deviation . Suppose that we would like
d IS, based
to select the strategy with optimal SR
on one year of observations. A risk we face is
choosing a strategy with a high Sharpe ratio IS but
zero Sharpe ratio OOS. So we ask the question,
how high is the expected maximum Sharpe ratio IS
among a set of strategy configurations where the
true Sharpe ratio is zero?
Bailey and Lpez de Prado [2] derived an
estimate of the Minimum Track Record Length
(MinTRL) needed to reject the hypothesis that an
estimated Sharpe ratio is below a certain threshold

Notices of the AMS

461

Figure 3. Performance IS vs. OOS before

introducing strategy selection.

Figure 3 shows the relation between SR IS (xaxis) and SR OOS (y-axis) for = 0, = 1, N =
1000, T = 1000. Because the process follows a
random walk, the scatter plot has a circular shape
centered at the point (0, 0). This illustrates the fact
that, in absence of compensation effects, overfitting
the IS performance (x-axis) has no bearing on the
OOS performance (y-axis), which remains around
zero.
(lets say zero). MinTRL was developed to evaluate
a strategys track record (a single realized path,
N = 1). The question we are asking now is different,
because we are interested in the backtest length
needed to avoid selecting a skill-less strategy
among N alternative specifications. In other words,
in this article we are concerned with overfitting
prevention when comparing multiple strategies,
not in evaluating the statistical significance of
a single Sharpe ratio estimate. Next, we will
derive the analogue to MinTRL in the context of
overfitting, which we will call Minimum Backtest
Length (MinBTL), since it specifically addresses the
problem of backtest overfitting.
a
d -
From (3), if = 0 and y = 1, then SR
N (0, 1).
Note that because SR = 0, increasing q does not
reduce the variance of the distribution. The proof
of the following proposition is left for the appendix.
Proposition 1. Given a sample of IID random variables, xn Z, n = 1, . . . , N, where Z is the CDF
of the Standard Normal distribution, the expected
maximum of that sample, E[max N ] = E[max{xn }],
can be approximated for a large N as


1
E[max ] (1 )Z 1 1
N
N


(4)
1 1
1
+ Z
1 e
N
462

where 0.5772156649 . . . is the Euler-Mascheroni

constant and N  1.
p
An upper bound to (4) is 2 ln[N].3 Figure 1
plots, for various values of N (x-axis), the expected
Sharpe ratio of the optimal strategy IS. For example,
if the researcher tries only N = 10 alternative
configurations of an investment strategy, she is
expected to find a strategy with a Sharpe ratio
IS of 1.57 despite the fact that all strategies are
expected to deliver a Sharpe ratio of zero OOS
(including the optimal one selected IS).
Proposition 1 has important implications. As
the researcher tries a growing number of strategy
configurations, there will be a nonnull probability
of selecting IS a strategy with null expected
performance OOS. Because the hold-out method
does not take into account the number of trials
attempted before selecting a model, it cannot
assess the representativeness of a backtest.

Minimum Backtest Length (MinBTL)

Let us consider now the case that = 0 but y 1.
Then, we can still apply Proposition 1 by rescaling
the expected maximum by the standard deviation
of the annualized Sharpe ratio, y 1/2 . Thus, the
researcher is expected to find an optimal strategy
with an IS annualized Sharpe ratio of
(5)
E[max ]
N





1
1
+Z 1 1 e1 .
y 1/2 (1 )Z 1 1
N
N
Equation (5) says that the more independent the
configurations a researcher tries (N), the more
likely she is to overfit, and therefore the higher the
acceptance threshold should be for the backtested
result to be trusted. This situation can be partially
mitigated by increasing the sample size (y). By
solving (5) for y, we obtain the following statement.
Theorem 2. The Minimum Backtest Length
(MinBTL, in years) needed to avoid selecting a
strategy with an IS Sharpe ratio of E[max N ]
among N independent strategies with an expected
OOS Sharpe ratio of zero is
(6)
MinBT L
h
i
h

1
(1)Z 1 1 N
+Z 1 1

E[maxN ]

1 1
Ne

i 2
<

2 ln[N]
E[max N ]

Equation (6) tells us that MinBTL must grow

as the researcher tries more independent model
configurations (N) in order to keep constant the
expected maximum Sharpe ratio at a given level
3

See Example 3.5.4 of Embrechts et al. [5] for a detailed

treatment of the derivation of upper bounds on the maximum of a Normal distribution.

