THE PROBABILITY OF BACKTEST

OVERFITTING
∗ †
David H. Bailey Jonathan M. Borwein
‡
Marcos López de Prado Qiji Jim Zhu§

February 27, 2015

Revised version: February 2015

∗
Lawrence Berkeley National Laboratory (retired), 1 Cyclotron Road, Berke-
ley, CA 94720, USA, and Research Fellow at the University of California,
Davis, Department of Computer Science. E-mail: david@davidhbailey.com; URL:
http://www.davidhbailey.com
†
Laureate Professor of Mathematics at University of Newcastle, Callaghan NSW
2308, Australia, and a Fellow of the Royal Society of Canada, the Australian
Academy of Science, the American Mathematical Society and the AAAS. E-mail:
jonathan.borwein@newcastle.edu.au; URL: http://www.carma.newcastle.edu.au/jon
‡
Senior Managing Director at Guggenheim Partners, New York, NY 10017, and Re-
search Affiliate at Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA.
E-mail: lopezdeprado@lbl.gov; URL: http://www.QuantResearch.info
§
Professor, Department of Mathematics, Western Michigan University, Kalamazoo, MI
49008, USA. Email: zhu@wmich.edu; URL: http://homepages.wmich.edu/~zhu/

Electronic copy available at: https://ssrn.com/abstract=2326253

 THE PROBABILITY OF BACKTEST OVERFITTING

Abstract
Many investment firms and portfolio managers rely on backtests
(i.e., simulations of performance based on historical market data) to
select investment strategies and allocate capital. Standard statistical
techniques designed to prevent regression overfitting, such as hold-
out, tend to be unreliable and inaccurate in the context of investment
backtests. We propose a general framework to assess the probabil-
ity of backtest overfitting (PBO). We illustrate this framework with
specific generic, model-free and nonparametric implementations in the
context of investment simulations, which implementations we call com-
binatorially symmetric cross-validation (CSCV). We show that CSCV
produces reasonable estimates of PBO for several useful examples.

Keywords. Backtest, historical simulation, probability of backtest over-

fitting, investment strategy, optimization, Sharpe ratio, minimum backtest
length, performance degradation.

JEL Classification: G0, G1, G2, G15, G24, E44.

AMS Classification: 91G10, 91G60, 91G70, 62C, 60E.

Acknowledgements. We are indebted to the Editor and three anony-

mous referees who peer-reviewed this article as well as a related article for the
Notices of the American Mathematical Society [1]. We are also grateful to
Tony Anagnostakis (Moore Capital), Marco Avellaneda (Courant Institute,
NYU), Peter Carr (Morgan Stanley, NYU), Paul Embrechts (ETH Zürich),
Matthew D. Foreman (University of California, Irvine), Ross Garon (SAC
Capital), Jeffrey S. Lange (Guggenheim Partners), Attilio Meucci (KKR,
NYU), Natalia Nolde (University of British Columbia and ETH Zürich) and
Riccardo Rebonato (PIMCO, University of Oxford) for many useful and
stimulating exchanges.

Sponsorship. Research supported in part by the Director, Office of Com-

putational and Technology Research, Division of Mathematical, Informa-
tion, and Computational Sciences of the U.S. Department of Energy, under
contract number DE-AC02-05CH11231 and by various Australian Research
Council grants.

Electronic copy available at: https://ssrn.com/abstract=2326253

 “This was our paradox: No course of action could be determined by
a rule, because every course of action can be made to accord with the
rule.” Ludwig Wittgenstein [36].

1 Introduction
Modern investment strategies rely on the discovery of patterns that can be
quantified and monetized in a systematic way. For example, algorithms can
be designed to profit from phenomena such as “momentum,” i.e., the ten-
dency of many securities to exhibit long runs of profits or losses, beyond
what could be expected from securities following a martingale. One advan-
tage of this systematization of investment strategies is that those algorithms
are amenable to “backtesting.” A backtest is a historical simulation of how
an algorithmic strategy would have performed in the past. Backtests are
valuable tools because they allow researchers to evaluate the risk/reward
profile of an investment strategy before committing funds.
Recent advances in algorithmic research and high-performance comput-
ing have made it nearly trivial to test millions and billions of alternative
investment strategies on a finite dataset of financial time series. While these
advances are undoubtedly useful, they also present a negative and often si-
lenced side-effect: The alarming rise of false positives in related academic
publications (The Economist [32]). This paper introduces a computational
procedure for detecting false positives in the context of investment strategy
research.
To motivate our study, consider a researcher who is investigating an al-
gorithm to profit from momentum. Perhaps the most popular technique
among Commodity Trading Advisors (CTAs) is to use so-called crossing-
moving averages to detect a change of trend in a security1 . Even for the
simplest case, there are at least five parameters that the researcher can fit:
Two sample lengths for the moving averages, entry threshold, exit threshold
and stop-loss. The number of combinations that can be tested over thou-
sands of securities is in the billions. For each of those billions of backtests,
we could estimate its Sharpe ratio (or any other performance statistic), and
determine whether that Sharpe ratio is indeed statistically significant at
a confidence level of 95%. Although this approach is consistent with the
Neyman-Pearson framework of hypothesis testing, it is highly likely that
false positives will emerge with a probability greater than 5%. The reason
1
Several technical tools are based on this principle, such as the Moving Average Con-
vergence Divergence (MACD) indicator.

Electronic copy available at: https://ssrn.com/abstract=2326253

is that a 5% false positive probability only holds when we apply the test
exactly once. However, we are applying the test on the same data multiple
times (indeed, billions of times), making the emergence of false positives
almost certain.
The core question we are asking is this: What constitutes a legitimate
empirical finding in the context of investment research? This may appear to
be a rather philosophical question, but it has important practical implica-
tions, as we shall see later in our discussion. Financial discoveries typically
involve identifying a phenomenon with low signal-to-noise ratio, where that
ratio is driven down as a result of competition. Because the signal is weak,
a test of hypothesis must be conducted on a large sample as a way of assess-
ing the existence of a phenomenon. This is not the typical case in scientific
areas where the signal-to-noise ratio is high. By way of example, consider
the apparatus of classical mechanics, which was developed centuries before
Neyman and Pearson proposed their theory of hypothesis testing. Newton
did not require statistical testing of his gravitation theory, because the signal
from that phenomenon dominates the noise.
The question of ‘legitimate empirical findings’ is particularly troubling
when researchers conduct multiple tests. The probability of finding false
positives increases with the number of tests conducted on the same data
(Miller [25]). As each researcher carries out millions of regressions (Sala-i-
Martin [28]) on a finite number of independent datasets without controlling
for the increased probability of false positives, some researchers have con-
cluded that ‘most published research findings are false’ (see Ioannidis [17]).
Furthermore, it is common practice to use this computational power to
calibrate the parameters of an investment strategy in order to maximize
its performance. But because the signal-to-noise ratio is so weak, often
the result of such calibration is that parameters are chosen to profit from
past noise rather than future signal. The outcome is an overfit backtest
[1]. Scientists at Lawrence Berkeley National Laboratory have developed
an online tool to demonstrate this phenomenon. This tool generates a time
series of pseudorandom returns, and then calibrates the parameters of an
optimal monthly strategy (i.e., the sequence of days of the month to be long
the security, and the sequence of days of the month to be short). After
a few hundred iterations, it is trivial to find highly profitable strategies
in-sample, despite the small number of parameters involved. Performance
out-of-sample is, of course, utterly disappointing. The tool is available at
http://datagrid.lbl.gov/backtest/index.php.
Backtests published in academic or practitioners’ publications almost
never declare the number of trials involved in a discovery. Because those

