volume 49

number 5

APRIL 2023
pm-research.com

A
N DING R T ICLE
TA
R NSTEIN FA
20
TS

24
OU

BO
B
THE

ZZI
JAC

OB
RD

S L A
EVY AW

Where Are the Factors

in Factor Investing?
Marcos López de Prado
Electronic copy available at: https://ssrn.com/abstract=4336002
L
everaging cutting-edge technologies in computational and data science—including sta-
tistics, artificial intelligence, machine learning, and high-performance computing—we
dedicate ourselves to addressing pressing questions in climate science, health, and the
digital economy. Rooted in our commitment to scientific excellence and the highest of stan-

It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.
dards, we champion interdisciplinary and international collaborations. Our drive is to break
down silos, foster global knowledge exchange, and aspire to be the best in our endeavors.

Marcos López de Prado

Marcos López de Prado is global head of quantitative research and development at the Abu
Dhabi Investment Authority (ADIA), a founding board member of ADIA Lab, and a professor of
practice at Cornell University’s School of Engineering. He has helped modernize finance for
the past 25 years, by pioneering machine learning and statistical inference methods, and by
implementing the Big Science paradigm of national laboratories at some of the largest invest-
ment corporations. In recognition of this work, Marcos has received various scientific and
industry awards, including the National Award for Academic Excellence (1999) by the Kingdom
of Spain, the Quant Researcher of the Year Award (2019) by Portfolio Management Research,
the Buy-Side Quant of the Year Award (2021) by Risk.net, and the Bernstein Fabozzi/Jacobs
Levy Award (2024) by The Journal of Portfolio Management. Since 2018, the Social Science
Research Network (SSRN) has ranked him as the most-read author in Economics.

Electronic copy available at: https://ssrn.com/abstract=4336002

 Where Are the Factors
in Factor Investing?
Marcos López de Prado

Marcos
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

López de Prado KEY FINDINGS

is global head of
Quantitative Research n The factor investing literature is logically incoherent: Factor models have a causal inter-
& Development at Abu pretation, and yet factor researchers never disclose the causal structure responsible for
Dhabi Investment Authority
(ADIA) in Abu Dhabi, United
the observed associations.
Arab Emirates; professor n As a consequence of this logical incoherence, factor models are likely misspecified and
of practice in the School
the reported anomalies are spurious.
of Engineering at Cornell
University in Ithaca, NY; n To address these pitfalls, factor researchers must embrace the modern tools of causal
and professor of practice inference.
in the Department of
Mathematics at Khalifa
University in Abu Dhabi,
ABSTRACT
United Arab Emirates. In this article, the author advocates for the use of causal graphs to modernize the field of
marcos.lopezdeprado
@adia.ae
factor investing and set it on a logically coherent foundation. To do this, first he introduces
the concepts of association and causation. Second, he explains the use of causal graphs
and the real (causal) meaning of the ceteris paribus assumption that is so popular among
economists. Third, he explains how causal graphs help us estimate causal effects in obser-
vational (nonexperimental) studies. Fourth, he illustrates all of the earlier concepts with
Monte Carlo experiments. He concludes that the field of factor investing must embrace
causal graphs in order to wake up from its associational slumber.

A
sk your favorite factor researcher: What is the meaning of the estimated b in
a factor model Y = a + bX + e? You are almost guaranteed to get the same
answer enshrined in the most popular econometrics textbooks: b is the slope
of a regression line, and it therefore has a distributional (associational) interpretation,
not a causal one (see, for example, Ruud 2000; Kennedy 2008; Woolbridge 2009;
Hill, Griffiths, and Lim 2011; and Greene 2012).1
However, if it is true that b is only an associational attribute of the data, why do
researchers choose a particular specification, Y = a + bX + e over X = g + dY + z, where
Y represents excess returns and X represents the factors? Associational properties
are nondirectional (“X is associated with Y” implies that “Y is associated with X”),
thus an associational interpretation of b implies that one model is recoverable from
the other: ˆγ = −α ˆ δˆ = 1/β,
ˆ /β, ˆ and ζˆ = −εˆ /β.
ˆ The problem is, this is virtually never the
case in the factor investing literature because researchers estimate b with the least-
squares method. Under least squares, the estimate of b cannot be recovered from
the estimate of d. By estimating Y = a + bX + e with least squares, the researcher
asserts that the direction of dependence is from X to Y and not the other way around.

1
For a critical analysis of causality in econometrics textbooks, see Chen and Pearl (2013).

Electronic copy available at: https://ssrn.com/abstract=4336002

2 | Where Are the Factors in Factor Investing? April 2023

Consequently, the correct interpretation is that b is the linear effect that X has on Y,
a causal (nonassociational) concept.
Why do factor researchers use the least-squares method? One reason is that
least-squares estimates are unbiased, subject to the model’s specification being
correct. Unbiasedness is a crucial property for factor researchers: Should the bias
be so large that b is estimated with the wrong sign (ββ ˆ < 0 ), the portfolio would be
exposed to a risk with negative expected return. If factor researchers were merely
interested in associations, they would minimize the cross-validated errors by exploit-
ing the bias-variance trade-off, like machine learners do. So not only do researchers