Notices of the AMS

Volume 61, Number 5

E[max N ]. Figure 2 shows how many years of backtest length (MinBTL) are needed so that E[max N ]
is fixed at 1. For instance, if only five years of data
are available, no more than forty-five independent
model configurations should be tried or we are
almost guaranteed to produce strategies with an
annualized Sharpe ratio IS of 1 but an expected
Sharpe ratio OOS of zero. Note that Proposition 1
assumed the N trials to be independent, which
leads to a quite conservative estimate. If the trials
performed were not independent, the number of
independent trials N involved could be derived
using a dimension-reduction procedure, such as
Principal Component Analysis.
We will examine this tradeoff between N and
T in greater depth later in the paper without
requiring such a strong assumption, but MinBTL
gives us a first glance at how easy it is to overfit by
merely trying alternative model configurations. As
an approximation, the reader may find it helpful
to remember the upper bound to the minimum
backtest length (in years), MinBT L < 2 ln[N] 2 .
E[maxN ]

Of course, a backtest may be overfit even if it is

computed on a sample greater than MinBTL. From
that perspective, MinBTL should be considered
a necessary, nonsufficient condition to avoid
overfitting. We leave to Bailey et al. [1] the derivation
of a more precise measure of backtest overfitting.

Model Complexity
How does the previous result relate to model
complexity? Consider a one-parameter model that
may adopt two possible values (like a switch
that generates a random sequence of trades) on
a sample of T observations. Overfitting will be
difficult, because N = 2. Lets say that we make
the model more complex by adding four more
parameters so that the total number of parameters
becomes 5, i.e., N = 25 = 32. Having thirty-two
independent sequences of random trades greatly
increases the possibility of overfitting.
While a greater N makes overfitting easier,
it makes perfectly fitting harder. Modern supercomputers can only perform around 250 raw
computations per second, or less than 258 raw
computations per year. Even if a trial could be
reduced to a raw computation, searching N = 2100
will take us 242 supercomputer-years of computation (assuming a 1 Pflop/s system, capable of
1015 floating-point operations per second). Hence,
a skill-less brute force search is certainly impossible. While it is hard to perfectly fit a complex
skill-less strategy, Proposition 1 shows that there
is no need for that. Without perfectly fitting a
strategy or making it overcomplex, a researcher
can achieve high Sharpe ratios. A relatively simple
strategy with just seven binomial independent parameters offers N = 27 = 128 trials, with an expected

May 2014

Figure 4. Performance IS vs. performance OOS

for one path after introducing strategy selection.

Figure 4 provides a graphical representation of

what happens when we select the random walk
with highest SR IS. The performance of the first
half was optimized IS, and the performance of the
second half is what the investor receives OOS. The
good news is, in the absence of memory, there is
no reason to expect overfitting to induce negative
performance.
maximum Sharpe ratio above 2.6 (see Figure 1).
We suspect, however, that backtested strategies
that significantly beat the market typically rely
on some combination of valid insight, boosted
by some degree of overfitting. Since believing in
such an artificially enhanced high-performance
strategy will often also lead to over leveraging,
such overfitting is still very damaging. Most
Technical Analysis strategies rely on filters, which
are sets of conditions that trigger trading actions,
like the random switches exemplified earlier.
Accordingly, extra caution is warranted to guard
against overfitting in using Technical Analysis
strategies, as well as in complex nonparametric
modeling tools, such as Neural Networks and
Kernel Estimators.
Here is a key concept that investors generally
miss:
A researcher that does not report the number of trials N used to identify the selected
backtest configuration makes it impossible
to assess the risk of overfitting.
Because N is almost never reported, the
magnitude of overfitting in published backtests is
unknown. It is not hard to overfit a backtest (indeed,
the previous theorem shows that it is hard not to), so
we suspect that a large proportion of backtests published in academic journals may be misleading. The

Notices of the AMS

463

Overfitting in Absence of Compensation

Effects

Figure 5. Performance degradation after

introducing strategy selection in absence of
compensation effects.

Figure 5 illustrates what happens once we add a

model selection procedure. Now the SR IS ranges
from 1.2 to 2.6, and it is centered around 1.7. Although the backtest for the selected model generates
the expectation of a 1.7 SR, the expected SR OOS
is unchanged and lies around 0.
situation is not likely to be better among practitioners.
In our experience, overfitting is pathological
within the financial industry, where proprietary
and commercial software is developed to estimate
the combination of parameters that best fits (or,
more precisely, overfits) the data. These tools allow
the user to add filters without ever reporting how
such additions increase the probability of backtest
overfitting. Institutional players are not immune
to this pitfall. Large mutual fund groups typically
discontinue and replace poorly performing funds,
introducing survivorship and selection bias. While
the motivation of this practice may be entirely
innocent, the effect is the same as that of hiding
experiments and inflating expectations.
We are not implying that those technical analysts, quantitative researchers, or fund managers
are snake oil salesmen. Most likely they most
genuinely believe that the backtested results are
legitimate or that adjusted fund offerings better represent future performance. Hedge fund
managers are often unaware that most backtests
presented to them by researchers and analysts may
be useless, and so they unknowingly package faulty
investment propositions into products. One goal
of this paper is to make investors, practitioners,
and academics aware of the futility of considering
backtest without controlling for the probability of
overfitting.

