From correlation to causation.

Correlation is not causation. This is the first thing every statistics student learns — and the last thing most practitioners remember when the p-values look convincing.

In finance and machine learning, the gap between "these two things move together" and "this thing causes that thing" is not academic pedantry. It is the difference between a strategy that earns alpha and one that blows up the next time the regime changes. Causal inference provides the mathematical machinery to close that gap.

Why causal inference matters in finance

Every quant strategy is, at its core, a causal claim. When a momentum signal "predicts" future returns, we are asserting that the past price movement causes the future one — not merely that they are correlated. When a factor model attributes returns to value, size, or quality, we are asserting that these characteristicscause higher expected returns. When we use a machine learning model to select trades, we are asserting that the features it uses have causal relevance to the outcome.

The problem is that most of these claims are not actually causal. They are associational. They capture correlations that happen to hold in-sample but break the moment the data-generating process shifts — which is exactly what happens in financial markets, where regimes change constantly.

The fundamental problem of causal inference

The fundamental problem of causal inference, articulated by Donald Rubin and others, is elegantly simple: we can never observe the same unit under both treatment and control simultaneously. For a single stock, we cannot know what its return would have been if a particular event had not occurred. We only observe one world.

This is why causal inference requires assumptions — structural assumptions about how the world works — that go beyond what the data alone can tell us. The art of causal inference lies in choosing assumptions that are plausible, testable, and sufficient to identify the causal effect of interest.

Potential outcomes framework

The potential outcomes framework (also called the Rubin causal model) formalizes this idea. For each unit i and treatment condition t, there exists a potential outcome$Y_i(t)$ — the outcome that would be observed if unit i received treatment t. The individual causal effect is $Y_i(1)-Y_i(0)$, but we only ever observe one of these two quantities for each unit.

The average treatment effect (ATE) — $E[Y(1)-Y(0)]$ — is the population-level quantity of interest. Identification strategies (randomization, matching, instrumental variables, difference-in-differences, etc.) are methods for estimating this quantity from observed data under specific assumptions.

A timeline of causal inference

An unobservable counterfactual is not a problem one paper solves. It took almost three centuries — mathematics, statistics, econometrics, and most recently machine learning, each era chipping away at a different piece — to assemble the toolkit that makes the rest of this essay possible.

Directed Acyclic Graphs (DAGs)

Sewall Wright's 1921 path diagrams were the first attempt to draw the assumptions instead of just stating them. A century of graph-theoretic development later, the DAG is that idea made rigorous: a directed graph in which nodes are variables and edges are direct causal effects. The graph encodes — in a compact, visual form — the assumptions that are necessary and sufficient for identification.

Why DAGs matter for quant research

In quant research, we routinely build models with dozens of features. Without a DAG, it is impossible to know which features are confounders (variables that cause both the treatment and the outcome, creating a spurious association), which are mediators (variables on the causal pathway), and which are instruments (variables that affect the outcome only through the treatment).

Consider a simple example: you observe that stocks with higher analyst coverage tend to have higher returns. Is analyst coverage causing higher returns? Or is there a confounder — perhaps firm size — that causes both higher analyst coverage and higher returns? A DAG makes this question precise and tells you exactly what data you need to answer it.

d-separation and identification

The graph theory of DAGs gives us d-separation: a criterion for determining whether two variables are conditionally independent given a set of covariates. If $X$ and $Y$ are d-separated given $Z$ in the DAG, then conditioning on $Z$ is sufficient to remove all backdoor paths between $X$ and $Y$, allowing unbiased estimation of the causal effect of $X$ on $Y$.

This is the backbone of the backdoor criterion — the most widely used identification strategy in applied causal inference. If you can find a set of covariates that blocks all backdoor paths, you can estimate the causal effect by conditioning on those covariates.

The do-calculus

The do-calculus, developed by Judea Pearl, is a formal system for reasoning about interventions. The key distinction is between conditioning ($P(Y|X=x)$) and intervening ($P(Y|do(X=x))$). Conditioning asks: among units that already have $X=x$, what is the distribution of $Y$? Intervening asks: if weforce $X=x$, what would the distribution of $Y$ be?

In finance, this distinction is crucial. A strategy that selects stocks with high momentum is conditioning on a variable that is itself a consequence of past returns. The do-operator forces the momentum signal to take a specific value regardless of its cause — which is what we mean when we say "what would happen if we implemented this strategy."

