SSRN Id4324450

Can Factor Investing Become Scientific?
Marcos López de Prado
Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450

Performance of Factor Investing
•
120
Bloomberg – Goldman Sachs Asset

Management US Equity Multi-Factor Index 115
(BBG code: BGSUSEMF <Index>)

–
110
It tracks the long/short performance of the
momentum, value, quality, and low-risk factors in
105
U.S. stocks
– Annualized Sharpe ratio: 0.29 (t-stat=1.16, p- 100
value=0.12), under 0% risk-free rate

– Average annualized return has been 1.13% 95
• This performance does not include:

Multi-Factor Index
– transaction costs
– market impact of order execution It takes over 31 years of daily observations for an
investment strategy with an annualized Sharpe
– cost of borrowing stocks for shorting positions
ratio of 0.29 to become statistically significant at
– management and incentive fees a 95% confidence level. After including costs,
performance is negative.
Electronic copy available at: https://ssrn.com/abstract=4324450 2
Classification: Public
Seminar’s Objective
• Every student of statistics learns that correlation does not imply causation
– Association is an observational property
– Causation is an interventional concept
• Causality plays a fundamental role in the scientific method
– Scientific theories are falsifiable statements of the form “𝑋 causes 𝑌 through mechanism 𝑀”
• Factor investing models remain at a pre-scientific stage
– Authors have failed to formulate falsifiable theories
• The “factor zoo” is a prime example of rampant spuriosity in investing:
– Type-A: Statistical flukes
– Type-B: Non-causal association
• This seminar proposes ways to solve the replication crisis that afflicts the factor
investing literature
– To read the full manuscript, visit: http://ssrn.com/abstract_id=4205613
3
Why Study Cause and Effect?

“Happy the man, who,
studying Nature’s laws,
thro’ known effects can
trace the secret cause”
The Second Book of the Georgics

Publius Vergilius Maro, “Virgil” (70 – 19 BC)

Around the year 1011
that Arab mathematician
Hasan Ibn Al Haytham
(965 - 1040) proposed a
scientific method for
deducing causal
mechanisms.

David Hume defined a cause-effect (causal)
relation as that "where, if the first object
had not been, the second never had
existed.“
An Enquiry concerning Human

Understanding. Sec. VII. (1748)
David Hume (1711 - 1776)

The Three Stages of the Scientific Method
Stage Statement Example
Phenomenological Smoking is associated with lung
𝑋 is associated with 𝑌
(induction) cancer
Smoking causes lung cancer
Theoretical 𝑋 causes 𝑌 through
through chemicals that damage
(abduction) mechanism 𝑀
the DNA
Refutation attempts:
Why do some smokers never
Falsification • 𝑋 does not cause 𝑌
suffer lung cancer?
Refutation • 𝑋 causes 𝑌, but not
Why do some non-smokers
(deduction) through mechanism 𝑀
suffer lung cancer?

Association vs. Causation

Paths
• A data-generating process can be represented
as a directed acyclic graph (DAG)
– Nodes are variables
– Arrows indicate the direction of dependence
• A path is a sequence of arrows and nodes that
A DAG is a directed graph with no directed
connect two variables 𝑋 and 𝑌, regardless of cycles.
the direction of causation
In the DAG above, there are two paths
• A directed path is a path where all arrows between 𝑋 and 𝑌:
point in the same direction
• In a directed path that starts in 𝑋 and ends in 𝑌 a) A directed path: 𝑋 → 𝑌
b) A non-directed path: 𝑋 ← 𝑍 → 𝑌
– 𝑋 is an ancestor of 𝑌, and
– 𝑌 is a descendant of 𝑋 In the DAG above, 𝑋 is a descendant of 𝑍,
and 𝑌 a descendant of 𝑋 and 𝑍.