It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.
believe that Y is a function of X (a causal concept), but they are also willing to sacrifice
as much predictive power (an associational concept) as necessary to remove all bias
from β̂. The catch is, to monetize the risk premium causally attributed to X, the model
must be correctly specified. At the very least, the errors must be exogenous causes
of Y, uncorrelated with X (the explicit exogeneity assumption). Assuming exogeneity
requires knowledge of the role that each variable plays in the causal graph. Specifica-
tion setting is the perilous maneuver in which the factor investing ship runs aground,
because factor investing researchers justify their specification choice through associ-
ational arguments, such as explanatory power, rather than through causal arguments.
There is a logical inconsistency at the heart of the factor investing literature:
On one hand, researchers attempt to compute unbiased β̂ and p-values in a way that
is consistent with a causal interpretation of the factors (López de Prado 2023, section
6.1). On the other hand, researchers almost never state a causal graph or falsifiable
causal mechanism under which the specification is correct, and the estimates are
unbiased (López de Prado 2023, section 6.3). The result of this causal confusion is
an academic literature where factors are not really factors (in the causal sense)2 and
where unfalsifiable spurious claims proliferate. Under these circumstances, investors
may feel compelled to ask: Where are the factors in factor investing?
In this article, I advocate for the use of causal graphs to modernize the field of
factor investing and set it on a logically coherent foundation. In order to do that, first
I must introduce the concepts of association and causation. Second, I explain the
use of causal graphs and the real (causal) meaning of the ceteris paribus assumption
that is so popular among economists. Third, I explain how causal graphs help esti-
mate causal effects in observational (nonexperimental) studies. Fourth, I illustrate
all of the earlier concepts with Monte Carlo experiments. Fifth, I conclude that the
field of factor investing must embrace causal graphs in order to wake up from its
associational slumber.

ASSOCIATION

Every student of statistics, and by extension econometrics, learns that associa-

tion does not imply causation. This statement, although superficially true, does not
explain why association exists and its relation to causation. Two discrete random
variables X and Y are statistically independent if and only if P[X = x, Y = y] = P[X = x]
P[Y = y], ∀x,y, where P[⋅] is the probability of the event described inside the squared
brackets. Conversely, two discrete random variables X and Y are said to be statistically
associated (or codependent) when, for some (x,y), they satisfy that P[X = x, Y = y] ≠
P[X = x]P[Y = y]. The conditional probability expression P[Y = y | X = x] = P[X = x, Y = y]/
P[X = x] represents the probability that Y = y among the subset of the population

2
The term factor investing is another misnomer. The word factor has its origin in the Latin language,
with the literal meaning of “doer” or “maker.” Semantically, a factor is a cause responsible, in total or
in part, for an effect. Ironically, the factor investing literature has not attempted to explain what does
or makes the observed cross section of expected returns.

Electronic copy available at: https://ssrn.com/abstract=4336002

 April 2023 The Journal of Portfolio Management | 3

where X = x. When two discrete variables are associated, observing the value of one
conveys information about the value of the other: P[Y = y | X = x] ≠ P[Y = y] or, equiv-
alently, P[X = x | Y = y] ≠ P[X = x]. For example, monthly drownings (Y) and ice cream
sales (X) are strongly associated because the probability that y people drown in a
month conditional on observing x ice cream sales on that same month does not equal
the unconditional probability of y drownings in a month for some (x,y). However, the
expression P[Y = y | X = x] ≠ P[Y = y] does not tell us whether ice cream sales cause
drownings. Answering that question requires the introduction of a more nuanced
concept than conditional probability: an intervention.
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

CAUSATION

A data-generating process is a physical process responsible for generating the

observed data, where the process is characterized by a system of structural equa-
tions. Within that system, a variable X is said to cause a variable Y when Y is a func-
tion of X. The structural equation by which X causes Y is called a causal mechanism.
Unfortunately, the data-generating process responsible for observations is rarely
known. Instead, researchers must rely on probabilities, estimated on a sample of
observations, to deduce the causal structure of a system. Probabilistically, a vari-
able X is said to cause a variable Y when setting the value of X to x increases the
likelihood that Y will take the value y. Econometrics lacks the language to represent
interventions, that is, setting the value of X (Chen and Pearl 2013). To avoid confusion
between conditioning by X = x and setting the value of X = x, Pearl (1995) introduced
the do-operator, do[X = x], which denotes the intervention that sets the value of X to x.
With this new notation, causation can be formally defined as follows: X = x causes
Y = y if and only if P[Y = y | do[X = x]] > P[Y = y].3 For example, setting ice cream sales
to x will not make y drownings more likely than its unconditional probability for any
pair (x,y); hence, ice cream sales are not a cause of drownings. In contrast, smoking
tobacco is a cause of lung cancer because the probability that y individuals develop
lung cancer among a collective where the level of tobacco smoking is set to x is
greater than the unconditional probability of y individuals developing lung cancer, for
some pair (x,y).4

CAUSAL GRAPHS

Variables X and Y may be part of a more complex system, involving additional

variables. The causal structure of a system can be represented through a directed
acyclic graph, also denoted a causal graph.5 Although a causal graph does not fully