Electronic copy available at: https://ssrn.com/abstract=2326253

researchers have most likely not controlled for the number of trials, it is
highly probable that their findings constitute false positives ([1, 3]). Even
though researchers at academic and investment institutions may be aware
of these problems, they have little incentive to expose them. Whether their
motivations are to receive tenure or raise funds for a new systematic fund,
those researchers would rather ignore this problem and make their investors
or managers believe that backtest overfitting does not affect their results.
Some may even pretend that they are controlling for overfitting using inap-
propriate techniques, exploiting the ignorance of their sponsors, as we will
see later on when discussing the ‘hold-out’ method.
The goal of our paper is to develop computational techniques to control
for the increased probability of false positives as the number of trials in-
creases, applied to the particular field of investment strategy research. For
instance, journal editors and investors could demand researchers to estimate
that probability when a backtest is submitted to them.

Our approach. First, we introduce a precise characterization of the event

of backtest overfitting. The idea is simple and intuitive: For overfitting to
occur, the strategy configuration that delivers maximum performance in
sample (IS) must systematically underperform the remaining configurations
out of sample (OOS). Typically the principal reason for this underperfor-
mance is that the IS “optimal” strategy is so closely tied to the noise con-
tained in the training set that further optimization of the strategy becomes
pointless or even detrimental for the purpose of extracting the signal.
Second, we establish a general framework for assessing the probability
of the event of backtest overfitting. We model this phenomenon of backtest
overfitting using an abstract probability space in which the sample space
consist of pairs of IS and OOS test results.
Third, we set as null hypothesis that backtest overfitting has indeed
taken place, and develop an algorithm that tests for this hypothesis. For
a given strategy, the probability of backtest overfitting (PBO) is then eval-
uated as the conditional probability that this strategy underperforms the
median OOS while remaining optimal IS. While the PBO provides a direct
way to quantify the likelihood of backtest overfitting, the general framework
also affords us information to look into the overfitting issue from different
perspectives. For example, besides PBO, this framework can also be used to
assess performance decay, probability of loss, and possible stochastic domi-
nance of a strategy.
It is worth clarifying in what sense do we speak of a probability of back-

Electronic copy available at: https://ssrn.com/abstract=2326253

test overfitting. Backtest overfitting is a deterministic fact (either the model
is overfit or it is not), hence it may seem unnatural to associate a probability
to a non-random event. Given some empirical evidence and priors, we can
infer the posterior probability that overfitting has taken place. Examples of
this line of reasoning abound in information theory and machine learning
treatises, e.g. [23]. It is in this Bayesian sense that we define and estimate
PBO.
A generic, model-free, and nonparametric testing algorithm is desirable,
since backtests are applied to trading strategies produced using a great va-
riety of different methods and models. For this reason, we present a specific
implementation, which we call a combinatorially symmetric cross-validation
(CSCV). We show that CSCV produces reasonable estimates of PBO for
several useful examples.
Our CSCV implementation draws from elements in experimental math-
ematics, information theory, Bayesian inference, machine learning and deci-
sion theory to address the very particular problem of assessing the represen-
tativeness of a backtest. This is not an easy problem, as evidenced by the
scarcity of academic papers addressing a dilemma that most investors face.
This gap in the literature is disturbing, given the heavy reliance on backtests
among practitioners. One advantage of our solution is that it only requires
time series of backtested performance. We avoid the credibility issue of pre-
serving a truly out-of-sample test-set by not requiring a fixed “hold-out,”
and swapping all in-sample (IS) and out-of-sample (OOS) datasets. Our
approach is generic in the sense of not requiring knowledge of either the
trading rule or forecasting equation. The output is a bootstrapped distribu-
tion of OOS performance measure. Although in our examples we measure
performance using the Sharpe ratio, our methodology does not rely on this
particular performance statistic, and it can be applied to any alternative
preferred by the reader.
We emphasize that the CSCV implementation is only one illustrative
technique. The general framework is flexible enough to accommodate other
task-specific methods for estimating the PBO.

Comparisons to other approaches. Perhaps the most common ap-

proach to prevent overfitting among practitioners is to require the researcher
to withhold a portion of the available data sample for separate testing and
validation as OOS performance (this is known as the “hold-out” or “test
set” method). If the IS and OOS performance levels are congruent, the in-
vestor might decide to “reject” the hypothesis that the backtest is overfit.

Electronic copy available at: https://ssrn.com/abstract=2326253

The main advantage of this procedure is its simplicity. This approach is,
however, unsatisfactory for multiple reasons.
First, if the data is publicly available, it is quite likely that the researcher
has used the “hold-out” as part of the IS dataset. Second, even if no “hold-
out” data was used, any seasoned researcher knows well how financial vari-
ables performed over the time period covered by the OOS dataset, and that
information may well be used in the strategy design, consciously or not (see
Schorfheide and Wolpin [29]).
Third, hold-out is clearly inadequate for small samples—the IS dataset
will be too short to fit, and the OOS dataset too short to conclude anything
with sufficient confidence. Weiss and Kulikowski [34] argue that hold-out
should not be applied to an analysis with less than 1, 000 observations. For
example, if a strategy trades on a weekly basis, hold-out should not be used
on backtests of less than 20 years. Along the same lines, Van Belle and
Kerr [33] point out the high variance of hold-out estimation errors. If one
is unlucky, the chosen hold-out section may be the one that refutes a valid
strategy or supports an invalid strategy. Different hold-outs are thus likely
to lead to different conclusions.
Fourth, even if the researcher works with a large sample, the OOS anal-
ysis will need to consume a large proportion of the sample to be conclusive,
which is detrimental to the strategy’s design (see Hawkins [15]). If the OOS
is taken from the end of a time series, we are losing the most recent obser-
vations, which often are the most representative going forward. If the OOS
is taken from the beginning of the time series, the testing has been done on
arguably the least representative portion of the data.
Fifth, as long as the researcher tries more than one strategy configura-
tion, overfitting is always present (see Bailey et al. [1] for a proof). The
hold-out method does not take into account the number of trials attempted
before selecting a particular strategy configuration, and consequently hold-
out cannot correctly assess a backtest’s representativeness.
In short, the hold-out method leaves the investor guessing to what degree
the backtest is overfit. The answer to the question “is this backtest overfit?”
is not a true-or-false, but a non-null probability that depends on the number
of trials involved (input ignored by hold-out). In this paper we will present
a way to compute this probability.
Another approach popular among practitioners consists in modeling the
underlying financial variable by generating pseudorandom scenarios and
measuring the performance of the resulting investment strategy for those
scenarios (see Carr and López de Prado [6] for a valid application of this
technique). This approach has the advantage of generating a distribution of

Electronic copy available at: https://ssrn.com/abstract=2326253

outcomes, rather than relying on a single OOS performance estimate, as the
“hold-out” method does. The disadvantages are that the model that gener-
ates random series of the underlying variable may also be overfit, or may not
contain all relevant statistical features, and may need to be customized to
every variable (with large development costs). Some retail trading platforms
offer backtesting procedures based on this approach, such as by pseudoran-
dom generation of tick data by fractal interpolation.
Several procedures have been proposed to determine whether an econo-
metric model is overfit. See White [35], Romano et al. [27], Harvey et
al. [13] for a discussion in the context of Econometric models. Essentially
these methods propose a way to adjust the p-values of estimated regression
coefficients to account for the multiplicity of trials. These are valuable ap-
proaches when the trading rule relies on an econometric specification. That
is not generally the case, as discussed in Bailey et al. [1]. Investment strate-
gies in general are not amenable to characterization through a system of
algebraic equations. Regression-tree decision making, for example, requires
a hierarchy that only combinatorial frameworks like graph theory can pro-
vide, and which are beyond the geometric arguments used in econometric
models (see Calkin and López de Prado [4, 5]). On the other hand, the
approach proposed here shares the same philosophy in that both are trying
to assess the probability of overfitting.