Blocked Paths
• In a DAG with three variables 𝑋, 𝑌, 𝑍 , a variable 𝑍
is
– a confounder when the causal relationships include a
structure 𝑋 ← 𝑍 → 𝑌
– a collider when the causal relationships are reversed, i.e.
𝑋→𝑍←𝑌
– a mediator when the causal relationships include a In the above DAG:
structure 𝑋 → 𝑍 → 𝑌 • 𝑊 is a confounder to 𝑍 and 𝑌
• 𝑍 is a collider to 𝑋 and 𝑊
• 𝑍 is a mediator between 𝑋 and 𝑌
A path between 𝑋 and 𝑌 is blocked if either:
a) the path traverses a collider, and the researcher has not
The path 𝑋 → 𝑍 ← 𝑊 → 𝑌 is blocked by
conditioned on that collider or its descendants; or
𝑍.
b) the researcher conditions on a variable in the path between 𝑋
and 𝑌, where the conditioned variable is not a collider
The only unblocked path between 𝑋
and 𝑌 is the causal path, 𝑋 → 𝑍 → 𝑌.

What is Association?
• Association flows along an unblocked path
between 𝑋 and 𝑌
– Association is symmetric because paths do not follow the
direction of causation
• Probabilistically, two variables 𝑋 and 𝑌 are
associated when knowing the value of one
conveys information about the value of the other
∃𝒙, 𝒚|𝑷 𝒀 = 𝒚|𝑿 = 𝒙 ≠ 𝑷 𝒀 = 𝒚 Weather (𝑊) influences ice cream sales

(𝑋) and the number of swimmers (𝑍),
hence the number of drownings (𝑌).
• Statistical association is merely an observational
statement on the joint distribution of probability There is no directed path between 𝑋 and
𝑌, however 𝑋 and 𝑌 are associated (red-
– 𝑃 𝑌 = 𝑦|𝑋 = 𝑥 does not measure the effect of 𝑋 on 𝑌
dashed undirected edge), because of the
unblocked path 𝑋 ← 𝑊 → 𝑍 → 𝑌. 12
Electronic copy available at: https://ssrn.com/abstract=4324450
What is Causation?
• Causal association flows along an unblocked
directed path that starts in treatment 𝑋 and
ends in outcome 𝑌, denoted the causal path
• Let 𝑑𝑜 𝑋 = 𝑥 represent the do-operator on 𝑋
– This is an intervention that sets the value of 𝑋 to 𝑥,
hence 𝑋 is not influenced by any other variable
• Definition: 𝑿 causes 𝒀 iff 𝑷 𝒀|𝒅𝒐 𝑿 > 𝑷 𝒀
• Association implies causation only if all non-
causal paths are blocked A do-operation on 𝑋 removes arrow (1),
because 𝑋 is no longer a function of 𝑊,
• Causality is while keeping all other things equal
– an interventional (beyond observational) concept (“ceteris paribus”). As a result, there is
– asymmetric (directional) no unblocked path between 𝑋 and 𝑌,
and 𝑃 𝑌|𝑑𝑜 𝑋 = 𝑃 𝑌 .
– sequential: 𝑋 happens first, and then 𝑌 adapts
Blocking Non-Causal Paths

Interventional Studies
• In a controlled experiment, scientists assess causality
by observing the effect on 𝑌 of changing the values of
𝑋, while keeping constant all other variables
– E.g., Ohm’s law of current, Newton’s law of gravitation, etc.
• When some of the variables are not under direct
experimental control, scientists may execute a
randomized controlled trial (RCT)
In the 1930s, Ronald Fisher
– E.g., Effectiveness of Pfizer’s COVID-19 vaccine popularized randomized experiments
• Under random assignment, subjects in the treatment as a way to de-confound variables.
group (𝑋 = 𝑥1) are assumed to be indistinguishable
The first published RCT appeared in
from subjects in the control group (𝑋 = 𝑥0 ) 1948. Today, well-blinded RCTs are
– Thanks to this assumption, the difference in outcomes can be considered the gold standard in
attributed to the treatment experimental research.