3
At first, it may seem counterintuitive that causality is defined in terms of a strict inequality (>),
in contrast to the difference (≠) used to define association. The reason is there is no need to consider
the < case due to complementary probabilities. For example, let X = 1 represent receiving a vaccine
against COVID-19 and Y = 1 represent developing COVID-19. For an effective vaccine, two causal
statements are true. First, P[Y = 1|do[X = 1]] < P[Y = 1], which means that receiving the vaccine (X = 1)
reduces the likelihood of developing the disease (Y = 1). Second, P[Y = 0|do[X = 1]] > P[Y = 0], which
means that receiving the vaccine (X = 1) increases the likelihood of not developing the disease (Y = 0).
One statement cannot be true without the other, and the redundancy is resolved by picking the latter.
4
A variable X may be a necessary cause of Y, a sufficient cause of Y, a necessary-and-sufficient
cause of Y, or neither a necessary-nor-sufficient cause of Y (also known as a contributory cause). I do
not explain the difference in this article because it is not required for the discussion that follows.
5
The causal structure of a system can also be encoded in a directed cyclical graph; however,
cyclicality comes at a heavy cost: The joint probability cannot be factorized as a product of conditional
probabilities between ancestors and descendants only. For this reason, directed acyclic graphs are
strongly preferred.

Electronic copy available at: https://ssrn.com/abstract=4336002

4 | Where Are the Factors in Factor Investing? April 2023

EXHIBIT 1
Causal Graph of a Confounder (Z)
Panel A: Before a Do-Operation Panel B: After a Do-Operation

1 X X

Z Z 2
2
Y Y

It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.
characterize the data-generating process, it conveys topological information essential
to estimate causal effects. Causal graphs declare the variables involved in a system,
which variables influence each other, and the direction of causality (Pearl 2009, 12).
Causal graphs help visualize do-operations as the action of removing all arrows
pointing toward X in the causal graph so that the full effect on Y can be attributed
to setting X = x. This is the meaning of the ceteris paribus assumption, which is of
critical importance to economists.
The causal graph in Exhibit 1 tells us that Z causes X and Z causes Y. In the
language of causal inference, Z is a confounder because this variable introduces an
association between X and Y, even though there is no arrow between X and Y. For this
reason, this type of association is denoted noncausal. Following with the previous
example, weather (Z) influences ice cream sales (X) and the number of swimmers,
hence drownings (Y). The intervention that sets ice cream sales removes arrow (1)
because it gives full control of X to the researcher (X is no longer a function of Z),
while keeping all other things equal (literally, ceteris paribus). And because X does
not cause Y, setting X = x (e.g., banning the sale of ice cream, X = 0) has no effect
on the probability of Y = y. As shown later, noncausal association can occur for a
variety of additional reasons that do not involve confounders.
Five conclusions can be derived from the exposition. First, causality is an
extra-statistical (in the sense of beyond observational) concept, connected to
mechanisms and interventions, and distinct from the concept of association. As a
consequence, researchers cannot describe causal systems with the associational
language of conditional probabilities. Failure to use the do-operator has led to confu-
sion between associational and causal statements in econometrics and elsewhere.
Second, association does not imply causation; however, causation does imply asso-
ciation, as evidenced by an intervention (do-operation).6 Third, unlike association,
causality is directional, as represented by the arrows of the causal graph. The state-
ment “X causes Y” implies that P[Y = y | do [X = x]] > P[Y = y] but not that P[X = x | do
[Y = y]] > P[X = x]. Fourth, unlike association, causality is sequential. “X causes Y”
implies that the value of X is set first, and only after that, Y adapts. Fifth, the ceteris
paribus assumption simulates an intervention (do-operation), whose implications can
only be understood with knowledge of the causal graph. The causal graph shows what
other things are kept equal by the intervention.

PATHS

In a graph with three variables {X, Y, Z}, a variable Z is a confounder with respect
to X and Y when the causal relationships include a structure X ← Z → Y. A variable
Z is a collider with respect to X and Y when the causal relationships are reversed

6
Here, I am referring to direct causes (a single link in the causal graph). There are causal structures
in which one cause may cancel another, resulting in total causation without association.

Electronic copy available at: https://ssrn.com/abstract=4336002

 April 2023 The Journal of Portfolio Management | 5

EXHIBIT 2
Example of a Causal Graph That Satisfies the Backdoor Criterion
Panel A: Before Conditioning on Z Panel B: After Conditioning on Z
(shaded node)

1 X 1 X

Z Z
3 3
2 2
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

Y Y

(i.e., X → Z ← Y). A variable Z is a mediator with respect to X and Y when the causal
relationships include a structure X → Z → Y.
A path is a sequence of arrows and nodes that connect two variables X and Y,
regardless of the direction of causation. A directed path is a path where all arrows
point in the same direction. In a directed path that starts in X and ends in Z, X is
an ancestor of Z, and Z is a descendant of X. A path between X and Y is blocked if
either: (1) the path traverses a collider and the researcher has not conditioned on
that collider or its descendants, or (2) the researcher conditions on a variable in the
path between X and Y, where the conditioned variable is not a collider. Association
flows along any paths between X and Y that are not blocked. Causal association flows
along an unblocked directed path that starts in treatment X and ends in outcome Y,
denoted the causal path. Association implies causation only if all noncausal paths are
blocked. This is the deeper explanation of why association does not imply causation
and why causal independence does not imply statistical independence.