464

Regardless of how realistic the prior being tested

is, there is always a combination of parameters
that is optimal. In fact, even if the prior is false, the
researcher is very likely to identify a combination of
parameters that happens to deliver an outstanding
performance IS. But because the prior is false, OOS
performance will almost certainly underperform
the backtests results. As we have described,
this phenomenon, by which IS results tend to
outperform the OOS results, is called overfitting.
It occurs because a sufficiently large number of
parameters are able to target specific data points
say by chance buying just before a rally and
shorting a position just before a sell-offrather
than triggering trades according to the prior.
To illustrate this point, suppose we generate
N Gaussian random walks by drawing from a
Standard Normal distribution, each walk having a
size T . Each performance path m can be obtained
as a cumulative sum of Gaussian draws
(7)

m = + ,

where the random shocks are IID distributed

Z, = 1, . . . , T . Suppose that each path has
been generated by a particular combination of
parameters, backtested by a researcher. Without
loss of generality, assume that = 0, = 1, and
T = 1000, covering a period of one year (with about
four observations per trading day). We divide these
paths into two disjoint samples of equal size, 500,
and call the first one IS and the second one OOS.
At the moment of choosing a particular parameter combination as optimal, the researcher
had access to the IS series, not the OOS. For each
model configuration, we may compute the Sharpe
ratio of the series IS and compare it with the
Sharpe ratio of the series OOS. Figure 3 shows the
resulting scatter plot. The p-values associated with
the intercept and the IS performance (SR a priori)
are respectively 0.6261 and 0.7469.
The problem of overfitting arises when the
researcher uses the IS performance (backtest) to
choose a particular model configuration, with the
expectation that configurations that performed
well in the past will continue to do so in the
future. This would be a correct assumption if the
parameter configurations were associated with a
truthful prior, but this is clearly not the case of the
simulation above, which is the result of Gaussian
random walks without trend ( = 0).
Figure 4 shows what happens when we select the
model configuration associated with the random
walk with highest Sharpe ratio IS. The performance
of the first half was optimized IS, and the performance of the second half is what the investor
receives OOS. The good news is that under these

Notices of the AMS

Volume 61, Number 5

conditions, there is no reason to expect overfitting

to induce negative performance. This is illustrated
in Figure 5, which shows how the optimization
causes the expected performance IS to range between 1.2 and 2.6, while the OOS performance will
range between -1.5 and 1.5 (i.e., around , which
in this case is zero). The p-values associated with
the intercept and the IS performance (SR a priori)
are respectively 0.2146 and 0.2131. Selecting an
optimal model IS had no bearing on the performance OOS, which simply equals the zero mean of
the process. A positive mean ( > 0) would lead
to positive expected performance OOS, but such
performance would nevertheless be inferior to the
one observed IS.

Overfitting in Presence of Compensation

Effects
Multiple causes create compensation effects
in practice, such as overcrowded investment
opportunities, major corrections, economic cycles,
reversal of financial flows, structural breaks,
bubble bursts, etc. Optimizing a strategys parameters (i.e., choosing the model configuration that
maximizes the strategys performance IS) does
not necessarily lead to improved performance
(compared to not optimizing) OOS, yet again
leading to overfitting.
In some instances, when the strategys performance series lacks memory, overfitting leads to no
improvement in performance OOS. However, the
presence of memory in a strategys performance
series induces a compensation effect, which increases the chances for that strategy to be selected
IS, only to underperform the rest OOS. Under those
circumstances, IS backtest optimization is in fact
detrimental to OOS performance.4
Global Constraint
Unfortunately, overfitting rarely has the neutral
implications discussed in the previous section. Our
previous example was purposely chosen to exhibit
a globally unconditional behavior. As a result, the
OOS data had no memory of what occurred IS.
Centering each path to match a mean removes
one degree of freedom:
(8)

m = m +

T
1 X
m .
T =1

Figure 6. Performance degradation as a result of

strategy selection under compensation effects
(global constraint).

Figure 6 shows that adding a single global constraint causes the OOS performance to be negative
even though the underlying process was trendless.
Also, a strongly negative linear relation between
performance IS and OOS arises, indicating that the
more we optimize IS, the worse the OOS performance of the strategy.
We may rerun the same Monte Carlo experiment
as before, this time on the recentered variables
m . Somewhat scarily, adding this single global
constraint causes the OOS performance to be
negative even though the underlying process was
trendless. Moreover, a strongly negative linear
relation between performance IS and OOS arises,
indicating that the more we optimize IS, the
worse the OOS performance. Figure 6 displays this
disturbing pattern. The p-values associated with
the intercept and the IS performance (SR a priori)
are respectively 0.5005 and 0, indicating that the
negative linear relation between IS and OOS Sharpe
ratios is statistically significant.
The following proposition is proven in the
appendix.
Proposition 3. Given two alternative configuraA
tions (A and B) of the same model, where IS
=
A
B
B
OOS = IS = OOS imposing a global constraint
A = B , implies that
(9)

A
B
A
B
SRIS
> SRIS
a SROOS
< SROOS
.