Natural Experiments
• Sometimes interventional studies are not possible,
because they are unfeasible, unethical, or prohibitively
expensive
• In a natural experiment, subjects are assigned to the
treatment and control groups determined randomly by
Nature or by other factors not controlled by scientists
• Examples of natural experiments include
– Regression discontinuity design (RDD): When treatment and
control groups are comparable in everything but the slight
difference in the assignment variable, attributed to noise
In 1854, Dr. John Snow found that exposure
– Crossover studies (COS): When the effect of confounders does
to contaminated water causes cholera. Sick
not change per subject over time and healthy neighbors of London’s Soho
– Difference-in-differences studies (DID): When factors other than district were comparable in all respects,
the treatment influence the outcome over time except by their use of different water
pumps.
Simulated Intervention: Backdoor Adjustment
• Under some conditions, we can simulate an intervention
• A backdoor path between 𝑋 and 𝑌 is an unblocked non-
causal path that connects those two variables
• A set of variables 𝑆 satisfies the backdoor criterion if the
following two conditions are true:
– conditioning on 𝑆 blocks all backdoor paths between 𝑋 and 𝑌
– 𝑆 does not contain any descendants of 𝑋 Conditioning on confounder Z
• Then, 𝑆 is a sufficient adjustment set, and the causal (shaded node) blocks the path
𝑋 ← 𝑍 → 𝑌, leaving the causal
effect of 𝑋 on 𝑌 can be estimated as: path 𝑋 → 𝑌 as the only unblocked
path.
𝑃 𝑌 = 𝑦|𝑑𝑜 𝑋 = 𝑥 = ෍ 𝑃 𝑌 = 𝑦|𝑋 = 𝑥, 𝑆 = 𝑠 𝑃 𝑆 = 𝑠
𝑠 Under those circumstances,
association does imply causation,
• Examples of other adjustments: Front-door, IV, etc. and we can simulate the outcome
of a do-operation through
Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450 conditional probabilities. 17
Causality in Econometrics

Authors Mistake Causation for Association
• On the joint distribution of 𝑋, 𝑌 , a researcher
can fit the linear model 𝑌𝑡 = 𝛽0 + 𝛽1 𝑋𝑡 + 𝜀𝑡
• Alternatively, one could fit 𝑋𝑡 = 𝛾0 + 𝛾1 𝑌𝑡 + 𝜁𝑡
• In general, for least-squares (LS) estimates:
– 𝛾ො0 ≠ −𝛽መ0 /𝛽መ1 , 𝛾ො1 ≠ 1/𝛽መ1 , and 𝜁መ ≠ −𝜀/Ƹ 𝛽መ1
• The reason for this asymmetry is, each model
defines the error as the portion of the effect that
cannot be adjudicated to the chosen cause
By computing LS estimates on a particular
– 𝑌𝑡 = 𝛽0 + 𝛽1 𝑋𝑡 + 𝜀𝑡 implies a causal graph 𝑋 → 𝑌 specification, an econometrician injects
– 𝑋𝑡 = 𝛾0 + 𝛾1 𝑌𝑡 + 𝜁𝑡 implies a causal graph 𝑌 → 𝑋 extra-statistical information consistent with a
particular causal graph. Alternatively,
• Econometric models rely on LS estimators, hence econometricians could have used a Deming
implying causal relationships, not associational (or orthogonal) regression, a type of errors-
relationships in-variables model that attributes errors to
both 𝑋 and 𝑌 (a non-causal association).
Authors Misunderstand the Meaning of 𝜷
• መ
For the LS estimate to be unbiased (𝐸 𝛽|𝑋 = 𝛽), it
must occur that 𝐸 𝜀|𝑋 = 0 (exogeneity condition)
• Econometricians typically address this requirement by
defining 𝜀 ≡ 𝑌 − 𝐸 𝑌|𝑋 , hence 𝐸 𝑌|𝑋 = 𝑋𝛽
– 𝛽 receives a distributional interpretation, as the slope of a line
– This associational interpretation of 𝜀 is inconsistent with the
causal meaning of LS
• The correct (causal) interpretation of 𝜀 is “all causes of
𝑌 that are uncorrelated to 𝑋”
– This is a consequence of random assignment (e.g., in an RCT)
– In purely observational studies, exogeneity is contingent on
correct model specification Economist Trygve Haavelmo was
among the first to recognize that 𝛽
• Then, the correct meaning of 𝛽 is: 𝐸 𝑌|𝒅𝒐 𝑿 = 𝑋𝛽 has causal meaning in Economics.
Unfortunately, his point was ignored.
Authors Mistake Association for Causation
• For stationary 𝑋𝑡 and 𝑌𝑡 , Granger [1969] Number of citations > 31,037
proposed an econometric test for (linear) causality
𝐼 𝐽
𝑌𝑡 = 𝛽0 + ෍ 𝛽𝑖 𝑋𝑡−𝑝 + ෍ 𝛾𝑗 𝑌𝑡−𝑗 + 𝜀𝑡
𝑖=𝑝 𝑗=1
• According to Granger, 𝑋 causes 𝑌 iif at least one of Granger attempted to define causality in
the estimated 𝛽መ𝑖 is statistically significant terms of predictability (a characteristic of
– This approach was later expanded to multivariate systems, in the joint distribution of probability).
the form of a VAR specification
Granger [1969] remains one of the most
• Granger causality is a misnomer highly cited articles in the econometrics
– For example, if 𝑋 and 𝑌 are caused by Z (a confounder), literature, with thousands of new
Granger’s test will still falsely conclude that 𝑋 causes 𝑌 citations each year. This abuse of the
term causality has led to numerous false
• The test itself is susceptible to selection bias claims in the factor literature.
– The specification search requires multiple testing
Causality in Factor Investing

