A Backtesting Protocol in the Era of Machine Learning *

Rob Arnott
Research Affiliates, Newport Beach, CA 92660, USA

Campbell R. Harvey
Fuqua School of Business, Duke University, Durham, NC 27708, USA
National Bureau of Economic Research, Cambridge, MA 02912, USA
Harry Markowitz
Harry Markowitz Company, San Diego, CA 92109, USA

ABSTRACT

Machine learning offers a set of powerful tools that holds considerable promise for investment
management. As with most quantitative applications in finance, the danger of misapplying
these techniques can lead to disappointment. One crucial limitation involves data availability.
Many of machine learning’s early successes originated in the physical and biological sciences, in
which truly vast amounts of data are available. Machine learning applications often require far
more data than are available in finance, which is of particular concern in longer-horizon
investing. Hence, choosing the right applications before applying the tools is important. In
addition, capital markets reflect the actions of people, which may be influenced by others’ actions
and by the findings of past research. In many ways, the challenges that affect machine learning
are merely a continuation of the long-standing issues researchers have always faced in
quantitative finance. While investors need to be cautious—indeed, more cautious than in past
applications of quantitative methods—these new tools offer many potential applications in
finance. In this article, the authors develop a research protocol that pertains both to the
application of machine learning techniques and to quantitative finance in general.

JEL: G11, G14, G17, C11, C58

Keywords: Machine Learning, Data Science, Data Mining, Backtesting, Overfitting,
Interpretable Classification, Interpretable Policy Design, Trading, Strategies,
Anomalies, Selection Bias, Research Protocol.

* Version: October 29, 2018. Corresponding author: Campbell R. Harvey

([email protected]). We have benefited from the comments of Frank Fabozzi, Marcos
López de Prado, and Joseph Simonian.

Electronic copy available at: https://ssrn.com/abstract=3275654

Introduction
Data mining is the search for replicable patterns, typically in large sets of data, from
which we can derive benefit. In empirical finance, “data mining” has a pejorative
connotation. We prefer to view data mining as an unavoidable element of research in
finance. We are all data miners, even if only by living through a particular history
that shapes our beliefs. In the past, data collection was costly and computing
resources were limited. As a result, researchers had to focus their efforts on
hypotheses that made the most sense. Today, both data and computing resources are
cheap, and in the era of machine learning, researchers no longer even need to specify
a hypothesis—the algorithm will supposedly figure it out.
Researchers are fortunate today to have a variety of statistical tools available, of
which machine learning, and the array of techniques it represents, is a prominent
and valuable one. Indeed, machine learning has already advanced our knowledge in
the physical and biological sciences, and has also been successfully applied to the
analysis of consumer behavior. All of these applications benefit from a vast amount
of data. With large data, patterns will emerge purely by chance. One of the big
advantages of machine learning is that it is hardwired to try to avoid overfitting by
constantly cross-validating discovered patterns. Again, this advantage performs well
in the presence of a large amount of data.
In investment finance, apart from tick data, the data are much more limited in scope.
Indeed, most equity-based strategies that purport to provide excess returns to a
passive benchmark rely on monthly and quarterly data. In this case, cross-validation
does not alleviate the curse of dimensionality. As a noted researcher remarked to one
of us:
[T]uning 10 different hyperparameters using k-fold cross-validation is a
terrible idea if you are trying to predict returns with 50 years of data (it
might be okay if you had millions of years of data). It is always necessary
to impose structure, perhaps arbitrary structure, on the problem you are
trying to solve.

Machine learning and other statistical tools, which have been impractical to use in
the past, hold considerable promise for the development of successful trading
strategies, especially in higher frequency trading. They might also hold great promise
in other applications such as risk management. Nevertheless, we need to be careful
in applying these tools. Indeed, we argue that given the limited nature of the standard
data that we use in finance, many of the challenges we face in the era of machine
learning are very similar to the issues we have long faced in quantitative finance in
general. We want to avoid backtest overfitting of investment strategies. And we want
a robust environment to maximize the discovery of new (true) strategies.
We believe the time is right to take a step back and to re-examine how we do our
research. Many have warned about the dangers of data mining in the past (e.g.,
Leamer, 1978; Lo and MacKinlay, 1990; and Markowitz and Xu, 1994), but the
problem is even more acute today. The playing field has leveled in computing
2