Structure of the paper. The rest of the study is organized as follows:

Section 2 sets the foundations of our framework: we describe our general
framework for the backtest overfitting probability in Subsection 2.1 and
present the CSCV method for estimate this probability in Subsection 2.2.
Section 3 discusses other ways that our general framework can be used to
assess a backtest. Section 4 further discusses some of the features of the
CSCV method, and how it relates to other machine learning methods. Sec-
tion 5 lists some of the limitations of this method. Section 6 discusses a
practical application, and Section 7 summarizes our conclusions. We have
carried out several test cases to illustrate how the PBO compares to differ-
ent scenarios, and to assess the accuracy of our method using two alterna-
tive approaches (Monte Carlo Methods and Extreme Value Theory). The
interested reader can find the details of those studies following this link:
http://ssrn.com/abstract=2568435

Electronic copy available at: https://ssrn.com/abstract=2326253

2 The framework
2.1 Definition of overfitting in the context of strategy
selection
We first establish a measure theoretic framework in which the probability
of backtest overfitting and other statistics related to the issue of overfitting
can be rigorously defined. Consider a probability space (T , F, P rob) where
T represents a sample space of pairs of IS and OOS samples. We aim at
estimating the probability of overfitting for the following backtest strategy
selection process: select from N strategies labeled as (1, 2, . . . , N ) the ‘best’
one using backtesting according to a given performance measure, say, the
Sharpe ratio. Fixing a performance measure, we will use random vectors
R = (R1 , R2 , . . . , RN ) and R = (R1 , R2 , . . . , RN ) on (T , F, P rob) to rep-
resent the IS and OOS performance of the N strategies, respectively. For
a given sample c ∈ T , that is a concrete pair of IS and OOS samples, we
c
will use Rc and R to signify the performances of the N strategies on the
IS and OOS pair given by c. For most applications T will be finite and one
can choose to use the power set T as F. Moreover, often it makes sense in
this case to assume that the P rob is uniform on elements in T . However,
we do not make specific assumptions at this stage of general discussion so
as to allow flexibility in particular applications.
The key observation here is to compare the ranking of the selected strate-
gies IS and OOS. Therefore we consider the ranking space Ω consists of the
N ! permutations of (1, 2, . . . , N ) indicating the ranking of the N strategies.
Then we use random vectors r, r̄ to represent the ranking of the components
of R, R, respectively. For example, if N = 3 and the performance measure
is the Sharpe ratio, for a particular sample c ∈ T , Rc = (0.5, 1.1, 0.7) and
c
R = (0.6, 0.7, 1.3), then we have rc = (1, 3, 2) and r̄c = (1, 2, 3). Thus, both
r and r̄ are random vectors mapping (T , F, P rob) to Ω.
Now, we define backtest overfitting, in the context of investment strategy
selection alluded to above. We will need to use the following subset of Ω:
Ω∗n = {f ∈ Ω | fn = N }.
Definition 2.1. (Backtest Overfitting) We say that the backtest strategy
selection process overfits if a strategy with optimal performance IS has an
expected ranking below the median OOS. By the Bayesian formula and using
the notation above that is
N
X
E[rn | r ∈ Ω∗n ]P rob[r ∈ Ω∗n ] ≤ N/2. (2.1)
n=1

Electronic copy available at: https://ssrn.com/abstract=2326253

Definition 2.2. (Probability of Backtest Overfitting) A strategy with op-
timal performance IS is not necessarily optimal OOS. Moreover, there is a
non-null probability that this strategy with optimal performance IS ranks be-
low the median OOS. This is what we define as the probability of backtest
overfit (PBO). More precisely,
N
X
P BO = P rob[rn < N/2 | r ∈ Ω∗n ]P rob[r ∈ Ω∗n ]. (2.2)
n=1

In other words, we say that a strategy selection process overfits if the

expected performance of the strategies selected IS is less than the median
performance rank OOS of all strategies. In that situation, the strategy se-
lection process becomes in fact detrimental. Note that in this context IS
corresponds to the subset of observations used to select the optimal strat-
egy among the N alternatives. With IS we do not mean the period on
which the investment model underlying the strategy was estimated (e.g.,
the period on which crossing moving averages are computed, or a forecast-
ing regression model is estimated). Consequently, in the above definition
we refer to overfitting in relation to the strategy selection process, not a
strategy’s model calibration (e.g., in the context of regressions). That is the
reason we were able to define overfitting without knowledge of the strategy’s
underlying models, i.e., in a model-free and non-parametric manner.

2.2 The CSCV procedure

The framework Subsection 2.1 is flexible in dealing with the probability
of backtest overfitting and other statistical characterizations related to the
issue of overfitting. However, in order to quantify say the PBO for con-
crete applications we need a method to estimate the probability that was
abstractly defined in the previous section. Estimating the probability in a
particular application relies on schemes for selecting samples of IS and OOS
pairs. This section is devoted to establishing such a procedure, which we
name combinatorially symmetric cross-validation, abbreviated as (CSCV)
for convenience of reference.
Suppose that a researcher is developing an investment strategy. She con-
siders a family of system specifications and parametric values to be back-
tested, in an attempt to uncover the most profitable incarnation of that
idea. For example, in a trend-following moving average strategy, the re-
searcher might try alternative sample lengths on which the moving averages
are computed, entry thresholds, exit thresholds, stop losses, holding periods,

10

Electronic copy available at: https://ssrn.com/abstract=2326253

sampling frequencies, and so on. As a result, the researcher ends up running
a number N of alternative model configurations (or trials), out of which one
is chosen according to some performance evaluation criterion, such as the
Sharpe ratio.

Algorithm 2.3 (CSCV). We proceed as follows.

First, we form a matrix M by collecting the performance series from
the N trials. In particular, each column n = 1, . . . , N represents a vector of
profits and losses over t = 1, . . . , T observations associated with a particular
model configuration tried by the researcher. M is therefore a real-valued
matrix of order (T × N ). The only conditions we impose are that:

i) M is a true matrix, i.e. with the same number of rows for each column,
where observations are synchronous for every row across the N trials,
and

ii) the performance evaluation metric used to choose the “optimal” strat-
egy can be estimated on subsamples of each column.

