Causal Factor Investing: Can Factor Investing Become Scientific? Elements in Quantitative Finance. 2023. Marcos M Lopez de Prado. Cambridge University Press.
Causal inference is one of the most important topics in finance today. It is not only a critical issue for quantitative analysts and econometricians, it is also relevant to any analyst who wants to unravel the complex causal links between market drivers and asset returns. Yet, the distinction between associational and causal relationships is rarely taught and seldom discussed in the broader asset management industry.
The phrases, “past performance is not indicative of future success” and “correlation does not mean causality,” have become clichés that have lost their meaning as warnings for models. Financial factor modeling is often too casual, with little thought of causal inference. Researchers will look for elaborate past associations and develop stories to create the illusion of causality, yet the testing process and modeling procedures are often spurious. The problem has been characterized as data mining and p-hacking, yet these terms fail to address causal inference’s fundamental issue.
Lopez de Prado’s short book, Causal Factor Investing: Can Factor Investing Become Scientific? is both a tutorial on conducting causal inference and a call to arms for the asset management industry. The status quo of modeling associations and treating them as causal is fraught with investment risk. Using factor models is big business, and there may be little desire among those marketing them to pursue causal inference to find the truth. One need only look at the “factor zoo” driving finance research to see an explosion of statistical models with theoretical explanations as an afterthought.
This book is divided into eight short chapters that revolve around three major topics: a foundational description of scientific discovery and causal inference; formal testing for causality in econometrics and factor investing; and Monte Carlo experiments in finance. Readers with a basic understanding of statistics and an econometric course should be able to appreciate the core arguments in this book. Those with a stronger quant background will be exposed to many concepts that are not usually given much attention in financial econometrics and programming classes.
Lopez de Prado focuses on the core concepts for distinguishing between association and causality. Causality can be visually described by the workhorse for displaying causal links, directed acyclic graphs (DAG). The critical rationale for using DAG is that causality is directional but extra-statistical (beyond observation) and hence distinct from association. Causality may imply association, but association does not mean causality. The problem is that many researchers mistake causality for association. Sequentiality is necessary but not sufficient for causality. Some effect, X, may occur before Y adapts but not cause Y. Granger causality, a finance workhorse, or ML models that train or fit behavior are areas of concern, given the lack of a causal map.
The search for causality is a process of scientific discovery that moves from observation (logical induction) to a theory or hypothesis (logical abduction) and then attempts to test it through falsification. Skepticism, linked to Popper’s view of falsification, should be part of any investment committee’s discussion of a new model. With this logic, hypotheses can never be truly accepted but only rejected. The bar for claims of causality is much higher than simple association. There should be a premium for simple models that are easier to refute.
This causality problem explains the poor out-of-sample performance. Too often, there is type-A spuriousness, when a researcher mistakes random noise for signal, resulting in false associations or type-I errors (false positives). The problem arises from p-hacking and back-test overfitting. However, even if an association is true and not type-A spurious, it can still be type-B spurious when there is a mistake of causality from an association arising from misspecification that creates false positives and false negatives. Type-B spurious results come from both over- and under-controlling for variables beyond those we consider causal. There is a hierarchy of evidence, and unfortunately, the lack of controlled trials, natural experiments, and simulated interventions means the rigor of testing is low.
The book shows how causal inference is often characterized through three types of testing: interventional studies, natural experiments, and simulated interventions. Unfortunately, finance does not lend itself to randomized controlled trials. Finance researchers must focus on natural experiments or look for ways to tease out causality; however, there is growing work on forming simulated interventions. This is a process of discovery where the researcher identifies do-operators (the causal drivers) based on conditional probabilities, often using instrumental variables to control for latent effects that intervene in the causality search.
Lopez de Prado reviews the classic Fama-French 3- and 5-variable versions, along with the Carhart 4-factor model. He notes that their rationale and testing do not support a causal map or testing. The findings may be type-A spurious due to p-hacking and type-B spurious due to both under- and over-controlling for key variables. Without a causal map, factor investing is a black box. When supported through a black box, factors that were effective in the past may fail, leaving investors with no explanation other than that a factor is not working. The disclaimer that past performance is not indicative of future results is proven, yet investors are left with a hollow feeling that there is no alternative.
The author concludes with a call for the extra work necessary for causal investing that is grounded in standards of efficiency, interpretability, transparency, reproducibility, adaptability, extrapolation, surveillance, and improvability. Researchers should take the extra time to think through causal discovery and the do-calculus associated with careful analysis of key factor drivers. A quote by Lopez de Prado sums up the issue: “Put simply, without a causal mechanism, there is no investment theory; without investment theory, there is no falsification; without falsification, investing cannot be scientific.”
While written for those who do the heavy lifting for model building, Causal Factor Investing should be read by a broader audience. If users of factor models are not causal inference-sensitive, quant builders will not take the extra care before moving to production. Associational models without causal drivers are risk time bombs. Setting a higher standard for model construction is not about adding complexity or slowing research but about doing the homework to build models that will perform better out-of-sample across all market conditions.
If you liked this post, don’t forget to subscribe to the Enterprising Investor.
All posts are the opinion of the author. As such, they should not be construed as investment advice, nor do the opinions expressed necessarily reflect the views of CFA Institute or the author’s employer.