Electronic copy available at: https://ssrn.com/abstract=3275654

resources, data, and statistical expertise. As a result, new ideas run the risk of
becoming very crowded very quickly. Indeed, the mere publishing of an anomaly may
well begin the process of arbitraging the opportunity away.
Our paper develops a protocol for empirical research in finance. Research protocols
are popular in other sciences and are designed to minimize obvious errors, which
might lead to false discoveries. Our protocol applies to both traditional statistical
methods and modern machine-learning methods.
How Did We Get Here?
The early days of quantitative investing brought many impressive successes. Severe
constraints on computing and data led research to be narrowly focused. In addition,
much of the client marketplace was skeptical of quantitative methods. Consequently,
given the limited capital deployed on certain strategies, the risk of crowding was
minimal. But today the playing field has changed. Now almost everyone deploys
quantitative methods—even discretionary managers—and clients are far less averse
to quantitative techniques.
The pace of transformation is striking. Consider the Cray 2, the fastest
supercomputer in the world in the late 1980s and early 1990s (Bookman, 2017). It
weighed 5,500 pounds and, adjusted for inflation, it cost over US$30 million, in 2019
dollars. The Cray 2 performed an extraordinary (at the time) 1.9 billion operations
per second (Anthony, 2012). Today’s iPhone Xs is capable of 5 trillion operations per
second, and weighs just six ounces. Whereas a gigabyte of storage cost $10,000 in
1990, it costs only about a penny today. Also, a surprising array of data and
application software is available for free, or very nearly free. The barriers to entry in
the data-mining business, once lofty, are now negligible.
Sheer computing power and vast data are only part of the story. We have witnessed
many advances in statistics, mathematics, and computer science, notably in the fields
of machine learning and artificial intelligence. In addition, the availability of open
source software has also changed the game: it is no longer necessary to invest in (or
create) costly software. Essentially, anyone can download software and data, and
potentially access massive cloud computing to join the data-mining game.
Given the low cost of entering the data-mining business, investors need to be wary.
Consider the long–short equity strategy whose results are illustrated in Exhibit 1.
This is not a fake exhibit. 1 It represents a market-neutral strategy developed on
NYSE stocks from 1963 to 1988, then validated out of sample with even stronger
results over the years 1989 through 2015. The Sharpe ratio is impressive—over a 50-
year span, far longer than most backtests—and the performance is both economically
meaningful, generating nearly 6% alpha a year, and statistically significant.
Better still, the strategy has five very attractive practical features. First, it relies on
a consistent methodology through time. Second, performance in the most recent

1 Harvey and Liu (2014) present a similar exhibit with purely simulated (fake) strategies.

Electronic copy available at: https://ssrn.com/abstract=3275654

period does not trail off, indicating that the strategy is not crowded. Third, the
strategy does well during the financial crisis, gaining nearly 50%. Fourth, the
strategy has no statistically significant correlations with any of the well-known
factors, such as value, size, and momentum, or with the market as a whole. Fifth,
the turnover of the strategy is extremely low, less than 10% a year, so the trading
costs should be negligible.

Exhibit 1: Long–Short Market-Neutral Strategy Based on NYSE Stocks, 1963-

2015

This strategy might seem too good to be true. And it is. This data-mined strategy
forms portfolios based on letters in a company’s ticker symbol. For example, A(1)-
B(1) goes long all stocks with “A” as the first letter of their ticker symbol and short
all stocks with “B” as the first letter, equally weighting in both portfolios. The strategy
in Exhibit 1 considers all combinations of the first three letters of the ticker symbol,
denoted as S(3)-U(3). With 26 letters in the alphabet, and with two pairings on three
possible letters in the ticker symbol, thousands of combinations are possible. In

Electronic copy available at: https://ssrn.com/abstract=3275654

searching all potential combinations, 2 the chances of finding a strategy that looks
pretty good are pretty high!
A data-mined strategy that has a nonsensical basis is, of course, unlikely to fool
investors. We do not see ETFs popping up that offer “alpha-bets,” each specializing
in a letter of the alphabet. While a strategy with no economic foundation might have
worked in the past by luck, any future success would be equally random luck.
The strategy detailed in Exhibit 1, as preposterous as it seems, holds important
lessons in both data mining and machine learning. First, the S(3)-U(3) strategy was
discovered by brute force, not machine learning. Machine learning implementations
would carefully cross-validate the data by training the algorithm on part of the data
and then validating on another part of the data. As Exhibit 1 shows, however, in a
simple implementation when the S(3)-U(3) strategy was identified in the first
quarter-century of the sample, it would be “validated” in the second quarter-century.
In other words, it is possible that a false strategy can work in the cross-validated
sample. In this case the cross-validation is not randomized and, as a result, a single
historical path can be found.
The second lesson is that the data are very limited. Today, we have about 55 years
of high-quality equity data (or less than 700 monthly observations) for many of the
metrics in each of the stocks we may wish to consider. This tiny sample is far too
small for most machine learning applications, and impossibly small for advanced
approaches such as deep learning. Third, we have a strong prior that the strategy is
false: if it works, it is only because of luck. Machine learning, and particularly
unsupervised machine learning, does not impose economic principles. If it works, it
works in retrospect, but not necessarily in the future.
When data are limited, economic foundations become more important. Chordia,
Goyal, and Saretto (CGS) (2017) examine 2.1 million equity-based trading strategies
that use different combinations of indicators based on data from Compustat. CGS
carefully take data mining into account by penalizing each discovery (i.e., by
increasing the hurdle for significance). They identify 17 strategies that “survive the
statistical and economic thresholds.”
One of the strategies is labeled (dltis-pstkr)/mrc4. This strategy sorts stocks as
follows: The numerator is long-term debt issuance minus preferred/preference stock
redeemable. The denominator is minimum rental commitments—four years into the
future! The statistical significance is impressive, nearly matching the high hurdle
established by researchers at CERN when combing through quintillions of
observations to discover the elusive Higgs Boson (ATLAS Collaboration, 2012, and
CMS Collaboration, 2012). All 17 of the best strategies CGS identify have a similarly
peculiar construction, which – in our view and in the view of the authors of the paper

