A Protocol for Causal Factor Investing

By Marcos López de Prado & Vincent Zoonekynd, ADIA Lab (May 2025)

This paper proposes a major shift in how factor investing is approached, arguing that most traditional methods are flawed because they rely on associational rather than causal inference. The authors introduce the concept of the factor mirage - strategies that appear statistically robust but are built on causal misunderstandings, leading to unreliable out-of-sample performance.

Why This Matters

Factor investing is a cornerstone of quantitative finance, with trillions managed via style-based models. Yet many strategies underperform in practice despite strong backtests. While p-hacking and overfitting are often blamed, this paper highlights a more fundamental issue: the use of econometric tools that detect correlations rather than causal relationships.

Key Concepts

• Factor Mirage: A factor model that passes conventional statistical tests but fails because it misrepresents causal relationships - particularly due to confounder bias and collider bias.

  • Confounder bias: Failing to adjust for a variable that influences both the factor and returns (e.g., leverage).

  • Collider bias: Adjusting for a variable that is a consequence of both the factor and returns, which introduces spurious correlations.

• Causal Misspecification: Traditional factor models often use variables that improve fit metrics (e.g., R²), but inadvertently introduce biases by including colliders or omitting confounders.

Practical Solution: A 7-Step Protocol for Causal Factor Investing

1. Variable Selection
   Use machine learning (e.g., SHAP values, mutual information) to detect relevant variables without pre-specifying a functional form.

2. Causal Discovery
   Apply algorithms like PC or LiNGAM, plus domain knowledge, to construct causal graphs showing relationships among variables.

3. Causal Adjustment
   Use do-calculus to identify the correct set of control variables (confounders only; avoid colliders) based on the causal graph.

4. Causal Estimation
   Estimate causal effects using robust methods (e.g., double machine learning), and assess model performance in terms of classification (direction, ranking) and regression (magnitude) tasks.

5. Portfolio Construction
   Translate causal insights into portfolio weights, ensuring alignment with economic rationale and minimizing exposure to non-causal or spurious factors. Account for real-world constraints and stress-test for causal robustness.

6. Backtesting
   Go beyond single historical backtests: use resampling and Monte Carlo simulations that respect the causal structure to evaluate performance variability.

7. Multiple Testing Adjustments
   Control false discovery rates using p-value corrections or deflated Sharpe ratios, particularly in research involving many hypotheses or features.

Broader Implications

• The paper stresses that causal reasoning is not optional in investment strategy design - it’s necessary to avoid systematic inefficiencies, excessive turnover, and misleading attributions of return.

• Asset owners and risk managers are encouraged to adopt the protocol as part of due diligence and oversight, not just portfolio construction.

Final Takeaway

Factor investing is at a crossroads. If it is to remain credible and effective, it must move from statistical association to causal understanding. This paper offers a transparent, falsifiable roadmap to do just that - grounding investment strategies in the same rigor expected in fields like medicine or economics.