BACKDOOR ADJUSTMENT

A backdoor path between X and Y is an unblocked noncausal path that connects

those two variables. The term backdoor is inspired by the fact that this kind of paths
have an arrow pointing into the treatment (X). For example, Exhibit 2 (Panel A) contains
a backdoor path (colored in red, Y ← Z → X) and a causal path (colored in green, X
→ Y). Backdoor paths can be blocked by conditioning on a set of variables S that
satisfies the backdoor criterion. The backdoor criterion is useful when controlling for
observable confounders.7
A set of variables S satisfies the backdoor criterion with regard to treatment X
and outcome Y if the following two conditions are true: (i) conditioning on S blocks
all backdoor paths between X and Y, and (ii) S does not contain any descendants
of X. Then, S is a sufficient adjustment set, and the causal effect of X on Y can be
estimated as

P[Y = y|do[ X = x]] = ∑ s P [Y = y| X = x, S = s]P[S = s]

Intuitively, condition (i) blocks all noncausal paths, whereas condition (ii) keeps
open all causal paths. In Exhibit 2, the only sufficient adjustment set S is {Z}, because
conditioning on Z blocks that backdoor path Y ← Z → X, and Z is not a descendant
of X. The result is that the only remaining association is the one flowing through the

7
I use here the nomenclature popularized by Pearl (2009); however, this form of adjustment was
fully developed by Robins (1986) under the term g-formula.

Electronic copy available at: https://ssrn.com/abstract=4336002

6 | Where Are the Factors in Factor Investing? April 2023

EXHIBIT 3
Causal Graph with a Confounder Z
Panel A: Before Control Panel B: After Control

1 X 1 X

Z Z
2 2
Y Y

It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.
causal path, thus adjusting the observations in a way that simulates a do-operation
on X. In general, there can be multiple sufficient adjustment sets that satisfy the
backdoor criterion for any given graph.

MONTE CARLO EXPERIMENTS

Factor model specification errors can lead to false positives and false negatives.
This section presents three instances of causal structures in which the application of
standard econometric procedures leads to mistaking association with causation and
ultimately to spurious factor claims. Standard econometric procedures are expected
to perform equally poorly on more complex causal structures. The code for these
experiments is available at ssrn.com/abstract_id=4205613.

EXPERIMENT 1: FORK

Three variables {X, Y, Z} form a fork when variable Z is a direct cause of variable
X and variable Y (see Exhibit 3). Consider a researcher who wishes to model Y as a
function of X. In that case, Z is said to be a confounding variable because not con-
trolling for the effect of Z on X and Y will bias the estimation of the effect of X on Y.
Given a probability distribution P, the application of Bayesian network factorization
on the fork represented by Exhibit 3 yields8

P[ X, Y , Z ] = P[ Z ]P[ X |Z ]P[Y |Z ]

which implies a (noncausal) association between X and Y because

P[ X, Y ] = ∑ Z P[ Z ]P[ X |Z ]P[Y |Z ] ≠ P[ X ]P[Y ]

This is an example of noncausal association because X and Y are associated

through the backdoor path Y ← Z → X, even though there is no causal path between
X and Y. The effect of conditioning by Z is equivalent to simulating a do-operation
(an intervention) because it blocks the backdoor path, resulting in the conditional
independence of X and Y,

P[ X, Y , Z ]
P[ X, Y |Z ] = = P[ X |Z ]P[Y |Z ]
P[ Z ]

8
For an introduction to the calculus of Bayesian network factorization, see Pearl, Glymour, and
Jewell (2016, 29–32) and Neal (2020, 20–22).

Electronic copy available at: https://ssrn.com/abstract=4336002

 April 2023 The Journal of Portfolio Management | 7

EXHIBIT 4
False Positive Due to Missing Confounder Z

OLS Regression Results

Dep. Variable: Y R-squared: 0.254
Model: OLS Adj. R-squared: 0.253
Method: Least Squares F-statistic: 1697.
Date: Sat, 21 Jan 2023 Prob(F-statistic): 1.03e-319
Time: 21:36:07 Log-Likelihood: –8157.8
No. Observations: 5000 AIC: 1.632e+04
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

Df Residuals: 4998 BIC: 1.633e+04

Df Model: 1
Covariance Type: nonrobust
coef std err t P>|t| [0.025 0.975]
const 0.0254 0.017 1.450 0.147 –0.009 0.060
X 0.5085 0.012 41.201 0.000 0.484 0.533
Omnibus: 1.899 Durbin-Watson: 2.077
Prob (Omnibus): 0.387 Jarque-Bera(JB): 1.865
Skew: –0.014 Prob(JB): 0.394
Kurtosis: 2.910 Cond. No. 1.42

NOTES: Results computed using python’s statsmodels package. For a description of the terms, consult the package documentation
at www.statsmodels.org. The code used in these experiments can be found in López de Prado (2023).