For example, if that metric was the Sharpe ratio, we would expect that the
IID Normal distribution assumption could be maintained on various slices
of the reported performance. If different model configurations trade with
different frequencies, observations should be aggregated to match a common
index t = 1, . . . , T .
Second, we partition M across rows, into an even number S of disjoint
submatrices of equal dimensions. Each of these submatrices Ms , with s =
1, . . . , S, is of order (T /S × N ).
Third, we form all combinations CS of Ms , taken in groups of size S/2.
This gives a total number of combinations
    S/2−1
S S−1 S Y S−i
= = ... = (2.3)
S/2 S/2 − 1 S/2 S/2 − i
i=0

For instance, if S = 16, we will form 12, 780 combinations. Each combination
c ∈ CS is composed of S/2 submatrices Ms .
Fourth, for each combination c ∈ CS , we:

a) Form the training set J, by joining the S/2 submatrices Ms that con-
stitute c in their original order. J is a matrix of order (T /S)(S/2) ×
N ) = T /2 × N .

11

Electronic copy available at: https://ssrn.com/abstract=2326253

 b) Form the testing set J, as the complement of J in M . In other words,
J is the T /2 × N matrix formed by all rows of M that are not part of
J also in their original order. (The order in forming J and J does not
matter for some performance measures such as the Sharpe ratio but
does matter for others e.g. return maximum drawdown ratio).

c) Form a vector Rc of performance statistics of order N , where the nth

component Rnc of Rc reports the performance associated with the nth
column of J (the testing set). As before rank of the components of Rc
is denoted by rc the IS ranking of the N strategies.
c
d) Repeat c) with J replaced by J (the test set) to derive R and r̄c the
OOS performance statistics and rank of the N strategies, respectively.

e) Determine the element n∗ such that rnc ∗ ∈ Ω∗n∗ . In other words, n∗ is

the best performing strategy IS.

f) Define the relative rank of r̄nc ∗ by ω̄c := r̄nc ∗ /(N + 1) ∈ (0, 1). This is
the relative rank of the OOS performance associated with the strategy
chosen IS. If the strategy optimization procedure is not overfitting, we
should observe that r̄nc ∗ systematically outperforms OOS, just as rnc ∗
outperformed IS.
ω̄c
g) We define the logit λc = ln (1−ω̄ c)
. High logit values imply a consis-
tency between IS and OOS performances, which indicates a low level
of backtest overfitting.
Fifth, we compute the distribution of ranks OOS by collecting all the
λc , for c ∈ CS . Define the relative frequency at which λ occurred across all
CS by
X χ{λ} (λc )
f (λ) = , (2.4)
#(CS )
c∈CS

where χ is the characterization

R∞ function and #(CS ) signifies the number of
elements in CS . Then −∞ f (λ)dλ = 1. This concludes the procedure.
Figure 1 schematically represents how the combinations in CS are used
to produce training and testing sets, where S = 4. It shows the six combi-
nations of four subsamples A, B, C, D, grouped in two subsets of size two.
The first subset is the training set (or in-sample). This is used to deter-
mine the optimal model configuration. The second subset is the testing set
(or out-of-sample), on which the in-sample optimal model configuration is

12

Electronic copy available at: https://ssrn.com/abstract=2326253

 Figure 1: Generating the CS symmetric combination.

tested. Running the N model configurations over each of these combinations

allows us to derive a relative ranking, expressed as a logit. The outcome is
a distribution of logits, one per combination. Note that each training sub-
set combination is re-used as a testing subset and vice-versa (as is possible
because we split the data in two equal parts).

3 Overfit statistics
The framework introduced in Section 2 allows us to characterize the relia-
bility of a strategy’s backtest in terms of four complementary analysis:

1. Probability of Backtest Overfitting (PBO): The probability that the

model configuration selected as optimal IS will underperform the me-
dian of the N model configurations OOS.

2. Performance degradation: This determines to what extent greater per-

formance IS leads to lower performance OOS, an occurrence associated
with the memory effects discussed in Bailey et al. [1].

3. Probability of loss: The probability that the model selected as optimal

IS will deliver a loss OOS.

4. Stochastic dominance: This analysis determines whether the proce-

dure used to select a strategy IS is preferable to randomly choosing
one model configuration among the N alternatives.

3.1 Probability of backtest overfitting (PBO)

The PBO definedR in Section 2.1 may now be estimated using the CSCV
0
method with φ = −∞ f (λ)dλ. This represents the rate at which optimal IS
strategies underperform the median of the OOS trials. The analogue of r̄ in

13

Electronic copy available at: https://ssrn.com/abstract=2326253

medical research is the placebo given to a portion of patients in the test set.
If the backtest is truly helpful, the optimal strategy selected IS should out-
perform most of the N trials OOS. That is the case when λc > 0. For φ ≈ 0,
a low proportion of the optimal IS strategy outperformed the median of
trials in most of the testing sets indicating no significant overfitting. On the
flip side, φ ≈ 1 indicates high likelihood of overfitting. We consider at least
three uses for PBO: i) In general the value of φ provides us a quantitative
sense about the likelihood of overfitting. In accordance with standard ap-
plications of the Neyman-Pearson framework, a customary approach would
be to reject models for which PBO is estimated to be greater than 0.05. ii)
PBO could be used as a prior probability in Bayesian applications, where
for instance the goal may be to derive the posterior probability of a model’s
forecast. iii) We could compute the PBO on a large number of investment
strategies, and use those PBO estimates to compute a weighted portfolio,
where the weights are given by (1 − P BO), 1/P BO or some other scheme.

3.2 Performance degradation and probability of loss

Section 2.2 introduced the procedure to compute, among other results, the
pair (Rn∗ , Rn∗ ) for each combination c ∈ CS . Note that while we know that
Rn∗ is the maximum among the components of R, Rn∗ is not necessarily
the maximum among the components of R. Because we are trying every
combination of Ms taken in groups of size S/2, there is no reason to expect
the distribution of R to dominate over R. The implication is that, generally,
c
Rn∗ < max{R} ≈ max{R} = Rn∗ . For a regression Rn∗ = α + βRnc ∗ + εc ,
the β will be negative in most practical cases, due to compensation effects
described in Bailey et al. [1]. An intuitive explanation for this negative
slope is that overfit backtests minimize future performance: The model is
so fit to past noise, that it is often rendered unfit for future signal. And
the more overfit a backtest is, the more memory is accumulated against its
future performance.
It is interesting to plot the pairs (Rn∗ , Rn∗ ) to visualize how strong is
such performance degradation, andto obtain a more realistic range of at-
tainable performance OOS (see Figure 8). A particularly useful statistic is
c
the proportion of combinations with negative performance, P rob[Rn∗ < 0].
c
Note that, even if φ ≈ 0, P rob[Rn∗ < 0] could be high, in which case
the strategy’s performance OOS is probably poor for reasons other than
overfitting.
Figure 2 provides a graphical representation of i) Out-Of-Sample Perfor-
mance Degradation, ii) Out-Of-Sample Probability of Loss, and iii) Proba-

14

Electronic copy available at: https://ssrn.com/abstract=2326253

 Figures

Figure 2: Performance degradation and distribution of logits. Note that

c
even if φ ≈ 0, P rob[Rn∗ < 0] could be high, in which case the strategy’s
performance OOS is poor for reasons other than overfitting.