2Online tools, such as those available at http://datagrid.lbl.gov/backtest/index.php, generate fake strategies that
are as impressive as the one illustrated in Exhibit 1.

Electronic copy available at: https://ssrn.com/abstract=3275654

– leaves them with little or no economic foundation, even though they are based on
financial metrics.
Our message on the use of machine learning in backtests is one of caution and is
consistent with the admonitions of López de Prado (2018). Machine learning
techniques have been widely deployed for uses ranging from detecting consumer
preferences to autonomous vehicles, all situations that involve big data. The large
amount of data allows for multiple layers of cross-validation, which minimizes the
risk of overfitting. We are not so lucky in finance. Our data are limited. We cannot
flip a 4TeV switch at a particle accelerator and create trillions of fresh (not simulated)
out-of-sample data. But we are lucky in that finance theory can help us filter out
ideas that lack an ex ante economic basis. 3
We also do well to remember that we are not investing in signals or data; we are
investing in financial assets which represent partial ownership of a business, or of
debt, or of real properties, or of commodities. The quantitative community is
sometimes so focused on its models that we seem to forget that these models are crude
approximations of the real world, and cannot possibly reflect all of the nuances of the
assets that actually comprise our portfolios. The amount of noise may dwarf the
signal. Finance is a world of human beings, with emotions, herding behavior, and
short memories. And market anomalies – opportunities that are the main source of
intended profit for the quantitative community and our clients – are hardly static.
They change with time and are often easily arbitraged away. We ignore the gaping
chasm between our models and the real world at our peril.
The Winner’s Curse
Most of the quantitative community will acknowledge the many pitfalls in model
development. Considerable incentives exist to “beat the market” and to outdo the
competition. Countless thousands of models are tried. In contrast to our example with
ticker symbols, most of this research explores variables that most would consider
reasonable. The overwhelming number of these models do not work and are routinely
discarded. But some of them do appear to work. Of the models that do appear to work,
how many really do, and how many are just the product of overfitting?
Many opportunities exist for quantitative investment managers to make mistakes.
The most common mistake is being seduced by the data into thinking a model is
better than it is. This mistake has a behavioral underpinning. Researchers want their
model to work. They seek evidence to support their hypothesis—and all of the
rewards that come with it. They believe if they work hard enough, they will find the
golden ticket. This induces a type of “selection problem” in which the models that
make it through are likely to be a result of a biased selection process.
Models with strong results will be tested, modified, and retested, while models with
poor results will be quickly expunged. This creates two problems. One is that some

3Economists have an advantage over physicists in that societies are human constructs. Economists research what
humans have created, and as humans, we know how we created it. Physicists are not so lucky.

Electronic copy available at: https://ssrn.com/abstract=3275654

good models will fail in the test period, perhaps for reasons unique to the data set,
and will be forgotten. The other problem is that researchers seek a narrative to
justify a bad model that works well in the test period, again perhaps for reasons
irrelevant to the future efficacy of the model. These outcomes are false negatives and
false positives, respectively. Even more common than a false positive is an
exaggerated positive, an outcome that seems stronger, perhaps much stronger, than
it is likely to be in the future.
In other areas of science, this phenomenon is sometimes called the “winner’s curse.”
This is not the same winner’s curse as in auction theory. The researcher who is first
to publish the results of a clinical trial is likely to face the following situation. Once
the trial is replicated, one of three different outcomes can occur. 4 First (sadly the least
common outcome), the trial stands up to many replication tests, even with a different
sample, different time horizons, and other out-of-sample tests, and continues to work
after its original publication, roughly as well as in the backtests. Second, after
replication, the effect is far smaller than the original finding (for example, if microcap
stocks are excluded or if the replication is out of sample). The third outcome is the
worst: there is no effect and the research is eventually discredited. Once published,
models rarely work as well as in the backtest. 5
Can we avoid the winner’s curse? Not entirely, but with a strong research culture it
is possible to mitigate the damage of the winner’s curse.