Bailey et al. [1] propose a method to determine the degree

to which a particular backtest may have been compromised
by the risk of overfitting.

May 2014

Notices of the AMS

465

Proposition 4. The half-life period of a first-order

autoregressive process with autoregressive coefficient (0, 1) occurs at
(12)

Figure 7. Performance degradation as a result of

strategy selection under compensation effects
(first-order serial correlation).

Figure 7 illustrates that a serially correlated performance introduces another form of compensation
effects, just as we saw in the case of a global constraint. For example, if = 0.995, it takes about
138 observations to recover half of the deviation
from the equilibrium. We have rerun the previous
Monte Carlo experiment, this time on an autoregressive process with = 0, = 1, = 0.995, and
have plotted the pairs of performance IS vs. OOS.
Recentering a series is one way to introduce
memory into a process, because some data points
will now compensate for the extreme outcomes
from other data points. By optimizing a backtest,
the researcher selects a model configuration that
spuriously works well IS and consequently is likely
to generate losses OOS.
Serial Dependence
Imposing a global constraint is not the only
situation in which overfitting actually is detrimental.
To cite another (less restrictive) example, the same
effect happens if the performance series is serially
conditioned, such as a first-order autoregressive
process,
(10)

m = (1 ) + ( 1)m1 +

or, analogously,
(11)

m = (1 ) + m1 + ,

where the random shocks are again IID distributed

as Z. The following proposition is proven in
the appendix. The number of observations that
it takes for a process to reduce its divergence
from the long-run equilibrium by half is known as
the half-life period, or simply half-life (a familiar
physical concept introduced by Ernest Rutherford
in 1907).

466

ln[2]
.
ln[]

For example, if = 0.995, it takes about 138

observations to retrace half of the deviation from
the equilibrium. This introduces another form of
compensation effect, just as we saw in the case of
a global constraint. If we rerun the previous Monte
Carlo experiment, this time for the autoregressive
process with = 0, = 1, = 0.995, and plot
the pairs of performance IS vs. OOS, we obtain
Figure 7.
The p-values associated with the intercept and
the IS performance (SR a priori) are respectively
0.4513 and 0, confirming that the negative linear
relation between IS and OOS Sharpe ratios is again
statistically significant. Such serial correlation
is a well-known statistical feature, present in
the performance of most hedge fund strategies.
Proposition 5 is proved in the appendix.
Proposition 5. Given two alternative configuraA
tions (A and B) of the same model, where IS
=
A
B
B
OOS = IS = OOS and the performance series follows the same first-order autoregressive stationary
process,
(13)

A
B
A
B
SRIS
> SRIS
a SROOS
< SROOS
.

Proposition 5 reaches the same conclusion as

Proposition 3 (a compensation effect) without
requiring a global constraint.

Is Backtest Overfitting a Fraud?

Consider an investment manager who emails his
stock market forecast for the next month to 2n x
prospective investors, where x and n are positive
integers. To half of them he predicts that markets
will go up, and to the other half that markets will
go down. After the month passes, he drops from
his list the names to which he sent the incorrect
forecast, and sends a new forecast to the remaining
2n1 x names. He repeats the same procedure n
times, after which only x names remain. These x
investors have witnessed n consecutive infallible
forecasts and may be extremely tempted to give this
investment manager all of their savings. Of course,
this is a fraudulent scheme based on random
screening: The investment manager is hiding the
fact that for every one of the x successful witnesses,
he has tried 2n unsuccessful ones (see Harris [8, p.
473] for a similar example).
To avoid falling for this psychologically compelling fraud, a potential investor needs to consider
the economic cost associated with manufacturing
the successful experiments, and require the investment manager to produce a number n for which

Notices of the AMS

Volume 61, Number 5

the scheme is uneconomic. One caveat is, even if n

is too large for a skill-less investment manager, it
may be too low for a mediocre investment manager
who uses this scheme to inflate his skills.
Not reporting the number of trials (N) involved
in identifying a successful backtest is a similar kind
of fraud. The investment manager only publicizes
the model that works but says nothing about all the
failed attempts, which as we have seen can greatly
increase the probability of backtest overfitting.
An analogous situation occurs in medical research, where drugs are tested by treating hundreds
or thousands of patients; however, only the best
outcomes are publicized. The reality is that the
selected outcomes may have healed in spite of
(rather than thanks to) the treatment or due to
a placebo effect (recall Proposition 1). Such behavior is unscientificnot to mention dangerous
and expensiveand has led to the launch of the
alltrials.net project, which demands that all results
(positive and negative) for every experiment are
made publicly available. A step forward in this
direction is the recent announcement by Johnson &
Johnson that it plans to open all of its clinical test
results to the public [14]. For a related discussion
of reproducibility in the context of mathematical
computing, see Stodden et al. [25].
Hiding trials appears to be standard procedure
in financial research and financial journals. As
an aggravating factor, we know from the section
Overfitting in Presence of Compensation Effects
that backtest overfitting typically has a detrimental
effect on future performance due to the compensation effects present in financial series. Indeed, the
customary disclaimer past performance is not an
indicator of future results is too optimistic in the
context for backtest overfitting. When investment
advisers do not control for backtest overfitting,
good backtest performance is an indicator of
negative future results.