Causal Content
• The objective of a factor model is not to use 𝑋 to
predict 𝑌, but to determine rewarded risk exposures
• A researcher who just wanted to predict 𝑌 would have
– used more powerful techniques than a linear model The objective of a factor model such
– minimized the mean-squared error, instead of using a MVUE as 𝑌 = 𝑋𝛽 + 𝑍𝛾 + 𝜀 is not to predict
– used MDA or Shapley values, instead of p-values 𝑌 conditioned on 𝑋 and 𝑍
(𝐸 𝑌|𝑋, 𝑍 ), but to estimate the
• Factor investors build portfolios that causal effect of 𝑋 on 𝑌 (𝐸 𝑌|𝑑𝑜 𝑋 ),
– overweight stocks with a high exposure to 𝑋, and which requires adjusting for the
– underweight stocks with a low exposure to 𝑋, confounding effect of 𝑍.
– at the tune of one separate portfolio for each factor,
– with no regard for the value of 𝜀 The model specification 𝑌 = 𝑋𝛽 +
𝑍𝛾 + 𝜀 is consistent with a particular
• By making these modelling decisions, researchers causal graph, of which the above is
have injected causal assumptions into their analyses just one possibility among several.

Omitted Mediation Analysis
• Factors are often justified with economic rationales
– E.g., reward for accepting a natural undiversifiable risk
• An economic rationale does not raise to the level of
scientific theory
– “𝑋 causes 𝑌 through a falsifiable mechanism 𝑀”
• An example of a causal theory would be the SEM
𝑂𝐼𝑡 ≔ 𝑓1 𝑝𝑡 − 𝑣𝑡 + 𝜀1,𝑡
𝐻𝑀𝐿𝑡
𝑝𝑡+ℎ − 𝑣𝑡 ≔ 𝑓2 𝑂𝐼𝑡 + 𝑓3 𝑀𝑂𝑀𝑡 + 𝜀2,𝑡+ℎ

𝑃𝐶𝑡+ℎ
𝐻𝑀𝐿𝑡 ≔ 𝑓4 𝑀𝑂𝑀𝑡 + 𝜀3,𝑡

• Absent a hypothesized causal mechanism, it is not
possible to design an experiment to assess the Hypothetical causal mechanism for
validity of factor investing claims value, controlling for momentum.
24
Causal Denial
• Despite of the causal content of factor investing
strategies, authors almost never
– Declare a causal mechanism, or
– Justify their model specification with a causal graph
• Without a declared causal structure,
– the estimated 𝛽 loses its causal meaning (the effect on 𝑌 of an
intervention on 𝑋), and
– p-values merely convey the strength of associations of unknown
origin (causal and non-causal combined).
If factor researchers only cared about
• The consequence of factor investing’s causal denial prediction, then they would minimize
– without a causal mechanism, there is no investment theory; the overall mean squared error, not just
– without investment theory, there is no falsification; the variance among unbiased estimators
– without falsification, investing cannot be scientific (BLUE). The use of BLUE and p-values
implies a causal interpretation of 𝛽.