Avoiding False Positives: A Protocol

The goal of investment management is to present strategies to clients that perform,
as promised, in live trading. Researchers want to minimize false positives, but to do
it in a way that does not miss too many good strategies. Protocols are widely used
both in scientific experiments and in practical applications. For example, every pilot
is now required to go through a protocol (sometimes called a checklist) before takeoff,
and airline safety has greatly improved in recent years. More generally, the use of
protocols has been shown to increase performance standards and prevent failure as
tasks become increasingly complex (e.g., Gawande, 2009). We believe that the use of
protocols for quantitative research in finance should become de rigueur, especially for
machine learning–based techniques, as computing power and process complexity
grow. Our goal is to improve investor outcomes in the context of backtesting.

4 In investing, two of these three outcomes have a twist to the winner’s curse: private gain and social loss. The
asset manager pockets the fees until the flaw of the strategy becomes evident, while the investor bears the losses
until the great reveal that it was a bad strategy all along.
5 See McLean and Pontiff (2016). Arnott, Beck, and Kalesnik (2016) examine eight of the most popular factors and

show an average return of 5.8% a year in the span before the factors’ publication and a return of only 2.4% after
publication. This loss of nearly 60% of the alpha on a long−short portfolio before any fees or trading costs is far
more slippage than most observers realize.

Electronic copy available at: https://ssrn.com/abstract=3275654

Many items in the protocol we suggest are not new (e.g., López de Prado, 2018), but
in this modern era of data science and machine learning, we believe it worthwhile to
specify best research practices in quantitative finance.
Category #1: Research Motivation
a) Establish an ex ante economic foundation.
Empirical research often provides the basis for the development of a theory. Consider
the relation between experimental and theoretical physics. Researchers in
experimental physics measure (generate data) and test the existing theories.
Theoretical physicists often use the results of experimental physics to develop better
models. This process is consistent with the concept of scientific method. A hypothesis
is developed, and the empirical tests attempt to find evidence inconsistent with the
hypothesis—so-called falsifiability. 6
The hypothesis provides a discipline that reduces the chance of overfitting.
Importantly, the hypothesis needs to have a logical foundation. For example, the
“alpha-bet” long–short trading strategy in Exhibit 1 has no theoretical foundation, let
alone a prior hypothesis. Bem (2011) published a study in a top academic journal that
“supported” the existence of extrasensory perception using over 1,000 subjects in 10
years of experiments. The odds of the results being a fluke were 74 billion to 1. They
were a fluke: the tests were not successfully replicated.
The researcher invites future problems by starting an empirical investigation without
an ex ante economic hypothesis. First, it is inefficient to even consider models or
variables without an ex ante economic hypothesis (such as scaling a predictor by
rental payments due in the fourth year). Second, no matter what the outcome,
without an economic foundation for the model, the researcher maximizes the chance
that the model will not work when taken into live trading. This is one of the
drawbacks of machine learning.
One of our recommendations is to carefully structure the machine learning problem
so that the inputs are guided by a reasonable hypothesis. Here is a simple example.
Suppose the researcher sets a goal of finding a long–short portfolio of stocks that
outperforms on a risk-adjusted basis, using the full spectrum of independent
variables available in Compustat and I/B/E/S. This is asking for trouble. With no
particular hypothesis, and even with the extensive cross-validation done in many
machine learning applications, the probability of a false positive is high.
b) Beware an ex post economic foundation.
It is also almost always a mistake to create an economic story—a rationale to justify
the findings—after the data mining has occurred. The story is often flimsy, and if the
data mining had delivered the opposite result, the after-the-fact story might easily
have been the opposite. An economic foundation should exist first, and a number of
6 One of the most damning critiques of theories in physics is to be “unfalsifiable.” Should we hold finance theories
to a lesser standard?

Electronic copy available at: https://ssrn.com/abstract=3275654

empirical tests should be designed to test how resilient that foundation is. Any
suspicion that the hypothesis was developed after looking at the data is an obvious
red flag.
Another subtle point: In other disciplines such as medicine, researchers often do not
have a pre-specified theory, and data exploration is crucial in shaping future clinical
trials. These trials provide the researcher with truly out-of-sample data. In finance
and economics, we do not have the luxury of creating a large out-of-sample test. It is
therefore dangerous to appropriate this exploratory approach into our field. We may
not jeopardize customer health, but we will jeopardize their wealth. This is
particularly relevant when it comes to machine learning methods, which were
developed for more data-rich disciplines.