A Practical Application
Institutional asset managers follow certain investment procedures on a regular basis, such as
rebalancing the duration of a fixed income portfolio (PIMCO); rolling holdings on commodities
(Goldman Sachs, AIG, JP Morgan, Morgan Stanley);
investing or divesting as new funds flow at the end
of the month (Fidelity, BlackRock); participating in
the regular U.S. Treasury Auctions (all major investment banks); delevering in anticipation of payroll,
FOMC or GDP releases; tax-driven effects around
the end of the year and mid-April; positioning for
electoral cycles, etc. There are a large number of
instances where asset managers will engage in
somewhat predictable actions on a regular basis.
It should come as no surprise that a very popular

May 2014

Figure 8. Backtested performance of a seasonal

strategy (Example 6).

We have generated a time series of 1000 daily prices

(about four years) following a random walk. The
PSR-Stat of the optimal model configuration is 2.83,
which implies a less-than 1 percent probability that
the true Sharpe ratio is below 0. Consequently, we
have been able to identify a plausible seasonal strategy with an SR of 1.27 despite the fact that no true
seasonal effect exists.
investment strategy among hedge funds is to profit
from such seasonal effects.
For example, a type of question often asked
by hedge fund managers follows the form: Is
there a time interval every [ ] when I would
have made money on a regular basis? You
may replace the blank space with a word like
day, week, month, quarter, auction, nonfarm
payroll (NFP) release, European Central Bank (ECB)
announcement, presidential election year, . . . . The
variations are as abundant as they are inventive.
Doyle and Chen [4] study the weekday effect and
conclude that it appears to wander.
The problem with this line of questioning
is that there is always a time interval that is
arbitrarily optimal regardless of the cause. The
answer to one such question is the title of a very
popular investment classic, Do Not Sell Stocks on
Monday, by Hirsch [12]. The same author wrote an
almanac for stock traders that reached its forty-fifth
edition in 2012 and he is also a proponent of the
Santa Claus Rally, the quadrennial political/stock
market cycle, and investing during the Best
Six Consecutive Months of the year, November
through April. While these findings may indeed
be caused by some underlying seasonal effect,
it is easy to demonstrate that any random data
contains similar patterns. The discovery of a
pattern IS typically has no bearing OOS, yet again is
a result of overfitting. Running such experiments

Notices of the AMS

467

without controlling for the probability of backtest

overfitting will lead the researcher to spurious
claims. OOS performance will disappoint, and the
reason will not be that the market has found
out the seasonal effect and arbitraged away the
strategys profits. Rather, the effect was never
there; instead, it was just a random pattern that
gave rise to an overfitted trading rule. We will
illustrate this point with an example.
Example 6. Suppose that we would like to identify the optimal monthly trading rule given four
customary parameters: Entry day, Holding period,
Stop loss, and Side. Side defines whether we will
hold long or short positions on a monthly basis.
Entry day determines the business day of the
month when we enter a position. Holding period
gives the number of days that the position is held.
Stop loss determines the size of the loss (as a multiple of the series volatility) that triggers an exit
for that months position. For example, we could
explore all nodes that span the set {1, . . . , 22} for
Entry day, the set {1, . . . , 20} for Holding period,
the set {0, . . . , 10} for Stop loss, and {1, 1} for
Side. The parameter combinations involved form
a four-dimensional mesh of 8,800 elements. The
optimal parameter combination can be discovered
by computing the performance derived by each
node.
First, we generated a time series of 1,000 daily
prices (about four years), following a random walk.
Figure 8 plots the random series, as well as the
performance associated with the optimal parameter combination: Entry day = 11, Holding period
= 4, Stop loss = 1 and Side = 1. The annualized
Sharpe ratio is 1.27.
Given the elevated Sharpe ratio, we could conclude that this strategys performance is significantly greater than zero for any confidence level.
Indeed, the PSR-Stat is 2.83, which implies a less
than 1 percent probability that the true Sharpe
ratio is below 0.5 Several studies in the practitioners and academic literature report similar results,
which are conveniently justified with some ex-post
explanation (the posterior gives rise to a prior).
What this analysis misses is an evaluation of the
probability that this backtest has been overfit to
the data, which is the subject of Bailey et al. [1].
In this practical application we have illustrated
how simple it is to produce overfit backtests when
answering common investment questions, such
as the presence of seasonal effects. We refer the
reader to the appendix section Reproducing the
5

The Probabilistic Sharpe Ratio (or PSR) is an extension to

the SR. Nonnormality increases the error of the variance
estimator, and PSR takes that into consideration when determining whether an SR estimate is statistically significant.
See Bailey and Lpez de Prado [2] for details.