15

Electronic copy available at: https://ssrn.com/abstract=2326253

bility of Overfitting (PBO).
The upper plot of Figure 2 shows that pairs of (SR IS, SR OOS) for the
optimal model configurations selected for each subset c ∈ CS , which corre-
sponds to the performance degradation associated with the backtest of an
investment strategy. We can once again appreciate the negative relationship
between greater SR IS and SR OOS, indicating that at some point seeking
the optimal performance becomes detrimental. Whereas 100% of the SR IS
are positive, about 78% of the SR OOS are negative. Also, Sharpe ratios
IS range between 1 and 3, indicating that backtests with high Sharpe ratios
tell us nothing regarding the representativeness of that result.
We cannot hope escaping the risk of overfitting by exceeding some SR
IS threshold. On the contrary, it appears that the higher the SR IS, the
lower the SR OOS. In this example we are evaluating performance using
the Sharpe ratio, however, we again stress that our procedure is generic
and can be applied to any performance evaluation metric R (Sortino ratio,
Jensen’s Alpha, Probabilistic Sharpe Ratio, etc.). The method also allows
us to compute the proportion of combinations with negative performance,
c
P rob[Rn∗ < 0], which corresponds to analysis ii).
The lower plot of Figure 2 shows the distribution of logits for the same
strategy, with a PBO of 74%. It displays the distribution of logits, which
allows us to compute the probability of backtest overfitting (PBO). This
represents the rate at which optimal IS strategies underperform the median
of the OOS trials.
Figure 3 plots the performance degradation and distribution of logits
of a real investment strategy. Unlike in the previous example, the OOS
probability of loss is very small (about 3%), and the proportion of selected
(IS) model configurations that performed OOS below the median of overall
model configurations was only 4%.
The upper plot of Figure 3 plots the performance degradation associated
with the backtest of a real investment strategy. The regression line that goes
through the pairs of (SR IS, SR OOS) is much less steep, and only 3% of the
SR OOS are negative. The lower plot of Figure 3 shows the distribution of
logits, with a PBO of 0.04%. According to this analysis, it is unlikely that
this backtest is significantly overfit. The chances that this strategy performs
well OOS are much greater than in the previous example.

3.3 Stochastic dominance

A further application of the results derived in Section 2.2 is to determine
whether the distribution of Rn∗ across all c ∈ CS stochastically dominates

16

Electronic copy available at: https://ssrn.com/abstract=2326253

Figure 3: Performance degradation and distribution of logits for a real in-
vestment strategy.

17

Electronic copy available at: https://ssrn.com/abstract=2326253

over the distribution of all R. Should that not be the case, it would present
strong evidence that strategy selection optimization does not provide con-
sistently better OOS results than a random strategy selection. One reason
that the concept of stochastic dominance is useful is that it allows us to
rank gambles or lotteries without having to make strong assumptions re-
garding an individual’s utility function. See Hadar and Russell [11] for an
introduction to these matters.
In the context of our framework, first-order stochastic dominance oc-
curs if P rob[Rn∗ ≥ x] ≥ P rob[M ean(R) ≥ x] for all x, and for some
x, P rob[Rn∗ ≥ x] > P rob[M ean(R) ≥ x]. It can be verified visually by
checking that the cumulative distribution function of Rn∗ is not above the
cumulative distribution function of R for all possible outcomes, and at least
for one outcome the former is strictly below the latter. Under such circum-
stances, the decision maker would prefer the criterion used to produce Rn∗
over a random sampling of R, assuming only that her utility function is
weakly increasing.
A less demanding criterion
Rx is second-order stochastic dominance. This
requires that SD2[x] = −∞ (P rob[M ean(R) ≤ x] − P rob[Rn∗ ≤ x])dx ≥ 0
for all x, and that SD2[x] > 0 at some x. When that is the case, the
decision maker would prefer the criterion used to produce Rn∗ over a random
sampling of R, as long as she is risk averse and her utility function is weakly
increasing.
Figure 4 complements the analysis presented in Figure 2, with analysis
of stochastic dominance. Stochastic dominance allows us to rank gambles
or lotteries without having to make strong assumptions regarding an indi-
vidual’s utility function.
Figure 4 also provides an example of the cumulative distribution function
of Rn∗ across all c ∈ CS (red line) and R (blue line), as well as the second
order stochastic dominance (SD2[x], green line) for every OOS SR. In this
example, the distribution of OOS SR of optimized (IS) model configurations
does not dominate (to first order) the distribution of OOS SR of overall
model configurations.
This can be seen in the fact that for every level of OOS SR, the proportion
of optimized model configurations is greater than the proportion of non-
optimized, thus the probabilistic mass of the former is shifted to the left of
the non-optimized. SD2 plots the second order stochastic dominance, which
indicates that the distribution of optimized model configurations does not
dominate the non-optimized even according to this less demanding criterion.
It has been computed on the same backtest used for Figure 2. Consistent
with that result, the overall distribution of OOS performance dominates the

18

Electronic copy available at: https://ssrn.com/abstract=2326253

 Figure 4: Stochastic dominance (example 1).

OOS performance of the optimal strategy selection procedure, a clear sign

of overfitting.
Figure 5 provides a counter-example, based on the same real investment
strategy used in Figure 3. It indicates that the strategy selection proce-
dure used in this backtest actually added value, since the distribution of
OOS performance for the selected strategies clearly dominates the overall
distribution of OOS performance. (First-order stochastic dominance is a
sufficient condition for second-order stochastic dominance, and the plot of
SD2[x] is consistent with that fact.)

4 Features of the CSCV sampling method

Our testing method utilises multiple developments in the fields of machine
learning (combinatorial optimization, jackknife, cross-validation) and de-
cision theory (logistic function, stochastic dominance). Standard cross-
validation methods include k-fold cross-validation (K-FCV) and leave-one-
out cross-validation (LOOCV).
Now, K-FCV randomly divides the sample of size T into k subsamples

19

Electronic copy available at: https://ssrn.com/abstract=2326253

 Figure 5: Stochastic dominance (example 2).

of size T /k. Then it sequentially tests on each of the k samples the model
trained on the T − T /k sample. Although a very valid approach in many
situations, we believe that our procedure is more satisfactory than K-FCV
in the context of strategy selection. In particular, we would like to compute
the Sharpe ratio (or any other performance measure) on each of the k testing
sets of size T /k. This means that k must be sufficiently small, so that the
Sharpe ratio estimate is reliable (see Bailey and López de Prado [2] for a
discussion of Sharpe ratio confidence bands). But if k is small, K-FCV
will essentially reduce to a “hold-out” method, which we have argued is
unreliable. Also, LOOCV is a K-FCV where k = T . We are not aware of
any reliable performance metric computed on a single OOS observation.
The combinatorially symmetric cross-validation (CSCV) method we have
proposed in Section
2.2 differs from both K-FCV and LOOCV. The key idea
S
is to generate S/2 testing sets of size T /2 by recombining the S slices of
the overall sample of size T . This procedure presents a number of advan-
tages. First, CSCV ensures that the training and testing sets are of equal
size, thus providing comparable accuracy to the IS and OOS Sharpe ratios
(or any other performance metric that is susceptible to sample size).