Category #2: Multiple Testing and Statistical Methods

a) Keep track of what is tried.
Given 20 randomly selected strategies, one will likely exceed the two-sigma threshold
(t-statistic of 2.0 or above) purely by chance. As a result, the t-statistic of 2.0 is not a
meaningful benchmark if more than one strategy is tested. Keeping track of the
number of strategies tried is crucial as well as is measuring their correlations
(Harvey, 2017, and López de Prado, 2018). A bigger penalty in terms of threshold is
applied to strategies that are relatively uncorrelated. For example, if the 20
strategies tested had a near 1.0 correlation, then the process is equivalent to trying
only one strategy.
b) Keep track of combinations of variables.
Suppose the researcher starts with 20 variables and experiments with some
interactions, say (variable 1 x variable 2) and (variable 1 x variable 3). This single
interaction does not translate into only 22 tests (the original 20, plus two additional
interactions), but into 190 possible interactions. Any declared significance should
take the full range of interactions into account. 7
c) Beware the parallel universe problem.
Suppose a researcher develops an economic hypothesis and tests the model once, that
is, the researcher decides on the data, variables, scaling, and type of test—all in
advance. Given the single test, the researcher believes the two-sigma rule is
appropriate. But perhaps not. Think of being in 20 different parallel universes. In
each the researcher chooses a different model informed on the identical history. In

7 There are “20 choose 2” interactions, which is 20!/(18!2!).

Electronic copy available at: https://ssrn.com/abstract=3275654

each the researcher performs a single test. One of them works. Is it significant at two
sigma? Probably not.
Another way to think about this is to suppose that (in a single universe) the
researcher compiles a list of 20 variables to test for predictive ability. The first one
“works.” The researcher stops and claims they have done a single test. True, but the
outcome may just be lucky. Think of another researcher with the same 20 variables
who tests in a different order and only the last variable “works”. In this case, a
discovery at two sigma would be discarded, because a two-sigma threshold is too low
for 20 different tests.

Category #3: Sample Choice and Data

a) Define the test sample ex ante.
The training sample needs to be justified in advance. The sample should never
change after the research begins. For example, suppose the model “works” if the
sample begins in 1970, but does not “work” if the sample begins in 1960—in such a
case, the model does not work. A more egregious example would be to delete the global
financial crisis data, the tech bubble or the 1987 market crash, because they hurt the
predictive ability of the model. The researcher cannot massage the data to make the
model “work.”
b) Ensure data quality.
Flawed data can lead researchers astray. Any statistical analysis of the data is only
as good as the quality of the data that are input, especially in the case of certain
machine learning applications that try to capture nonlinearities. A nonlinearity
might simply be a bad data point.
The idea of garbage in/garbage out is hardly new. In the past, researchers would
literally “eyeball” smaller data sets and look for anomalous observations. Given the
size of today’s data sets, the human eyeball is insufficient. Cleaning the data before
employing machine learning techniques in the development of investment models is
crucial. Interestingly, some valuable data science tools have been developed to check
data integrity. These need to be applied as a first step.
c) Document choices in data transformations.
Manipulation of the input data, for example, volatility scaling or standardization, is
a choice, and is analogous to trying extra variables. The choices need to be
documented and ideally decided in advance. Furthermore, results need to be robust
to minor changes in the transformation. For example, given 10 different volatility-
scaling choices, if the one the researcher chose is the one that performed the best, this
is a red flag.

10

Electronic copy available at: https://ssrn.com/abstract=3275654

d) Do not arbitrarily exclude outliers.
By definition, outliers are influential observations for the model. Inclusion or
exclusion of influential observations can make or break the model. Ideally, a solid
economic case should be made for exclusion—before the model is estimated. In
general, no influential observations should be deleted. Assuming the observation is
based on valid data, the model should explain all data, not just a select number of
observations.
e) Select Winsorization level before constructing the model.
Winsorization is related to data exclusion. Winsorized data are truncated at a certain
threshold (e.g., truncating outliers to the 1% or 2% tails) rather than deleted.
Winsorization is a useful tool, because outliers can have an outsize influence on any
model. But, the choice to winsorize, and at which level, should be decided before
constructing the model. An obvious sign of a faulty research process is that the model
“works” at a winsoration level of 5%, but fails at 1%, and the 5% level is then chosen.

Category #4: Cross-Validation

a) Acknowledge out of sample is not really out of sample.
Researchers have lived through the hold-out sample and thus understand the history,
are knowledgeable about when markets rose and fell, and associate leading variables
with past experience. As such, no true out-of-sample data exists; the only true out of
sample is the live trading experience.
A better out-of-sample application is on freshly uncovered historical data; for
example, some researchers have tried to backfill the historical database of US
fundamental data to the 1920s. It is reasonable to assume these data have not been
data mined, because the data were not previously available in machine readable form.
But beware. Although these data were not previously available, well-informed
researchers are aware of how history unfolded and how macroeconomic events were
correlated with market movements. For those well versed on the history of markets,
these data are in sample in their own experience and in shaping their own prior
hypotheses. Even for those less knowledgeable, today’s conventional wisdom is
informed by past events.
As with the deep historical data, applying the model in different settings is a good
idea, but should be done with caution because correlations exist across countries. For
example, a data-mined (and potentially fake) anomaly that works in the US market
over a certain sample, may also work in Canada or the United Kingdom over the same
time span, given the correlation between these markets.