Conditioning by variable Z deconfounds P[X, Y] in this causal graph, however, not

in other causal graphs. The widespread notion that econometricians should condition
(or control) for all variables involved in a phenomenon is misleading. The precise
deconfounding variables are determined by do-calculus rules in general and by the
backdoor adjustment in this particular example. These conclusions can be verified
through the following numerical experiment. First, draw 5,000 observations from the
data-generating process characterized by the structural equation model

Zt := ξ t

X t := Zt + t

Yt := Zt + ζt

where {ξ t , t , ζt } are three independent random variables that follow a standard normal
distribution. Second, fit on the 5,000 observations the linear equation

Yt = α + βX t + ε t

Exhibit 4 reports the results of the least-squares estimate. Following the econo-
metric canon, a researcher will conclude that β̂ is statistically significant. Given the
causal content injected by the researcher through the least-squares model specifi-
cation, a statistically significant β̂ implies the statement “X causes Y,” not the state-
ment “X is associated with Y.” If the researcher intended to establish association, he
should have used an associational model, such as Pearson’s correlation coefficient
or orthogonal regression. At the same time, Exhibit 3 shows that there is no causal
path from X to Y. The claim of statistical significance is spurious because Y is not a
function of X, as implied by the model’s specification. This is the effect of missing
a single confounder.

Electronic copy available at: https://ssrn.com/abstract=4336002

8 | Where Are the Factors in Factor Investing? April 2023

EXHIBIT 5
Deconfounding through the Partial Correlations Method

OLS Regression Results

Dep. Variable: Y R-squared: 0.509
Model: OLS Adj. R-squared: 0.509
Method: Least Squares F-statistic: 2593.
Date: Sat, 21 Jan 2023 Prob(F-statistic): 0.00
Time: 21:37:35 Log-Likelihood: –7109.0
No. Observations: 5000 AIC: 1.422e+04

It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.
Df Residuals: 4997 BIC: 1.424e+04
Df Model: 2
Covariance Type: nonrobust
coef std err t P>|t| [0.025 0.975]
const 0.0105 0.014 0.740 0.460 –0.017 0.038
X –0.0100 0.014 –0.699 0.484 –0.038 0.018
Z 1.0291 0.020 51.036 0.000 0.990 1.069
Omnibus: 0.284 Durbin-Watson: 2.064
Prob (Omnibus): 0.868 Jarque-Bera(JB): 0.290
Skew: –0.018 Prob(JB): 0.865
Kurtosis: 2.993 Cond. No. 2.65

It is possible to remove the confounder-induced bias by adding Z as a regressor

(the partial correlations method)

Yt = α + βX t + γZt + ε t

Exhibit 5 reports the result of this adjustment. With the correct model specifica-
tion, the researcher will conclude that X does not cause Y.

EXPERIMENT 2: IMMORALITY

Three variables {X, Y, Z} form an immorality when variable Z is directly caused by

variable X and variable Y (see Exhibit 6). Consider a researcher who wishes to model
Y as a function of X. In that case, Z is said to be a collider variable.
Colliders should be particularly concerning to econometricians because controlling
for the effect of Z on X and Y biases the estimation of the effect of X on Y. Given a
probability distribution P, the application of Bayesian network factorization on the
immorality represented by Exhibit 6 yields

P[ X, Y , Z ] = P[ X ]P[Y ]P[ Z| X, Y ]

There is no association between X and Y because

P[ X, Y ] = ∑ Z P[ X ]P[Y ]P[ Z| X, Y ] = P[ X ]P[Y ]∑ Z P[ Z| X, Y ] = P[ X ]P[Y ]

However, conditioning by Z opens the backdoor path between X and Y that Z was
blocking (Y → Z ← X). The following analytical example illustrates this fact. Consider
the data-generating process

X t := t

Yt := ζt

Electronic copy available at: https://ssrn.com/abstract=4336002

 April 2023 The Journal of Portfolio Management | 9

EXHIBIT 6
Causal Graph with a Collider Z
Panel A: With Control Panel B: Without Control

X 1 X 1
Z Z
2 2
Y Y
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

Zt := X t + Yt + ξ t

where {ξ t , t , ζt } are three independent random variables that follow a standard normal
distribution. Then, the covariance between X and Y is

Cov [ X, Y ] = E [( X − E [ X ])(Y − E [Y ])] = E [ XY ] = E [ X ]E [Y ] = 0

The problem is, a researcher who (wrongly) conditions on Z will find a negative
covariance between X and Y, even though there is no causal path between X and Y
because

1
Cov [ X, Y |Z ] = −
3

Compare the causal graph in Exhibit 6 with the causal graph in Exhibit 3. Exhibit 3 has
a structure X ← Z → Y, where not controlling for confounder Z results in undercontrolling.
The direction of causality is reversed in Exhibit 6, transforming the confounder into a
collider. In the structure X → Z ← Y, controlling for Z results in overcontrolling. This is an
instance of Berkson’s fallacy, whereby a noncausal association is observed between two
independent variables as a result of conditioning on a collider (Pearl 2009, 17).
This finding is problematic for econometricians because the direction of causal-
ity cannot always be solely determined by observational studies (Peters, Janzing,
and Scholkopf 2017, 44–45), and solving the confounder-collider conundrum often
requires the injection of extra-statistical (beyond observational) information. Causal
graphs inject the required extra-statistical information by making explicit assumptions
that complement the information contributed by observations.9 Accordingly, the statis-
tical and econometric mantra “data speaks for itself” is in fact misleading, because
two econometricians who rely solely on observational evidence can consistently reach
contradicting conclusions from the analysis of the same data. With a careful selection
of colliders, a researcher can present evidence in support of any spurious investment
factor. The correct causal treatment of a collider is to indicate its presence and explain
why researchers should not control for it. A key takeaway is that researchers must
declare and justify the hypothesized causal graph that supports their chosen model
specification or else submit to the healthy scepticism of their peers.
We can verify the above conclusions with the following numerical experiment.
First, draw 5,000 observations from the preceding data-generating process.
Second, fit on the 5,000 observations the linear equation

Yt = α + βX t + γZt + ε t

9
In the absence of an interventional study or a natural experiment, the statement X causes Y is an
assumption, which may be consistent with, however not proved by, observational evidence.