20

Electronic copy available at: https://ssrn.com/abstract=2326253

 This is important, because making the testing set smaller than the train-
ing set (as hold-out does) would mean that we are evaluating with less accu-
racy OOS than the was used to choose the optimal strategy. Second, CSCV
is symmetric, in the sense that all training sets are re-used as testing sets
and vice versa. In this way, the decline in performance can only result from
overfitting, not arbitrary discrepancies between the training and testing sets.
Third, CSCV respects the time-dependence and other season-dependent
features present in the data, because it does not require a random allocation
of the observations to the S subsamples. We
avoid that requirement by
S
recombining the S subsamples into the S/2 testing sets. Fourth, CSCV
derives a non-random distribution of logits, in the sense that each logit is
deterministically derived from one item in the set of combinations CS . As
with jackknife resampling, running CSCV twice on the same inputs generates
identical results. Therefore, for each analysis, CSCV will provide a single
result, φ, which can be independently replicated and verified by another
user. Fifth, the dispersion of the distribution of logits conveys relevant
information regarding the robustness of the strategy selection procedure. A
robust strategy selection leads to a consistent OOS performance rankings,
which translate into similar logits.
Sixth, our procedure to estimate PBO is model-free, in the sense that it
does not require the researcher to specify a forecasting model or the defini-
tions of forecasting errors. It is also non-parametric, as we are not making
distributional assumptions on PBO. This is accomplished by using the con-
cept of logit, λc . A logit is the logarithm of odds. In our problem, the odds
are represented by relative ranks (i.e., the odds that the optimal strategy
chosen IS happens to underperform OOS). The logit function presents the
advantage of being the inverse of the sigmoidal logistic distribution, which
resembles the cumulative Normal distribution.
As a consequence, if ωc are distributed close to uniformly (the case when
the backtest appears to be informationless), the distribution of the logits
will approximate the standard Normal. This is important, because it gives
us a baseline of what to expect in the threshold case where the backtest
does not seem to provide any insight into the OOS performance. If good
backtesting results are conducive to good OOS performance, the distribution
of logits will be centered in a significantly positive value, and its left tail will
marginally cover the region of negative logit values, making φ ≈ 0.
A key parameter of our procedure is the value of S. This regulates the
number of submatrices Ms that will be generated, each of order
(T /S × N ),
S
and also the number of logit values that will be computed, S/2 . Indeed, S
must be large enough so that the number of combinations suffices to draw

21

Electronic copy available at: https://ssrn.com/abstract=2326253

inference. If S is too small, the left tail of the distribution of logits will be
underrepresented. On the other hand, if we believe that the performance
series is time-dependent and incorporates seasonal effects, S cannot be too
large, or the relevant time structure may be shuttered across the partitions.
For example, if the backtest includes more than six years of data, S =
24 generates partitions spanning over a quarter each, which would pre-
serve daily, weekly and monthly effects, while producing a distribution of
2, 704, 156 logits. By contrast, if we are interested in quarterly effects, we
have two choices: i) Work with S = 12 partitions, which will give us 924
logits, and/or ii) double T , so that S does not need to be reduced. The ac-
curacy of the procedure relies on computing a large number of logits, where
that number is derived in Equation (2.3). Because f (λ) is estimated as a
proportion of the number of logits, S needs to be large enough to generate
sufficient logits. For a proportion p̂ estimated on a sample of sizeqN , the
standard deviation of its expected value can be computed as σ[p̂] = p(1−p)
N
(see Gelman and Hill [10]).
q In other words, the standard deviation is highest
for p = 21 , with σ[p̂] = 1
4N . Fortunately, even a small number S gener-
ates a large number of logits. For example, S = 16 we will obtain 12, 780
logits (see Equation (2.3)), and σ[f (λ)] < 0.0045, with less than a 0.01 es-
timation error at 95% confidence level. Also, if M contains 4 years of daily
data, S = 16 would equate to quarterly partitions, and the serial correlation
structure would be preserved. For these two reasons, we believe that S = 16
is a reasonable value to use in most cases.
Another key parameter is the number of trials (i.e., the number of
columns in Ms ). Hold-out’s disregard for the number of trials attempted
was the reason we concluded it was an inappropriate method to assess a
backtest’s representativeness (see Bailey et al. [1] for a proof). N must be
large enough to provide sufficient granularity to the values of the relative
rank, ωc . If N is too small, ωc will take only a very few values, which will
translate into a very discrete number of logits, making f (λ) too discontinu-
ous, and adding estimation error to the evaluation of φ. For example, if the
investor is sensitive to values of φ < 1/10, it is clear that the range of values
that the logits can adopt must be greater than 10, and so N >> 10 is re-
quired. Other considerations regarding N will be discussed in the following
Section.
Finally, PBO is evaluated by comparing combinations of T /2 observa-
tions with their complements. But the backtest works with T observations,
rather than only T /2. Therefore, T should be chosen to be double of the
number of observations used by the investor to choose a model configuration

22

Electronic copy available at: https://ssrn.com/abstract=2326253

or to determine a forecasting specification.

5 Limitations and misuse

The general framework in Subsection 2.1 can be flexibly used to assess back-
test overfitting probability. Quantitative assessment, however, also relies on
methods for estimating the probability measure. In this paper, we focus
on one of such implementations: the CSCV method. This procedure was
designed to evaluate PBO under minimal assumptions and input require-
ments. In doing so, we have attempted to provide a very general (in fact,
model-free and non-parametric) procedure against which IS backtests can be
benchmarked. However, any particular implementation has its limitations
and the CSCV method is no exception. Below is a discussion of some of the
limitations of this method from the perspective of design and application.

5.1 Limitation in design

First, a key feature of the CSCV implementation is symmetry. In dividing
the total sample of the testing results into IS and OOS both the size and
method of division in CSCV are symmetric. The advantage of such an
symmetric division has been elaborated above. However, the complexity of
investment strategies and performance measures makes it unlikely that any
particular method will be a one size fits all solution. For some backtests
other methods, for example K-FCV, may well be better suited.
Moreover, symmetrically dividing the sample performance in to S sym-
metrically layered sub-samples also may not suitable for certain strategies.
For example, if the performance measure as a time series has a strong auto-
correlation, then such a division may obscure the characterization especially
when S is large.
Finally, the CSCV estimate of the probability measure assumes all the
sample statistics carries the same weight. Without knowing any prior infor-
mation on the distribution of the backtest performance measure this is, of
course, a natural and reasonable choice. If, however, one does have knowl-
edge regarding the distribution of the backtest performance measure, then
model-specific methods of dividing the sample performance measure and as-
signing different weights to different strips of the subdivision are likely to be
more accurate. For instance, if a forecasting equation was used to generate
the trials, it would be possible to develop a framework that evaluates PBO
particular to that forecasting equation.