468

Results in Example 6 for the implementation of

this experiment in the Python language. Similar
experiments can be designed to demonstrate
overfitting in the context of other effects, such
as trend following, momentum, mean-reversion,
event-driven effects, etc. Given the facility with
which elevated Sharpe ratios can be manufactured
IS, the reader would be well advised to remain
highly suspicious of backtests and of researchers
who fail to report the number of trials attempted.

Conclusions
While the literature on regression overfitting is
extensive, we believe that this is the first study
to discuss the issue of overfitting on the subject
of investment simulations (backtests) and its
negative effect on OOS performance. On the
subject of regression overfitting, the great Enrico
Fermi once remarked (Mayer et al. [20]):
I remember my friend Johnny von Neumann
used to say, with four parameters I can fit
an elephant, and with five I can make him
wiggle his trunk.
The same principle applies to backtesting, with
some interesting peculiarities. We have shown that
backtest overfitting is difficult indeed to avoid. Any
perseverant researcher will always be able to find a
backtest with a desired Sharpe ratio regardless of
the sample length requested. Model complexity is
only one way that backtest overfitting is facilitated.
Given that most published backtests do not report
the number of trials attempted, many of them
may be overfitted. In that case, if an investor
allocates capital to them, performance will vary: It
will be around zero if the process has no memory,
but it may be significantly negative if the process
has memory. The standard warning that past
performance is not an indicator of future results
understates the risks associated with investing
on overfit backtests. When financial advisors do
not control for overfitting, positive backtested
performance will often be followed by negative
investment results.
We have derived the expected maximum Sharpe
ratio as a function of the number of trials (N) and
sample length. This has allowed us to determine
the Minimum Backtest Length (MinBTL) needed to
avoid selecting a strategy with a given IS Sharpe
ratio among N trials with an expected OOS Sharpe
ratio of zero. Our conclusion is that the more trials
a financial analyst executes, the greater should
be the IS Sharpe ratio demanded by the potential
investor.
We strongly suspect that such backtest overfitting is a large part of the reason why so many
algorithmic or systematic hedge funds do not live
up to the elevated expectations generated by their
managers.

Notices of the AMS

Volume 61, Number 5

We would feel sufficiently rewarded in our efforts

if this paper succeeds in drawing the attention of
the mathematical community to the widespread
proliferation of journal publications, many of
them claiming profitable investment strategies on
the sole basis of in-sample performance. This is
understandable in business circles, but a higher
standard is and should be expected from an
academic forum.
A depressing parallel can be drawn between
todays financial academic research and the situation denounced by economist and Nobel Laureate
Wassily Leontief writing in Science (see Leontief
[16]):
A dismal performance
. . . What economists revealed most
clearly was the extent to which their
profession lags intellectually. This editorial comment by the leading economic
weekly (on the 1981 annual proceedings of
the American Economic Association) says,
essentially, that the king is naked. But no
one taking part in the elaborate and solemn
procession of contemporary U.S. academic
economics seems to know it, and those who
do dont dare speak up.
[. . .]
[E]conometricians fit algebraic functions
of all possible shapes to essentially the
same sets of data without being able to
advance, in any perceptible way, a systematic understanding of the structure and the
operations of a real economic system.
[. . .]
That state is likely to be maintained as
long as tenured members of leading economics departments continue to exercise
tight control over the training, promotion,
and research activities of their younger
faculty members and, by means of peer
review, of the senior members as well.

May 2014

Appendices
Proof of Proposition 1
Embrechts et al. [5, pp. 138147] show that the
maximum value (or last order statistic) in a sample
of independent random variables following an
exponential distribution converges asymptotically
to a Gumbel distribution. As a particular case, the
Gumbel distribution covers the Maximum Domain
of Attraction of the Gaussian distribution, and
therefore it can be used to estimate the expected
value of the maximum of several independent
random Gaussian variables.
To see how, suppose there is a sample of IID
random variables, zn Z, n = 1, . . . , N, where Z is
the CDF of the Standard Normal distribution. To
derive an approximation for the sample maximum,
max n = max{zn }, we apply the Fisher-TippettGnedenko theorem to the Gaussian distribution
and obtain that
"
#
max N
(14)
lim P r ob
x = G[x],
N

where
x

G[x] = ee is the CDF for the Standard

Gumbel distribution.
h
i
i
h
1
1
= Z 1 1 N , = Z 1 1 N e1 ,
and Z 1 corresponds to the inverse of the
Standard Normals CDF.

We hope that our distinguished colleagues will

follow this humble attempt with ever-deeper and
more convincing analysis. We did not write this
paper to settle a discussion. On the contrary, our
wish is to ignite a dialogue among mathematicians
and a reflection among investors and regulators.
We should do well also to heed Newtons comment
after he lost heavily in the South Seas bubble; see
[21]:
For those who had realized big losses or
gains, the mania redistributed wealth. The
largest honest fortune was made by Thomas
Guy, a stationer turned philanthropist, who
owned 54,000 of South Sea stock in April
1720 and sold it over the following six weeks
for 234,000. Sir Isaac Newton, scientist,

master of the mint, and a certifiably rational

man, fared less well. He sold his 7,000 of
stock in April for a profit of 100 percent.
But something induced him to reenter the
market at the top, and he lost 20,000. I
can calculate the motions of the heavenly
bodies, he said, but not the madness of
people.