Electronic copy available at: https://ssrn.com/abstract=4336002

10 | Where Are the Factors in Factor Investing? April 2023

EXHIBIT 7
False Positive Due to Adding Collider Z

OLS Regression Results

Dep. Variable: Y R-squared: 0.491
Model: OLS Adj. R-squared: 0.491
Method: Least Squares F-statistic: 2410.
Date: Sat, 21 Jan 2023 Prob(F-statistic): 0.00
Time: 21:38:49 Log-Likelihood: –5380.5
No. Observations: 5000 AIC: 1.077e+04

It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.
Df Residuals: 4997 BIC: 1.079e+04
Df Model: 2
Covariance Type: nonrobust
coef std err t P>|t| [0.025 0.975]
const –0.0077 0.010 –0.768 0.443 –0.027 0.012
X –0.5017 0.012 –40.421 0.000 –0.526 –0.477
Z 0.4959 0.007 69.420 0.000 0.482 0.510
Omnibus: 1.629 Durbin-Watson: 2.058
Prob (Omnibus): 0.443 Jarque-Bera(JB): 1.614
Skew: 0.010 Prob(JB): 0.446
Kurtosis: 2.914 Cond. No. 2.46

Exhibit 7 reports the results of the least-squares estimate. Following the econo-
metric canon, a researcher will conclude that β̂ is statistically significant. This claim
of statistical significance is spurious because Y is not a function of X, as implied by
the model’s specification. This is the effect of controlling for a collider.
We can remove the bias induced by collider Z by excluding that variable from the
model’s specification

Yt = α + βX t + ε t

Exhibit 8 reports the results of this adjustment. Note that the misspecified model
delivered higher explanatory power; hence, specification-searching would have mis-
led the researcher into a false positive. With the correct model specification, the
researcher will conclude that X does not cause Y.

EXPERIMENT 3: CHAIN

Three variables {X, Y, Z} form a chain when variable Z mediates the causal flow
from variable X to variable Y (see Exhibit 9). Consider a researcher who wishes to
model Y as a function of X. In that case, Z is said to be a mediator variable.
Given a probability distribution P, the application of Bayesian network factorization
on the chain represented by Exhibit 9 yields

P[ X, Y , Z ] = P[ X ]P[ Z| X ]P[Y |Z ]

which implies an association between X and Y because

P[ X, Y ] = ∑ Z P[ X ] P[ Z| X ]P[Y |Z ] ≠ P[ X ]P[Y ]

Electronic copy available at: https://ssrn.com/abstract=4336002

 April 2023 The Journal of Portfolio Management | 11

EXHIBIT 8
Debiasing by Removing Collider Z

OLS Regression Results

Dep. Variable: Y R-squared: 0.000
Model: OLS Adj. R-squared: –0.000
Method: Least Squares F-statistic: 0.2664
Date: Sat, 21 Jan 2023 Prob(F-statistic): 0.606
Time: 21:38:45 Log-Likelihood: –7068.5
No. Observations: 5000 AIC: 1.414e+04
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

Df Residuals: 4998 BIC: 1.415e+04

Df Model: 1
Covariance Type: nonrobust
coef std err t P>|t| [0.025 0.975]
const –0.0049 0.014 –0.350 0.726 –0.033 0.023
X 0.0072 0.014 0.516 0.606 –0.020 0.035
Omnibus: 4.168 Durbin-Watson: 2.023
Prob (Omnibus): 0.124 Jarque-Bera(JB): 4.157
Skew: 0.056 Prob(JB): 0.125
Kurtosis: 3.086 Cond. No. 1.02

EXHIBIT 9
Causal Graph with Mediator Z
Panel A: Before Control Panel B: After Control
1 2 1 2
X Z Y X Z Y

There is no backdoor path in Exhibit 9. This is an example of association with

causation because X and Y are associated only through the causal path mediated by Z.
Like in the case of a fork, the effect of conditioning by Z is equivalent to simulating
a do-operation (an intervention), resulting in the conditional independence of X and Y

P[ X, Y , Z ] P[ X ]P[ Z| X ]P[Y |Z ] P[ X, Z ]
P[ X, Y |Z ] = = = P[Y |Z ] = P[ X |Z ]P[Y |Z ]
P[ Z ] P[ Z ] P[ Z ]

The problem with conditioning on a mediator is that it may disrupt the very causal
association that the researcher wants to estimate (an instance of overcontrolling),
leading to a false negative. To make matters more complex, conditioning on a mediator
can also lead to a false positive. This statement can be verified through the follow-
ing numerical experiment. First, draw 5,000 observations from the data-generating
process characterized by the structural equation model