Type-A Spuriosity
• Type-A spuriosity occurs when a researcher
mistakes random variability (noise) for signal,
resulting in a false association
• Type-A spuriosity has several attributes:
a) it results in type-1 errors (false positives)
b) for the same number of trials, it has a lower
probability to take place as the sample size grows
c) it can be corrected through multiple-testing
adjustments
• Two main reasons for Type-A spuriosity
– p-hacking, e.g., Hochberg [1988]
– Backtest overfitting, Bailey and López de Prado
[2014]
Distribution of the maximum Sharpe ratio as a function of
the number of trials, where the true Sharpe ratio is zero.
See “The False Strategy Theorem.”
Type-B Spuriosity
• Type-B spuriosity occurs when a researcher mistakes
association for causation (e.g., due to misspecification)
• Type-B spuriosity has several attributes:
a) it results in type-1 errors and type-2 errors (false positives and
false negatives);
b) it can occur with a single trial;
c) it has a greater probability to take place as the sample size The top graph is an example of false
grows, because the non-causal association can be estimated association (type-A spuriosity). The
with lower error; and bottom graph is an example of
association mistaken for causation
d) it cannot be corrected through multiple-testing adjustments. Its
(type-B spuriosity).
correction requires the injection of extra-statistical information,
in the form of a causal theory Type-A and type-B spuriosity are
• Type-B spurious factors exhibit mutually exclusive. For type-B spuriosity
to take place, the association must be
– misattributed risk and returns non-causal but true, which precludes
– time-varying risk premia that association from being type-A
spurious.
Type-B(1) Spuriosity: Under-Controlling
• Consider a researcher who fits 𝑌 = 𝑋𝛽 + 𝜀 on data
generated by 𝑌: = 𝑋𝛽 + 𝑍𝛾 + 𝑢, where 𝛾 ≠ 0 and
𝑢 is white noise
– As a consequence, 𝐸 𝜀|𝑋 = 𝛾𝐸 𝑍|𝑋
– 𝐸 𝑍|𝑋 ≠ 0 ⇒ 𝐸 𝜀|𝑋 ≠ 0 (exogeneity is not satisfied)
• Case 1: 𝑍 is a mediator (𝑍 ≔ 𝑋𝛿 + 𝑣, with 𝛿 ≠ 0)
– the chosen specification biases 𝛽መ
– however 𝛽መ can still be interpreted as a total causal effect
Econometric textbooks treat all missing
• Case 2: 𝑍 is a confounder (𝑋 ≔ 𝑍𝛿 + 𝑣, with 𝛿 ≠ 0) variables as equal. This is a mistake.
– the chosen specification biases 𝛽መ
In the top graph, 𝑍 is a mediator, and
– 𝛽መ cannot be interpreted as a causal effect (direct or total)
missing 𝑍 has mild consequences. In the
• Wrong risk attribution means wrong allocations! bottom graph, 𝑍 is a confounder, and
missing 𝑍 will likely lead to false positives
or false negatives.
A Mystery Solved: Time-Varying Risk Premia
• The time-varying nature of risk premia has puzzled
researchers for 3+ decades. Explanations include:
– changes in expected market returns
– temporary changes in investor or market behavior
A leading explanation for the time-
• There is an easier explanation: A missing confounder 𝑍
varying nature of risk premia is that 𝛽
መ
𝐸 𝛽|𝑋 = 𝛽 + 𝛾𝛿 1 + 𝛿 2 −1 changes in response to changes in
investors’ expectations or behavior.
• Consider the case where the market rewards exposure
to 𝑋 and 𝑍 (𝛽 > 0, 𝛾 > 0) A more likely explanation is that 𝛽መ
changes (but not 𝛽) due to changes
– Even if the two risk premia remain constant, changes over time in 𝛾 or even worse, changes in 𝛿.
in 𝛿 will change 𝛽መ
– In particular, for a sufficiently negative value of 𝛿, then 𝛽መ < 0 Time-varying risk premia is not a
feature of the markets, it is a bug in
• Time-varying risk premia is due to Type-B(1) spuriosity factor investing.