23

Electronic copy available at: https://ssrn.com/abstract=2326253

5.2 Limitation in application
First, the researcher must provide full information regarding the actual trials
conducted, to avoid the file drawer problem (the test is only as good as
the completeness of the underlying information), and should test as many
strategy configurations as is reasonable and feasible. Hiding trials will lead to
an underestimation of the overfit, because each logit will be evaluated under
a biased relative rank ωc . This would be equivalent to removing subjects
from the trials of a new drug, once we have verified that the drug was not
effective on them. Likewise, adding trials that are doomed to fail in order to
make one particular model configuration succeed biases the result. If a model
configuration is obviously flawed, it should have never been tried in the first
place. A case in point is guided searches, where an optimization algorithm
uses information from prior iterations to decide what direction should be
followed next. In this case, the columns of matrix M should be the final
outcome of each guided search (i.e., after it has converged to a solution),
and not the intermediate steps.2 This procedure aims at evaluating how
reliable a backtest selection process is when choosing among feasible strategy
configurations. As a rule of thumb, the researcher should backtest as many
theoretically reasonable strategy configurations as possible.
Second, this procedure does nothing to evaluate the correctness of a
backtest. If the backtest is flawed due to bad assumptions, such as incorrect
transaction costs or using data not available at the moment of making a
decision, our approach will be making an assessment based on flawed infor-
mation.
Third, this procedure only takes into account structural breaks as long
as they are present in the dataset of length T . If a structural break occurs
outside the boundaries of the available dataset, the strategy may be over-
fit to a particular data regime, which our PBO has failed to account for
because the entire set belongs to the same regime. This invites the more
general warning that the dataset used for any backtest is expected to be
representative of future states of the modeled financial variable.
Fourth, although a high PBO indicates overfitting in the group of N
tested strategies, skillful strategies can still exists in these N strategies. For
example, it is entirely possible that all the N strategies have high but similar
Sharpe ratios. Since none of the strategies is clearly better than the rest,
PBO will be high. Here overfitting is among many ‘skillful’ strategies.
Fifth, we must warn the reader against applying CSCV to guide the
2
We thank David Aronson and Timothy Masters (Baruch College) for asking for this
clarification.

24

Electronic copy available at: https://ssrn.com/abstract=2326253

search for an optimal strategy. That would constitute a gross misuse of our
method. As Strathern [31] eloquently put it, “when a measure becomes a
target, it ceases to be a good measure.” Any counter-overfitting technique
used to select an optimal strategy will result in overfitting. For example,
CSCV can be employed to evaluate the quality of a strategy selection pro-
cess, but PBO should not be the objective function on which such selection
relies.

6 A practical application
Bailey et al. [1] present an example of an investment strategy that attempts
to profit from a seasonal effect. For the reader’s convenience, we reiterate
here how the strategy works. 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 month’s position. For example, we
could explore all nodes that span the interval [1, . . . , 22] for Entry day, the
interval [1, . . . , 20] for Holding period, the interval [0, . . . , 10] for Stop loss,
and {−1, 1} for Sign. The parameters 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, as discussed in the above cited paper, a time series of 1, 000 daily
prices (about 4 years) was generated by drawing from a random walk. Pa-
rameters were optimized (Entry day = 11, Holding period = 4, Stop loss =
-1 and Side = 1), resulting in an annualized Sharpe ratio of 1.27. Given the
elevated Sharpe ratio, we may conclude that this strategy’s performance is
significantly greater than zero for any confidence level. Indeed, the PSR-
Stat is 2.83, which implies a less than 1% probability that the true Sharpe
ratio is below 0 (see Bailey and López de Prado [2] for details). Figure 6
gives a graphical illustration of this example.
We have estimated the PBO using our CSCV procedure, and obtained
the results illustrated below. Figure 7 shows that approx. 53% of the SR
OOS are negative, despite all SR IS being positive and ranging between
1 and 2.2. Figure 8 plots the distribution of logits, which implies that,
despite the elevated SR IS, the PBO is as high as 55%. Consequently,

25

Electronic copy available at: https://ssrn.com/abstract=2326253

 Figure 6: Backtested performance of a seasonal strategy (example 1).

Figure 9 shows that the distribution of optimized OOS SR does not dominate
the overall distribution of OOS SR. This is consistent with the fact that
the underlying series follows a random walk, thus the serial independence
among observations makes any seasonal patterns coincidental. The CSCV
framework has succeeded in diagnosing that the backtest was overfit.
Second, we generated a time series of 1, 000 daily prices (about 4 years),
following a random walk. But unlike the first case, we have shifted the
returns of the first 5 random observations of each month to be centered at
a quarter of a standard deviation. This simulates a monthly seasonal effect,
which the strategy selection procedure should discover. Figure 10 plots
the random series, as well as the performance associated with the optimal
parameter combination: Entry day = 1, Holding period = 4, Stop loss =
-10 and Side = 1. The annualized Sharpe ratio at 1.54 is similar to the
previous (overfit) case (1.54 vs. 1.3).
The next three graphs report the results of the CSCV analysis, which
confirm the validity of this backtest in the sense that performance inflation
from overfitting is minimal. Figure 11 shows only 13% of the OOS SR to
be negative. Because there is a real monthly effect in the data, the PBO for

26

Electronic copy available at: https://ssrn.com/abstract=2326253

Figure 7: CSCV analysis of the backtest of a seasonal strategy (example 1):
Performance degradation.

this second case should be substantially lower than the PBO of the first case.
Figure 12 shows a distribution of logits with a PBO of only 13%. Figure
13 evidences that the distribution of OOS SR from IS optimal combinations
clearly dominates the overall distribution of OOS SR. The CSCV analysis
has this time correctly recognized the validity of this backtest, in the sense
that performance inflation from overfitting is small.
In this practical application we have illustrated how simple is to produce
overfit backtests when answering common investment questions, such as the
presence of seasonal effects. We refer the reader to [1, Appendix 4] for the
implementation of this experiment in 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,
and the like. Given the facility with which elevated Sharpe ratios can be
manufactured IS, the reader would be well advised to remain critical of
backtests and researchers that fail to report the PBO results.

27

Electronic copy available at: https://ssrn.com/abstract=2326253

Figure 8: CSCV analysis of the backtest of a seasonal strategy (example 1):
logit distrubution.

7 Conclusions
In [2] Bailey and López de Prado developed methodologies to evaluate the
probability that a Sharpe ratio is inflated (PSR), and to determine the
minimum track record length (MinTRL) required for a Sharpe ratio to be
statistically significant. These statistics were developed to assess Sharpe
ratios based on live investment performance and backtest track records. This
paper has extended this approach to present formulas and approximation
techniques for finding the probability of backtest overfitting.
To that end, we have proposed a general framework for modeling the
IS and OOS performance using probability. We define the probability of
backtested overfitting (PBO) as the probability that an optimal strategy IS
underperforms the mean OOS. To facilitate the evaluation of PBO for par-
ticular applications, we have proposed a combinatorially symmetric cross-
validation (CSCV) implementation framework for estimating this probabil-
ity. This estimate is generic, symmetric, model-free and non-parametric.
We have assessed the accuracy of CSCV as an approximation of PBO in

28

Electronic copy available at: https://ssrn.com/abstract=2326253

Figure 9: CSCV analysis of the backtest of a seasonal strategy (example 1):
Absent of dominance.

two different ways, on a wide variety of test cases. Monte Carlo simula-
tions show that CSCV applied on a single dataset provides similar results to
computing PBO on a large number of independent samples. We have also
directly computed PBO by deriving the Extreme Value distributions that
model the performance of IS optimal strategies. These results indicate that
CSCV provides reasonable estimates of PBO, with relatively small errors.
Besides estimating PBO, our general framework and its CSCV imple-
mentation scheme can also be used to deal with other issues related to
overfitting, such as performance degeneration, probability of loss and pos-
sible stochastic dominance of a strategy. On the other hand, the CSCV
implementation also has some limitations. This suggests that other imple-
mentation frameworks may well be more suitable, particularly for problems
with structure information.
Nevertheless, we believe that CSCV provides both a new and powerful
tool in the arsenal of an investment and financial researcher, and that it also

29

Electronic copy available at: https://ssrn.com/abstract=2326253

 Figure 10: Backtested performance of a seasonal strategy (example 2).

constitutes a nice illustration of our general framework for quantitatively

studying issues related to backtest overfitting. We certainly hope that this
study will raise greater awareness concerning the futility of computing and
reporting backtest results, without first controlling for PBO and MinBTL.