The normalizing constants (, ) are derived in

Resnick [22] and Embrechts et al. [5]. The limit of
the expectation of the normalized maxima from a
distribution in the Gumbel Maximum Domain of
Attraction (see Proposition 2.1(iii) in Resnick [22])
is
"
#
max N
(15)
lim E
= ,
N

where is the Euler-Mascheroni constant,

0.5772156649 . . .. Hence, for N sufficiently large,
the mean of the sample maximum of standard
normally distributed random variables can be
approximated by
(16)
E[max ]
N




1
1
+ = (1 )Z 1 1
+Z 1 1 e1
N
N
where N > 1.

Notices of the AMS

469

Proof of Proposition 3
Suppose there are two random samples (A and
B) of the same process {m }, where A and B
are of equal size and have means and standard
deviations A , B , A , B . A fraction of each
sample is called IS, and the remainder is called
OOS, where for simplicity we have assumed that
A
A
B
B
IS
= OOS
= IS
= OOS
. We would like to
understand the implications of a global constraint
A = B .
A
A
First, we note that A = IS
+ (1 )OOS
B
B
A
A
and B = IS
+ (1 )OOS
. Then, IS
> OOS
a
A
A
B
B
IS
> A a OOS
< A . Likewise, IS
> OOS
a
B
B
IS
> B a OOS
< B .
Second, because of the global constraint A = B ,
(1) A
(1) B
B
OOS = IS +
OOS
(1)
B
A
A
(OOS OOS ). Then, IS >

A
IS
+

A
B
and IS
IS
=
B
A
IS
a OOS
<

B
A
OOS
. We can divide this expression by IS
> 0,
with the implication that

(17)

A
B
A
B
SRIS
> SRIS
a SROOS
< SROOS

A
where we have denoted SRIS
=

A
IS
A
IS

, etc. Note that

Proof of Proposition 5
Suppose that we draw two samples (A and B) of
a first-order autoregressive process and generate
two subsamples of each. The first subsample is
called IS and is comprised of = 1, . . . , T , and the
second subsample is called OOS and is comprised
of = T + 1, . . . , T , with (0, 1) and T an
integer multiple of . For simplicity, let us assume
A
A
B
B
that IS
= OOS
= IS
= OOS
. From Proposition 4,
(18), we obtain
(22)

ET [mT ] mT = (1 T )( mT ).

A
B
A
B
Because 1 T > 0, IS
= IS
, SRIS
> SRIS
a
A
B
mT > mT . This means that the OOS of A
begins with a seed that is greater than the seed
A
that initializes the OOS of B. Therefore, mT
>
B
A
A
B
B
mT a ET [mT ] mT < ET [mT ] mT . Because
B
B
IS
= OOS
, we conclude that

(23)

A
B
A
B
SRIS
> SRIS
a SROOS
< SROOS

Reproducing the Results in Example 6

we did not have to assume that m is IID, thanks

to our assumption of equal standard deviations.
The same conclusion can be reached without
assuming equality of standard deviations; however,
the proof would be longer but no more revealing
(the point of this proposition is the implication of
global constraints).

Python code implementing the experiment described in A Practical Application can be found
at http://www.quantresearch.info/Software.
htm and at http://www.financial-math.org/
software/.

Proof of Proposition 4

We are indebted to the editor and two anonymous

referees who peer-reviewed this article for the
Notices of the American Mathematical Society. We
are also grateful to Tony Anagnostakis (Moore
Capital), Marco Avellaneda (Courant Institute, NYU),
Peter Carr (Morgan Stanley, NYU), Paul Embrechts
(ETH Zrich), Matthew D. Foreman (University
of California, Irvine), Jeffrey Lange (Guggenheim
Partners), Attilio Meucci (KKR, NYU), Natalia Nolde
(University of British Columbia ad ETH Zrich), and
Riccardo Rebonato (PIMCO, University of Oxford).
The opinions expressed in this article are the
authors and they do not necessarily reflect the
views of the Lawrence Berkeley National Laboratory,
Guggenheim Partners, or any other organization.
No particular investment or course of action is
recommended.

This proposition computes the half-life of a firstorder autoregressive process. Suppose there is a
random variable m that takes values of a sequence
of observations {1, . . . , }, where
(18)

m = (1 ) + m1 +

such that the random shocks are IID distributed

as N(0, 1). Then
lim E0 [m ] =

if and only if (1, 1). In particular, from Bailey

and Lpez de Prado [3] we know that the expected
value of this process at a particular observation
is
(19)

E0 [m ] = (1 ) + m0 .