Type-B(2) Spuriosity: Over-Controlling
• Statisticians have been trained for decades to
control for any variable 𝑍 associated with 𝑌 that is
not 𝑋
– Econometrics textbooks dismiss as a harmless error the
inclusion of an irrelevant variable, regardless of the variable’s
role in the causal graph
• Case 1: 𝑍 is a mediator
– Controlling for a mediator interferes with the mediated effect
and the total effect, which the researcher may wish to assess Greene [2012, section 4.3.3] states that
– 𝛽መ measures only the direct effect the only downside to adding superfluous
variables is a reduction in the precision
• Case 2: 𝑍 is a collider of the estimates. This is a mistake.
– Controlling for a collider opens a backdoor path, 𝑋 → 𝑍 ← 𝑌
(Berkson’s fallacy) Over-controlling for a collider has the
same consequences as under-controlling
for a confounder: an open backdoor.
Type-B(3) Spuriosity: Specification-Searching
• The use of explanatory power (an associational, non-
causal concept) for selecting the specification of a
factor model is inconsistent with that model’s causal
content
• Specification-searching commingles two separate and
sequential stages of the causal analysis:
1) Causal discovery: Finding the causal graph Econometric studies often justify
2) Control: Use the graph to determine the correct specification the chosen specification in terms of
explanatory power. This comingles
• Stage (2) should be informed by stage (1), not the causal discovery with controlling,
other way around and all but ensures that the
– A researcher may achieve higher explanatory power by regressors will include colliders
combining multiple causes of 𝑌, at the expense of biasing the (Berkson’s fallacy).
multiple parameters’ estimates due to multicollinearity or over-
controlling for a collider
Monte Carlo Experiments

Type-B(1) Spuriosity :
Forks

Confounders
• Consider the fork structure in the right graph
• Applying Bayesian network factorization
𝑃 𝑋, 𝑌, 𝑍 = 𝑃 𝑍 𝑃 𝑋|𝑍 𝑃 𝑌|𝑍
• 𝑋 and 𝑌 are associated, since
𝑃 𝑋, 𝑌 = ෍ 𝑃 𝑍 𝑃 𝑋|𝑍 𝑃 𝑌|𝑍 ≠ 𝑃 𝑋 𝑃 𝑌
𝑌𝑡 = 𝛼 + 𝛽𝑋𝑡 + 𝜀𝑡
𝑍
• This is an example of non-causal association
– 𝑋 and 𝑌 are associated through the backdoor path
𝑌←𝑍→𝑋
• Given the causal content of the factor model, a
statistically significant 𝛽መ implies that 𝑋 causes 𝑌
– This claim of statistical significance is type-B spurious

The Backdoor Adjustment
• The effect of conditioning by 𝑍 is equivalent to
simulating a do-operation (an intervention)
– 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 𝑌𝑡 = 𝛼 + 𝛽𝑋𝑡 + 𝛾𝑍𝑡 + 𝜀𝑡
𝑃𝑍
• It is possible to remove the confounder-induced
bias by adding 𝑍 as a regressor (the partial
correlations method)
• With the correct model specification, the
researcher concludes that 𝑋 does not cause 𝑌

Immoralities

Colliders
• This causal graph shows a collider:
– Variable 𝑍 is influenced by both, the treatment 𝑋 and the
outcome 𝑌
• If a researcher controls for 𝑍, the result is a false
positive (bottom table)
• Compare the fork structure with the immorality
𝑌𝑡 = 𝛼 + 𝛽𝑋𝑡 + 𝛾𝑍𝑡 + 𝜀𝑡
structure
– When the direction of causality is reversed, a confounder
becomes a collider
– The direction of causality is critical for specification
• One problem is, the direction of causality cannot
always be determined from data
– Causal graphs incorporates extra-statistical (beyond
observational) information
Berkson’s Fallacy
• Berkson’s fallacy occurs when a spurious
association is observed between two independent
variables, as a result of conditioning on a collider
• With a careful selection of colliders, a researcher
can present evidence in support of any spurious
investment factor 𝑌𝑡 = 𝛼 + 𝛽𝑋𝑡 + 𝜀𝑡
• The correct causal treatment of a collider is to
indicate its presence, and justify why researchers
should not control for it
• Over-controlling leads to
– false positives, in the presence of colliders
– false negatives, in the presence of mediators
• E.g., controlling for 𝑍 in 𝑋 → 𝑍 → 𝑌