X t := t

Wt := ηt

Zt := X t + Wt + ξ t

Yt := Zt + Wt + ζt

Electronic copy available at: https://ssrn.com/abstract=4336002

12 | Where Are the Factors in Factor Investing? April 2023

where {ξ t , t , ζt , ηt } are four independent random variables that follow a standard

normal distribution. Exhibit 10 displays the relevant causal graph. Second, fit on the
5,000 observations the linear equation

Yt = α + βX t + γZt + ε t

Exhibit 11 reports the results of the least-squares estimate. Although it is true

that X causes Y (through Z), this result is still a false positive because the reported
association did not flow through the causal path X → Z → Y. The reason is Z also

It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.
operates as a collider to X and W, and controlling for Z has opened the backdoor path
X → Z ← W → Y. This is the reason β ˆ  0, despite of all effects being positive. This
phenomenon is known as the mediation fallacy, which involves conditioning on the
mediator when the mediator and the outcome are confounded (Pearl and MacKenzie
2018, 315). This experiment also illustrates Simpson’s paradox, which occurs when
an association is observed in several groups of data, but it disappears or reverses
when the groups are combined (Pearl, Glymour, and Jewell 2016, 1–6).

EXHIBIT 10
A Confounded Mediator (Z)
Panel A: With Control Panel B: Without Control

1 Z 1 Z
X X
3 3
2 2

W 4 W 4
Y Y

EXHIBIT 11
False Positive Due to Adding a Confounded Mediator Z

OLS Regression Results

Dep. Variable: Y R-squared: 0.786
Model: OLS Adj. R-squared: 0.786
Method: Least Squares F-statistic: 9160.
Date: Sat, 21 Jan 2023 Prob(F-statistic): 0.00
Time: 21:40:59 Log-Likelihood: –8114.0
No. Observations: 5000 AIC: 1.623e+04
Df Residuals: 4997 BIC: 1.625e+04
Df Model: 2
Covariance Type: nonrobust
coef std err t P>|t| [0.025 0.975]
const 0.0007 0.017 0.040 0.968 –0.033 0.035
X –0.5388 0.021 –25.130 0.000 –0.581 –0.497
Z 1.5075 0.012 122.161 0.000 1.483 1.532
Omnibus: 2.003 Durbin-Watson: 2.048
Prob (Omnibus): 0.367 Jarque-Bera(JB): 1.952
Skew: –0.045 Prob(JB): 0.377
Kurtosis: 3.036 Cond. No. 2.46

Electronic copy available at: https://ssrn.com/abstract=4336002

 April 2023 The Journal of Portfolio Management | 13

EXHIBIT 12
Deconfounding by Removing the Confounded Mediator

OLS Regression Results

Dep. Variable: Y R-squared: 0.146
Model: OLS Adj. R-squared: 0.145
Method: Least Squares F-statistic: 852.2
Date: Sat, 21 Jan 2023 Prob(F-statistic): 4.03e-173
Time: 21:44:28 Log-Likelihood: –11571.
No. Observations: 5000 AIC: 2.315e+04
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

Df Residuals: 4998 BIC: 2.316e+04

Df Model: 1
Covariance Type: nonrobust
coef std err t P>|t| [0.025 0.975]
const 0.0092 0.035 0.265 0.791 –0.059 0.077
X 1.0083 0.035 29.192 0.000 0.941 1.076
Omnibus: 2.574 Durbin-Watson: 2.024
Prob (Omnibus): 0.276 Jarque-Bera(JB): 2.605
Skew: –0.026 Prob(JB): 0.272
Kurtosis: 3.098 Cond. No. 1.02

Following the rules of do-calculus, the effect of X on Y in this causal graph can be
estimated without controls. The reason is the noncausal path through W is already
blocked by Z. Controlling for W is not strictly necessary to debias β̂; however, it
can help improve the precision of the estimates. The following model specification
produces an unbiased estimate of b:

Yt = α + βX t + ε t

Exhibit 12 reports the results. Note that the correct model specification has
much lower explanatory power: The adjusted R-squared drops from 0.786 to 0.146,
and the F-statistic drops from 9,160 to 852.2. A specification-searching researcher
would have chosen and reported the wrong model because it has higher explanatory
power, resulting in a misspecified model that misattributes risk and performance.
With the proper model specification, as informed by the declared causal graph, the
researcher correctly concludes that X causes Y and that βˆ  0.

CONCLUSIONS

Scientific theories should matter to investors for at least three reasons: First,
theories are a deterrent against selection bias because they force scientists to
justify their modeling choices, thus curtailing efforts to explain random variation.
A researcher who engages in p-hacking or backtest overfitting may build an ad hoc
theory that explains an observed random variation. However, other researchers will
use the theory to design an experiment in which the original random variation is not
observed. Second, causality is a necessary condition for investment efficiency. Causal
models allow investors to attribute risk and performance to the variables responsible
for a phenomenon. With proper attribution, investors can build a portfolio exposed
only to rewarded risks and aim for investment efficiency. In contrast, associational
models misattribute risks and performance, thus preventing investors from building
efficient portfolios. Third, causal models enable counterfactual reasoning, hence the

Electronic copy available at: https://ssrn.com/abstract=4336002

14 | Where Are the Factors in Factor Investing? April 2023

stress-testing of investment portfolios in a coherent and forward-looking manner (see