Granger causality

In time series and financial econometrics, Granger causality is the workhorse identification strategy. A variable $X$ is said to Granger-cause $Y$ if past values of $X$ contain information that helps predict $Y$ beyond the information contained in past values of $Y$ alone.

Granger causality is not "true" causality in the Pearl sense — it is a predictive, not an ontological, concept. But it is practically useful: if past values of a signal help predict future returns after controlling for the return history itself, that signal has information content that is not redundant with the return history.

The limitation is that Granger causality is sensitive to the choice of lag length and can be confounded by omitted variables. A variable that Granger-causes returns might do so because it proxies for an omitted confounder, not because it has a direct causal effect. This is where DAGs come in: they tell you what you need to control for to get from Granger-causality to genuine causal identification.

Causal discovery

While most of the causal inference literature assumes the DAG is known (from domain expertise), causal discovery algorithms attempt to learn the DAG from observational data. Algorithms like PC, FCI, and GES use conditional independence tests to recover the causal structure.

In finance, causal discovery is particularly appealing because it promises to let the data — rather than the researcher's prior beliefs — reveal which factors actually drive returns. The caveat is that causal discovery from observational data is fundamentally limited without strong assumptions. Different algorithms can recover different DAGs from the same data, and the equivalence class of DAGs consistent with the data is often too large to be useful without domain knowledge.

Applications in quantitative finance

Factor investing and causal factor models

Traditional factor models (Fama-French, Carhart) are associational: they describe which factors are correlated with returns, not which factors cause them. The book Causal Factor Investing by Marcos López de Prado argues that this distinction matters profoundly. A factor that is merely correlated with returns may stop working when the correlation breaks; a factor that causally drives returns is more likely to persist across regimes.

The approach uses DAGs to identify the causal structure of factor returns, then applies the do-calculus to estimate the causal effect of each factor on portfolio returns. This leads to different factor selection criteria than traditional correlational approaches — and, in practice, to more robust portfolios.

Meta-labeling and causal feature selection

In ML for trading, meta-labeling uses a second model to predict whether the primary model's prediction is correct. Causal inference can improve this process by ensuring that the features used in the meta-labeler are causally relevant to the prediction accuracy, not merely correlated with it. This reduces overfitting and improves out-of-sample performance.

Treatment effect estimation for strategy evaluation

When evaluating a trading strategy, the question is not "what were the returns?" but "what would the returns have been if we had not employed this strategy?" This is a causal question. Methods like synthetic control and difference-in-differences can provide causal estimates of strategy performance by constructing a counterfactual — what would have happened in the absence of the strategy.

Common pitfalls

• Confounding by omitted variables. The most common error in quant research is estimating an associational effect and calling it causal. If you do not control for a confounder, your estimate is biased — and the bias can be in any direction.
• Collider bias. Conditioning on a variable that is a common effect of two other variables (a collider) induces a spurious association between them. This is a subtle but pervasive source of bias in factor research.
• M-bias. A variant of collider bias where the collider is unobserved but its proxies are included in the model, inducing bias through a different path.
• Time-travel confounding. Using future information to predict the past. In finance, this happens when a feature is constructed using data that would not have been available at the time of the prediction.
• Survivorship bias. Estimating causal effects on a sample that excludes units that dropped out of the study (e.g., delisted stocks), leading to overly optimistic estimates.

The path forward

Causal inference is not a replacement for machine learning or statistical modeling — it is a layer on top of them. The workflow looks like this:

1. Draw a DAG that encodes your causal assumptions about the problem.
2. Use the DAG to identify which variables to control for (the backdoor criterion) and which to exclude (to avoid collider bias).
3. Estimate the causal effect using an appropriate method (regression, matching, instrumental variables, difference-in-differences, etc.).
4. Validate the causal estimate using out-of-sample data, cross-validation, and — ideally — interventional experiments (paper trading, A/B tests).

This workflow does not guarantee correct causal estimates — no methodology does. But it dramatically reduces the risk of confusing correlation with causation, which is the single most expensive error in quantitative research.

This is where the essay ends, but not where the toolkit does. The same identification problem, asked of a sequence instead of a snapshot, is what Granger causality and PCMCI exist to answer. Pointed at a trading signal instead of a treatment, it is what separates a feature that causes returns from one that merely correlated with them in-sample. That extension — causal inference stretched across time and stress-tested against live capital — is the throughline of the research on this site.