Chains

Confounded Mediators
• This causal graph shows a mediator and a
confounder:
– Variable 𝑍 mediates the causal flow from the treatment
𝑋 to the outcome 𝑌
– Variable 𝑊 confounds 𝑍 and 𝑌
• If a researcher controls for 𝑍, the outcome is a
false positive (bottom table) 𝑌𝑡 = 𝛼 + 𝛽𝑋𝑡 + 𝛾𝑍𝑡 + 𝜀𝑡
– The reason is that 𝑍 also operates as a collider to 𝑋 and
𝑊
• Controlling for 𝑍 opens a backdoor path 𝑋 → 𝑍 ← 𝑊 → 𝑌
– While it is true that 𝑋 causes 𝑌 (through 𝑍), the
collider’s bias is so strong that the sign of the
relationship is reversed (𝛽መ ≪ 0)
• In the absence of link 3, controlling for 𝑍 would
have led to a false negative
Mediation Fallacy & Simpson’s Paradox
• The Mediation Fallacy involves conditioning on
the mediator when the mediator and the
outcome are confounded
• Simpson’s paradox occurs when there is an
association in several groups, but it disappears
or reverses when the groups are combined 𝑌𝑡 = 𝛼 + 𝛽𝑋𝑡 + 𝜀𝑡
• The solution to Simpson’s paradox is to inject
extra-statistical information in the form of a
causal graph
• We can estimate the unbiased effect (𝛽መ ≫ 0)
– Specification-searching would have returned a
misspecified model (R2 drops from 0.78 to 0.14!)
– Adding 𝑊 increases R2 to 0.71 (still below 0.78)

Conclusions

Spurious Investment Factors (1/2)
• Consider the influential three-factor (FF93) and
five-factor (FF15) models proposed by Fama and
French
• Type-A Spuriosity: p-hacking
– These authors do not adjust p-values for multiple trials
– In the year 2023, it is unfortunate that financial
econometricians still deny the need to adjust for
selection bias under multiple testing
• Type-B(1) Spuriosity: Under-controlling
– Missing confounders (e.g., momentum, macro variables)
may explain the puzzle of time-varying risk premia
– Carhart (C97) added momentum to FF93, however his Performance of active ETFs launched in the
rationale was greater explanatory power (specification- U.S. between 1993 and 2014. Cumulative
searching), not de-confounding returns flatten 6 months before launch,
coinciding with the end of the in-sample
period. This points to backtest overfitting.
43
Spurious Investment Factors (2/2)
• Type-B(2) Spuriosity: Over-controlling
– Controlled colliders (e.g., HML) may explain false
associations, or true associations with wrong signs
• Type-B(3) Spuriosity: Specification-searching
– A correctly specified model can deliver lower
explanatory power than a misspecified model
– Improving on FF93’s explanatory power does not make
C97’s model better specified, or its estimates less biased
• To summarize, the findings in FF93, FF15 and
C97 are likely type-A or type-B spurious
Under this graph, the estimates in FF93,
• To address these issues, factor researchers must FF15 and C97 are biased. This particular
– adjust for selection bias under multiple testing graph may be incorrect, however the
– justify their specification choices through a causal graph burden of the proving it wrong belongs to
– propose a falsifiable causal mechanism the authors claiming the existence of
investment factors.
The Dawn of Causal Factor Investing
• A scientific theory is a falsifiable statement of Type Rigor Example
the form “𝑋 causes 𝑌 through mechanism 𝑀” Randomized Very high Algo-wheel
controlled trials experiments
• Scientific theories matter to investors because
Natural High Market-maker reaction
– causality is a necessary condition for investment experiments to random spikes in
efficiency order imbalance
• associational models misattribute risks and performance, thus
preventing investors from building efficient portfolios
Simulated Medium Estimate effect of HML
interventions using a causal graph
– causal models enable counterfactual reasoning,
hence the stress-testing of investment portfolios in a Econometric Low Factor investing
(observational) literature; backtested
coherent and forward-looking manner studies investment strategies
• associational models cannot answer counterfactual questions, such
as what would be the effect of 𝑌 on a not-yet-observed scenario 𝑋, Case studies Very low Broker report /
thus exposing those relying on associations to black-swan events analysis
• Financial economists’ adoption of causal Expert opinion Anecdotal Investment guru’s
prediction
inference has the potential to transform
investing into a truly scientific discipline Hierarchy of evidence: Pre-scientific vs.
scientific evidence in financial research. 45
Call For Papers
ADIA Lab encourages researchers to move
factor investing beyond its current pre-
scientific stage.
To help dawn the discipline of Causal