Suppose that the process is initialized or reset

at some value m0 . We ask the question, how
many observations must pass before
+ m0
(20)
E0 [m ] =
?
2
Inserting (20) into (19) and solving for , we obtain
(21)

470

which implies the additional constraint that

(0, 1).

ln[2]
,
ln[]

Acknowledgments

References
[1] D. Bailey, J. Borwein, M. Lpez de Prado
and J. Zhu, The probability of backtest overfitting, 2013, working paper, available at
http://ssrn.com/abstract=2326253.
[2] D. Bailey and M. Lpez de Prado, The Sharpe ratio
efficient frontier, Journal of Risk 15(2) (2012), 344.
Available at http://ssrn.com/abstract=1821643.

Notices of the AMS

Volume 61, Number 5

[3]

[4]

[5]

[6]
[7]

[8]

[9]
[10]

[11]

[12]
[13]
[14]

[15]

[16]
[17]

[18]

[19]

[20]

[21]
[22]
[23]

[24]

[25]

, Drawdown-based stop-outs and the triple

penance rule, 2013, working paper. Available at
http://ssrn.com/abstract=2201302.
J. Doyle and C. Chen, The wandering weekday effect in major stock markets, Journal of Banking and
Finance 33 (2009), 13881399.
P. Embrechts, C. Klueppelberg and T. Mikosch,
Modelling Extremal Events, Springer-Verlag, New York,
2003.
R. Feynman, The Character of Physical Law, The MIT
Press, 1964.
J. Hadar and W. Russell, Rules for ordering uncertain
prospects, American Economic Review 59 (1969), 25
34.
L. Harris, Trading and Exchanges: Market Microstructure for Practitioners, Oxford University Press,
2003.
C. Harvey and Y. Liu, Backtesting, working paper,
SSRN, 2013, http://ssrn.com/abstract=2345489.
C. Harvey, Y. Liu, and H. Zhu, . . . and the crosssection of expected returns, working paper, SSRN,
2013, http://ssrn.com/abstract=2249314.
D. Hawkins, The problem of overfitting, Journal of
Chemical Information and Computer Science 44 (2004),
112.
Y. Hirsch, Dont Sell Stocks on Monday, Penguin Books,
1st edition, 1987.
J. Ioannidis, Why most published research findings
are false, PLoS Medicine 2(8), August 2005.
H. Krumholz Give the data to the people, New
York Times, February 2, 2014. Available at
http://www.nytimes.com/2014/02/03/opinion/
give-the-data-to-the-people.html.
D. Leinweber and K. Sisk, Event driven trading and
the new news, Journal of Portfolio Management 38(1)
(2011), 110124.
W. Leontief, Academic economics, Science Magazine
(July 9, 1982), 104107.
A. Lo, The statistics of Sharpe ratios, Financial Analysts Journal 58 4 (Jul/Aug, 2002). Available at
http://ssrn.com/abstract=377260.
M. Lpez de Prado and A. Peijan, Measuring the
loss potential of hedge fund strategies, Journal of
Alternative Investments 7(1), (2004), 731. Available
at http://ssrn.com/abstract=641702.
M. Lpez de Prado and M. Foreman, A mixture
of Gaussians approach to mathematical portfolio oversight: The EF3M algorithm, working paper,
RCC at Harvard University, 2012. Available at
http://ssrn.com/abstract=1931734.
J. Mayer, K. Khairy, and J. Howard, Drawing an
elephant with four complex parameters, American
Journal of Physics 78(6) (2010).
C. Reed, The damnd South Sea, Harvard Magazine
(MayJune 1999).
S. Resnick, Extreme Values, Regular Variation and
Point Processes, Springer, 1987.
J. Romano and M. Wolf, Stepwise multiple testing as
formalized data snooping, Econometrica 73(4) (2005),
12731282.
F. Schorfheide and K. Wolpin, On the use of holdout samples for model selection, American Economic
Review 102(3) (2012), 477481.
V. Stodden, D. Bailey, J. Borwein, R. LeVeque,
W. Rider, and W. Stein, Setting the default to
reproducible: Reproducibility in computational and

May 2014

[26]
[27]
[28]

[29]

experimental mathematics, February 2013. Available

at
http://www.davidhbailey.com/dhbpapers/
icerm-report.pdf.
G. Van Belle and K. Kerr, Design and Analysis of
Experiments in the Health Sciences, John Wiley & Sons.
H. White, A reality check for data snooping,
Econometrica 68(5), 10971126.
S. Weiss and C. Kulikowski, Computer Systems That
Learn: Classification and Prediction Methods from
Statistics, Neural Nets, Machine Learning and Expert
Systems, 1st edition, Morgan Kaufman,1990.
A. Wiles, Financial greed threatens the good name
of maths, The Times (04 Oct 2013). Available online
at http://www.thetimes.co.uk/tto/education/
article3886043.ece.

Notices of the AMS

471