Rebonato 2010, Rebonato and Denev 2014, Denev 2015, and Rodríguez-Domínguez
2023). In contrast, associational models cannot answer counterfactual questions,
such as what would be the effect of Y on a not-yet-observed scenario X, thus exposing
those relying on associations to black-swan events.
Financial economists’ adoption of causal inference methods has the potential to
transform investing into a truly scientific discipline. Economists are best positioned to
inject, make explicit, and argue the extra-statistical information that complements and
enriches the work of statisticians. Machine learning methods can assist economists

It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.
in identifying the set of variables involved in the data-generating process, from which
a causal graph can be postulated, which in turn will support a chosen specification
(López de Prado 2022).
The new discipline of causal factor investing will be characterized by the adapta-
tion and adoption of tools from causal discovery and do-calculus to the study of the
risk characteristics that are responsible for differences in asset returns. Every year,
new alternative datasets become available at an increasing rate, allowing research-
ers to conduct natural experiments and other types of causal inference that were
not possible in the 20th century. Causal factor investing will serve a social purpose
beyond the reach of (associational) factor investing, helping asset managers fulfill
their fiduciary duties with the transparency and confidence that only the scientific
method can deliver. To achieve this noble goal, the dawn of scientific investing, the
factor investing community must first wake up from its associational slumber.

ACKNOWLEDGMENTS
I thank Alexander Lipton, Jean-Paul Villain, and Vincent Zoonekynd for numerous comments
and contributions. I also benefited from conversations with more ADIA colleagues than I can cite
here, as well as David H. Bailey (University of California, San Diego), David Easley (Cornell Univer-
sity), Campbell Harvey (Duke University), Miguel Hernán (Harvard University), John Hull (University
of Toronto), Maureen O’Hara (Cornell University), Riccardo Rebonato (EDHEC), Alessio Sancetta
(Royal Holloway, University of London), Horst Simon (Berkeley Lab), Sasha Stoikov (Cornell Uni-
versity), and Josef Teichmann (ETH Zurich). This article is an abridged version of the monograph
“Causal Factor Investing,” available at ssrn.com/abstract_id=4205613.

REFERENCES
Chen, B., and J. Pearl. 2013. “Regression and Causation: A Critical Examination of Six Economet-
rics Textbooks.” Real-World Economics Review 65: 2–20.

Denev, A. 2015. Probabilistic Graphical Models: A New Way of Thinking in Financial Modelling.
London, UK: Risk Books.

Greene, W. 2012. Econometric Analysis, 7th ed. Hoboken, NJ: Pearson Education.

Hill, R., W. Griffiths, and G. Lim. 2011. Principles of Econometrics, 4th ed. New York, NY: John
Wiley & Sons.

Kennedy, P. 2008. A Guide to Econometrics, 6th ed. Cambridge, MA: MIT Press.

López de Prado, M. 2022. “Machine Learning for Econometricians: The Readme Manual.”
The Journal of Financial Data Science 4 (3): 10–30.

——. 2023. Causal Factor Investing. Elements in Quantitative Finance. Cambridge University Press
(forthcoming). Pre-print available in https://ssrn.com/abstract_id=4205613.

Electronic copy available at: https://ssrn.com/abstract=4336002

 April 2023 The Journal of Portfolio Management | 15

Neal, B. 2020. “Introduction to Causal Inference: From a Machine Learning Perspective.” Course
lecture notes, December 17, https://www.bradyneal.com/causal-inference-course.

Pearl, J. 1995. “Causal Diagrams for Empirical Research.” Biometrika 82: 669–710.

——. 2009. Causality: Models, Reasoning and Inference, 2nd ed. New York, NY: Cambridge Uni-
versity Press.

Pearl, J., M. Glymour, and N. Jewell. 2016. Causal Inference in Statistics: A Primer. Hoboken, NJ:
John Wiley & Sons.
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

Pearl, J., and D. MacKenzie. 2018. The Book of Why. New York, NY: Basic Books.

Peters, L., D. Janzing, and B. Scholkopf. 2017. Elements of Causal Inference: Foundations and
Learning Algorithms. Cambridge, MA: MIT Press.

Rebonato, R. Coherent Stress Testing. New York, NY: John Wiley & Sons, 2010.

Rebonato, R., and A. Denev. 2014. Portfolio Management under Stress: A Bayesian-Net Approach
to Coherent Asset Allocation. Cambridge, UK: Cambridge University Press.

Robins, J. 1986. “A New Approach to Causal Inference in Mortality Studies with a Sustained Expo-
sure Period: Application to Control of a Healthy Worker Survivor Effect.” Mathematical Modelling
7 (9–12): 1393–1512.

Rodríguez-Domínguez, A. 2023. “Portfolio Optimization Based On Neural Networks Sensitivities

From Asset Dynamics Respect Common Drivers.” Machine Learning with Applications 11: 100447.

Ruud, P. 2000. An Introduction to Classical Econometric Theory. Oxford, UK: Oxford University Press.

Wooldridge, J. 2009. “Should Instrumental Variables Be Used as Matching Variables?” Techni-

cal report, Michigan State University, https://www.msu.edu/~ec/faculty/wooldridge/current%20
research/treat1r6.pdf.

Disclaimer
The views expressed in this article are the author’s and do not necessarily represent the opinions of the
organizations with which he is affiliated.

Electronic copy available at: https://ssrn.com/abstract=4336002