11

Electronic copy available at: https://ssrn.com/abstract=3275654

b) Understand iterated out of sample is not out of sample.
Suppose a model is successful in the in-sample period, but fails out of sample. The
researcher observes that the model fails for a particular reason. The researcher
modifies the initial model so it then works both in sample and out of sample. This is
no longer an out-of-sample test. It is overfitting.
c) Do not ignore trading costs and fees.
Almost all of the investment research published in academic finance ignores
transactions costs. 8 Even with modest transactions costs, the statistical
“significance” of many published anomalies essentially vanishes. Any research on
historical data needs to take transactions costs and, more generally, implementation
shortfall into account in both the in-sample and out-of-sample analysis (Arnott, 2006).

Category #5: Model Dynamics

a) Be aware of structural changes.
Certain machine applications have the ability to adapt through time. In economic
applications, structural changes, or nonstationarities, exist. This concern is largely
irrelevant in the physical and biological sciences. In finance, we are not dealing with
physical constants; we are dealing with human beings, and with changing preferences
and norms. Once again, the amount of available data is limiting, and the risk of
overfitting the dynamics of a relation through time is high.
b) Acknowledge the Heisenberg Uncertainty Principle and overcrowding.
In physics, the Heisenberg Uncertainty Principle states that we cannot know a
particle’s position and momentum simultaneously with precision. The more
accurately we know one characteristic, the less accurately we can know the other. A
similar principle can apply in finance. As we move from the study of past data into
the live application of research, market inefficiencies are hardly static. The cross-
validated relations of the past may seem powerful for reasons that no longer apply or
may dissipate merely because we are now aware of them and are trading based on
them.
Indeed, the mere act of studying and refining a model serves to increase the mismatch
between our expectations of a model’s efficacy and the true underlying efficacy of the
model—and that’s before we invest live assets, moving asset prices and shrinking the
efficacy of the models through our own collective trading.

8See Asness and Frazzini (2013). Hou, Xue, and Zhang (2017) show that most anomaly excess returns
disappear once microcaps are excluded.

12

Electronic copy available at: https://ssrn.com/abstract=3275654

c) Refrain from tweaking the model.
Suppose the model is running, but not doing as well as expected. Such a case should
not be a surprise because the backtest of the model is likely overfit to some degree.
It may be tempting to tweak the model, especially as a means to improve its fit in
recent, now in-sample data. While these modifications are a natural response to
failure, we should be fully aware that they will generally lead to further overfitting
of the model, and may lead to even worse live-trading performance.

Category #6: Model Complexity

a) Beware the curse of dimensionality.
Multidimensionality works against the viability of machine learning applications; the
reason is related to the limitations of data. Every new piece of information increases
dimensionality and requires more data. Recall the research of Chordia, Goyal and
Saretto (2017) who examine 2.1 million equity models based on Compustat data.
There are orders of magnitude more models than assets. With so many models, some
will work very well in sample.
Consider a model to predict the cross-section of stock prices. One reasonable variable
to explore is past stock prices (momentum), but many other variables, such as volume,
trailing volatility, bid–ask spread, and option skew, could be considered. As each
possible predictor variable is added, more data are required, but history is limited
and new data cannot be created or simulated. 9
Macroeconomic analysis provides another example. While most believe that certain
economic state variables are important drivers of market behavior and expected
returns, macroeconomic data, generally available on a monthly or quarterly basis,
are largely offside for most machine learning applications. Over the post-1960
period, 10 just over 200 quarterly observations and fewer than 700 monthly
observations exist.
Although the number of historical observations are limited for each time series, a
plethora of macroeconomic variables are available. If we select one or two to be
analyzed, we create an implicit data-mining problem, especially given that we have
lived through the chosen out-of-sample period.
b) Pursue simplicity and regularization.
Given data limitations, regularizing by imposing structure on the data is important.
Regularization is a key component of machine learning. It might be the case that a
machine learning model decides that a linear regression is the best model. But if a
more elaborate machine learning model beats the linear regression model, then it

9 Monte Carlo simulations are part of the toolkit, perhaps less used today than in the past. Of course,
simulations will produce results entirely consonant with the assumptions that drive the simulations.
10 Monthly macroeconomic data generally became available in 1959.