References
[1] Bailey, D., J. Borwein, M. López de Prado and J. Zhu, “Pseudo-mathematics
and financial charlatanism: The effects of backtest over fitting on out-of-sample
performance,” Notices of the AMS, 61 May (2014), 458–471. Online at http:
//www.ams.org/notices/201405/rnoti-p458.pdf.

[2] Bailey, D. and M. López de Prado, “The Sharpe Ratio Efficient Frontier,” Journal
of Risk, 15(2012), 3–44. Available at http://ssrn.com/abstract=1821643.

[3] Bailey, D. and M. López de Prado, “The Deflated Sharpe Ratio: Correcting for
Selection Bias, Backtest Overfitting and Non-Normality”, Journal of Portfolio Man-
agement, 40 (5) (2014), 94-107.

[4] Calkin, N. and M. López de Prado, “Stochastic Flow Diagrams”, Algorithmic Fi-
nance, 3(1-2) (2014) Available at http://ssrn.com/abstract=2379314.

30

Electronic copy available at: https://ssrn.com/abstract=2326253

Figure 11: CSCV analysis of the backtest of a seasonal strategy (example
2): monthly effect.

[5] Calkin, N. and M. López de Prado, “The Topology of Macro Financial Flows: An Ap-
plication of Stochastic Flow Diagrams”, Algorithmic Finance, 3(1-2) (2014). Avail-
able at http://ssrn.com/abstract=2379319.

[6] Carr, P. and M. López de Prado, “Determining Optimal Trading Rules without
Backtesting”, (2014) Available at http://arxiv.org/abs/1408.1159.

[7] Doyle, J. and C. Chen, “The wandering weekday effect in major stock markets,”
Journal of Banking and Finance, 33 (2009), 1388–1399.

[8] Embrechts, P., C. Klueppelberg and T. Mikosch, Modelling Extremal Events,

Springer Verlag, New York, 2003.

[9] Feynman, R., The Character of Physical Law, 1964, The MIT Press.

[10] Gelman, A. and J. Hill, Data Analysis Using Regression and Multilevel/Hierarchical
Models, 2006, Cambridge University Press, First Edition.

[11] Hadar, J. and W. Russell, “Rules for Ordering Uncertain Prospects,” American
Economic Review, 59 (1969), 25–34.

[12] Harris, L., Trading snf Exchanges: Market Microstructure for Practitioners, Oxford
University Press, 2003.

31

Electronic copy available at: https://ssrn.com/abstract=2326253

Figure 12: CSCV analysis of the backtest of a seasonal strategy (example
2): logit distribution.

[13] Harvey, C. and Y. Liu, “Backtesting”, SSRN, working paper, 2013. Available at
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2345489.

[14] Harvey, C., Y. Liu and H. Zhu, “...and the Cross-Section of Expected Returns,”
SSRN, 2013. Available at http://papers.ssrn.com/sol3/papers.cfm?abstract_
id=2249314.

[15] Hawkins, D., “The problem of overfitting,” Journal of Chemical Information and
Computer Science, 44 (2004), 10–12.

[16] Hirsch, Y., Don’t Sell Stocks on Monday, Penguin Books, 1st Edition, 1987.

[17] Ioannidis, J.P.A., “Why most published research findings are false.” PloS Medicine,
Vol. 2, No. 8,(2005) 696-701.

[18] Leinweber, D. and K. Sisk,“Event Driven Trading and the ‘New News’,” Journal of
Portfolio Management, 38(2011), 110–124.

[19] Leontief, W., “Academic Economics”, Science, 9 Jul 1982, 104–107.

[20] Lo, A., “The Statistics of Sharpe Ratios,” Financial Analysts Journal, 58 (2002),
July/August.

32

Electronic copy available at: https://ssrn.com/abstract=2326253

Figure 13: CSCV analysis of the backtest of a seasonal strategy (example
2): dominance.

[21] López de Prado, M. and A. Peijan, “Measuring the Loss Potential of Hedge Fund
Strategies,” Journal of Alternative Investments, 7 (2004), 7–31. Available at http:
//ssrn.com/abstract=641702.

[22] López de Prado, M. and M. Foreman, ”A Mixture of Gaussians Approach to Math-

ematical Portfolio Oversight: The EF3M Algorithm”, Quantitative Finance, forth-
coming, 2014. Available at http://ssrn.com/abstract=1931734.

[23] MacKay, D.J.C. “Information Theory, Inference and Learning Algorithms”, Cam-
bridge University Press, First Edition, 2003.

[24] Mayer, J., K. Khairy and J. Howard, “Drawing an Elephant with Four Complex
Parameters,” American Journal of Physics, 78 (2010), 648–649.

[25] Miller, R.G., Simultaneous Statistical Inference, 2nd Ed. Springer Verlag, New York,
1981. ISBN 0-387-90548-0.

[26] Resnick, S., Extreme Values, Regular Variation and Point Processes, Springer, 1987.

[27] Romano, J. and M. Wolf, “Stepwise multiple testing as formalized data snooping”,
Econometrica, 73 (2005), 1273–1282.

33

Electronic copy available at: https://ssrn.com/abstract=2326253

[28] Sala-i-Martin, X., “I just ran two million regressions.” American Economic Review.
87(2), May (1997).

[29] Schorfheide, F. and K. Wolpin, “On the Use of Holdout Samples for Model Selec-
tion,” American Economic Review, 102 (2012), 477–481.

[30] Stodden, V., Bailey, D., Borwein, J., LeVeque, R, Rider, W. and Stein, W., “Set-
ting the default to reproducible: Reproduciblity in computational and experimen-
tal mathematics,” February, 2013. Available at http://www.davidhbailey.com/
dhbpapers/icerm-report.pdf.

[31] Strathern, M., “Improving Ratings: Audit in the British University System,” Euro-
pean Review, 5, (1997) pp. 305-308.

[32] The Economist, “Trouble at the lab”, Oct. 2013 Available at

http://www.economist.com/news/briefing/21588057 -scientists -think-science-
self -correcting-alarming-degree-it-not-trouble.

[33] Van Belle, G. and K. Kerr, Design and Analysis of Experiments in the Health Sci-
ences, John Wiley and Sons, 2012.

[34] Weiss, S. and C. Kulikowski, Computer Systems That Learn: Classification and Pre-
diction Methods from Statistics, Neural Nets, Machine Learning and Expert Systems,
Morgan Kaufman, 1st Edition, 1990.

[35] White, H., “A Reality Check for Data Snooping,” Econometrica, 68 (2000), 1097–
1126.

[36] Wittgenstein, L.: Philosophical Investigations, 1953. Blackwell Publishing. Section

201.

34

Electronic copy available at: https://ssrn.com/abstract=2326253