Factor Investing, ADIA Lab has called for
papers that promote the use of the formal
language of causal inference in investing.
To learn more, visit:

https://www.adialab.ae/call-for-papers
We look forward to your papers!

For More Information
Download for free
Causal Factor Investing
(Cambridge University Press, 2023)
Available at:
https://www.cambridge.org/core/elements/
causal-factor-investing/
9AFE270D7099B787B8FD4F4CBADE0C6E

Disclaimer
• The views expressed in this presentation are my own, and do not necessarily
reflect the views of Cornell University, the Abu Dhabi Investment Authority, or
ADIA Lab
• No investment decision or particular course of action is recommended by this
presentation
• All Rights Reserved. © 2020 - 2023 by Marcos López de Prado
www.adialab.ae


SSRN Id4324450

Uploaded by

Copyright:

Available Formats

SSRN Id4324450

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SSRN Id4324450

Uploaded by

Copyright:

Available Formats

Can Factor Investing Become Scientific?

Marcos López de Prado

Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450

Bloomberg – Goldman Sachs Asset

(BBG code: BGSUSEMF <Index>)

value=0.12), under 0% risk-free rate

• This performance does not include:

Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450

The Second Book of the Georgics

Electronic copy available at: https://ssrn.com/abstract=4324450 5

Electronic copy available at: https://ssrn.com/abstract=4324450 6

An Enquiry concerning Human

Electronic copy available at: https://ssrn.com/abstract=4324450 7

Electronic copy available at: https://ssrn.com/abstract=4324450 8

Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450

Electronic copy available at: https://ssrn.com/abstract=4324450 10

Electronic copy available at: https://ssrn.com/abstract=4324450 11

∃𝒙, 𝒚|𝑷 𝒀 = 𝒚|𝑿 = 𝒙 ≠ 𝑷 𝒀 = 𝒚 Weather (𝑊) influences ice cream sales

Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450

Electronic copy available at: https://ssrn.com/abstract=4324450 15

Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450

Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450

Electronic copy available at: https://ssrn.com/abstract=4324450 23

𝑝𝑡+ℎ − 𝑣𝑡 ≔ 𝑓2 𝑂𝐼𝑡 + 𝑓3 𝑀𝑂𝑀𝑡 + 𝜀2,𝑡+ℎ

𝐻𝑀𝐿𝑡 ≔ 𝑓4 𝑀𝑂𝑀𝑡 + 𝜀3,𝑡

Electronic copy available at: https://ssrn.com/abstract=4324450 25

Electronic copy available at: https://ssrn.com/abstract=4324450 29

Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450

Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450

Electronic copy available at: https://ssrn.com/abstract=4324450 34

Electronic copy available at: https://ssrn.com/abstract=4324450 35

Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450

Electronic copy available at: https://ssrn.com/abstract=4324450 38

Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450

Electronic copy available at: https://ssrn.com/abstract=4324450 41

Classification: Public Electronic copy available at: https://ssrn.com/abstract=4324450

To help dawn the discipline of Causal

To learn more, visit:

We look forward to your papers!

Download for free

Causal Factor Investing

(Cambridge University Press, 2023)

Electronic copy available at: https://ssrn.com/abstract=4324450 47

Electronic copy available at: https://ssrn.com/abstract=4324450 48

You might also like