13

Electronic copy available at: https://ssrn.com/abstract=3275654

better win by an economically significant amount before switching to a more complex
model.
A simple analogy is a linear regression model of Y on X. The model in-sample fit can
almost always be improved by adding higher powers of X to the model. In out-of-
sample testing, the model with the higher powers of X will often perform poorly.
Current machine learning tools are designed to minimize the in-sample overfitting
by extensive use of cross-validation. Nevertheless, these tools may add complexity
(which is potentially nonintuitive) which leads to disappointing performance in the
true out-of-sample live trading. The greater the complexity, and the greater the
reliance on nonintuitive relationships, the greater the likely slippage between
backtest simulations and live results.
c) Seek interpretable machine learning.
It is important to look under the hood of any machine learning application. It cannot
be a black box. Investment managers should know what to expect with any machine
learning–based trading system. Indeed, an interesting new subfield in computer
science focuses on “interpretable classification” and “interpretable policy design” (e.g.,
Wang et al., 2016).

Category #7: Research Culture

a) Establish a research culture that rewards quality.
The investment industry rewards research that produces backtests with winning
results. If we do this in actual asset management, we create a toxic culture that
institutionalizes incentives to hack the data, to produce a seemingly good strategy.
Researchers should be rewarded for good science, not good results. A healthy culture
will also set the expectation that most experiments will fail to uncover a positive
result. Both management and researchers must have this common expectation.
b) Be careful with delegated research.
No one can perform every test that could potentially render an interesting result, so
researchers will often delegate. Delegated research needs to be carefully monitored.
Research assistants have an incentive to please their supervisor by presenting results
that support the supervisor’s hypothesis. This incentive can lead to a free-for-all data-
mining exercise that is likely to produce a failure when applied to live data.

Exhibit 2 condenses the foregoing discussion into a seven-point protocol for research
in quantitative finance.

14

Electronic copy available at: https://ssrn.com/abstract=3275654

Conclusions
The nexus of unprecedented computing power, free software, widely available data,
and advances in scientific methods provide us with unprecedented opportunities for
quantitative research in finance. Given these unprecedented capabilities, we believe
it is useful to take a step back and reflect on the investment industry’s research
process. It is naïve to think we no longer need economic models in the era of machine
learning. Given that the quantity (and quality) of data is relatively limited in finance,
machine learning applications face many of the same issues quantitative finance
researchers have struggled with for decades.
In this article, we have developed a research protocol for investment strategy
backtesting. The list is applicable to most research tools used in investment strategy
research—from portfolio sorts to machine learning. Our list of prescriptions and
proscriptions is long, but hardly exhaustive.
Importantly, the goal is not to eliminate all false positives. Indeed, that is easy – just
reject every single strategy. One of the important challenges that we face is satisfying
the dual objectives of minimizing false strategies but not missing too many good
strategies at the same time. The optimization of this trade-off is the subject of on-
going research (see Harvey and Liu, 2018).
At first reading, our observations may seem trivial and obvious. Importantly, our goal
is not to criticize quantitative investing. Our goal is to encourage humility, to
recognize that we can easily deceive ourselves into thinking we have found the “holy
grail.” Hubris is our enemy. A protocol is a simple step. Protocols can improve
outcomes, whether in a machine shop, an airplane cockpit, a hospital, or for an
investment manager. For the investment manager, our presumptive goal is an
investment process that creates the best possible opportunity to match or exceed
expectations when applied in live trading. Adopting this process is good for the client
and good for the reputation of the investment manager.

15

Electronic copy available at: https://ssrn.com/abstract=3275654

Exhibit 2. Seven-Point Protocol for Research in Quantitative Finance

1. Research Motivation
a) Does the model have a solid economic foundation?
b) Did the economic foundation or hypothesis exist before the research was conducted?

2. Multiple Testing and Statistical Methods

a) Did the researcher keep track of all models and variables that were tried (both successful and
unsuccessful) and are the researchers aware of the multiple-testing issue?
b) Is there a full accounting of all possible interaction variables if interaction variables are used?
c) Did the researchers investigate all variables set out in the research agenda or did they cut the
research as soon as they found a good model?

3. Data and Sample Choice

a) Do the data chosen for examination make sense? And, if other data are available, does it make
sense to exclude these data?
b) Did the researchers take steps to ensure the integrity of the data?
c) Do the data transformations, such as scaling, make sense? Were they selected in advance? And
are the results robust to minor changes in these transformations?
d) If outliers are excluded, are the exclusion rules reasonable?
e) If the data are winsorized, was there a good reason to do it? Was the winsorization rule chosen
before the research was started? Was only one winsorization rule tried (as opposed to many)?

4. Cross-Validation
a) Are the researchers aware that true out-of-sample tests are only possible in live trading?
b) Are steps in place to eliminate the risk of out-of-sample “iterations” (i.e., an in-sample model
that is later modified to fit out-of-sample data)?
c) Is the out-of-sample analysis representative of live trading? For example, are trading costs and
data revisions taken into account?

5. Model Dynamics
a) Is the model resilient to structural change and have the researchers taken steps to minimize
the overfitting of the model dynamics?
b) Does the analysis take into account the risk/likelihood of overcrowding in live trading?
c) Do researchers take steps to minimize the tweaking of a live model?

6. Complexity
a) Does the model avoid the curse of dimensionality?
b) Have the researchers taken steps to produce the simplest practicable model specification?
c) Is an attempt made to interpret the predictions of the machine learning model rather than
using it as a black box?

7. Research Culture
a) Does the research culture reward quality of the science rather than finding the winning
strategy?
b) Do the researchers and management understand that most tests will fail?
c) Are expectations clear (that researchers should seek the truth not just something that works)
when research is delegated?

16

Electronic copy available at: https://ssrn.com/abstract=3275654

References

Anthony, Sebastian. 2012. “The History of Supercomputers.” ExtremeTech.com (April 10).

Arnott, Robert. 2006. “Implementation Shortfall.” Financial Analysts Journal, vol. 62, no. 3
(May/June): 6–8.
Arnott, Robert, Noah Beck, and Vitali Kalesnik. 2016. “Timing ‘Smart Beta’ Strategies? Of Course!
Buy Low, Sell High!” Research Affiliates Publications (September).
Asness, Clifford, and Andrea Frazzini. 2013. “The Devil in HML’s Details.” Journal of Portfolio
Management, vol. 39, no. 4 (Summer): 49-68.
ATLAS Collaboration. 2012. “Observation of a New Particle in the Search for the Standard Model
Higgs Boson with the ATLAS Detector at the LHC.” Physics Letters B, vol. 716, no. 1 (September): 1–
29.
Bem, Daryl. 2011. “Feeling the Future: Experimental Evidence for Anomalous Retroactive Influences
on Cognition and Affect.” Journal of Personality and Social Psychology, vol. 100, no. 3 (March): 407–
425.
Bookman, Samantha. 2017. “15 Huge Supercomputers That Were Less Powerful Than Your
Smartphone.” TheClever.com (April 18).
Chordia, Tarun, Amit Goyal, and Alessio Saretto. 2017. “p-Hacking: Evidence from Two Million
Trading Strategies.” Swiss Finance Institute Research Paper No. 17-37. Available at SSRN.
CMS Collaboration. 2012. “Observation of a New Boson at a Mass of 125 GeV with the CMS
Experiment at the LHC.” Physics Letters B, vol. 716, no. 1 (September): 30–61.
Fabozzi, Frank, and Marcos López de Prado. 2018. “Being Honest in Backtest Reporting: A Template
for Disclosing Multiple Tests.” Journal of Portfolio Management, 45, No. 1 (Winter), 114-147.
Gawande, Atul. 2009. The Checklist Manifesto: How to Get Things Right. New York, NY: Henry Holt
and Sons.
Harvey, Campbell R., and Yan Liu. 2014. “Evaluating Trading Strategies.” Journal of Portfolio
Management, vol. 40, no. 5 (Special 40th Anniversary Issue):108–118.
Harvey, Campbell R. and Yan Liu, 2018. “False (and Missed) Discoveries in Financial Economics.”
Available on SSRN. https://ssrn.com/abstract=3073799
Harvey, Campbell R., 2017, Presidential Address: The Scientific Outlook in Financial
Economics, Journal of Finance 72, 1399-1440.
Hou, Kewei, Chen Xue, and Lu Zhang. 2017. “Replicating Anomalies. Available on SSRN.
https://www.ssrn.com/abstract=2961979
Leamer, Edward. 1978. Specification Searches: Ad Hoc Inference with Nonexperimental Data. New
York, NY: John Wiley & Sons.
Lo, Andrew, and A. Craig MacKinlay. 1990. “Data-Snooping Biases in Tests of Financial Asset Pricing
Models.” Review of Financial Studies, vol. 3, no. 3 (July):431–467.

López de Prado, Marcos. 2018. “The 10 Reasons Most Machine Learning Funds Fail.” Journal of
Portfolio Management, vol. 44, no. 6 (Special Issue Dedicated to Stephen A. Ross):120−133.
Markowitz, Harry, and Ban Lin Xu. 1994. “Data Mining Corrections.” Journal of Portfolio
Management, vol. 21, no. 1 (Fall): 60−69.
McLean, R. David, and Jeffrey Pontiff. 2016. “Does Academic Research Destroy Stock Return
Predictability?” Journal of Finance, vol. 71, no. 1 (February):5–32.

17

Electronic copy available at: https://ssrn.com/abstract=3275654

Wang, Tong, Cynthia Rudin, Finale Doshi-Velez, Yimin Liu, Erica Klampfl, and Perry MacNeille.
2017. “A Bayesian Framework for Learning Rule Sets for Interpretable Classification.” Journal of
Machine Learning Research, vol. 18: 1–37.

18

Electronic copy available at: https://ssrn.com/abstract=3275654