Double Machine Learning (DML)

DML constructs Neyman-orthogonal scores that are locally insensitive to nuisance parameter estimation errors. This means small mistakes in modeling the outcome or treatment functions do not contaminate the causal effect estimate.

• Removes first-order bias from the moment conditions
• Decouples the target parameter from nuisance function estimation
• Enables the use of flexible, high-dimensional ML models (random forests, neural nets) for nuisance functions without sacrificing root-n consistency of the treatment effect

Cross-fitting eliminates overfitting bias introduced when the same data is used to both estimate nuisance functions and the parameter of interest. The sample is partitioned into K folds.

• Nuisance models are trained on K-1 folds and used to generate predictions on the held-out fold
• The final estimator averages scores across all folds
• This sample-splitting ensures the estimator remains asymptotically normal and centered at the true parameter value

The foundational DML setup assumes a partially linear structural equation: Y = θ·D + g(X) + ε, where θ is the causal effect of treatment D on outcome Y, and g(X) is an unknown, potentially complex function of confounders X.

• Treatment D is also modeled as D = m(X) + ν
• DML first residualizes both Y and D by subtracting predicted g(X) and m(X)
• The causal parameter θ is then estimated by regressing the outcome residuals on the treatment residuals

DML provides doubly robust properties in many specifications. The estimator for the causal parameter remains consistent if at least one of the two nuisance models—the outcome model or the treatment propensity model—is correctly specified.

• Protects against systematic errors in a single ML model
• Allows practitioners to use aggressive, black-box algorithms for nuisance estimation
• Reduces reliance on fragile parametric assumptions common in classical econometrics

Traditional causal methods break down when the number of confounders approaches or exceeds the sample size. DML thrives in these high-dimensional regimes by leveraging ML regularization.

• Handles hundreds or thousands of potential confounders simultaneously
• Uses Lasso, gradient boosting, or deep networks to automatically select and transform relevant controls
• Enables causal inference from rich datasets like tick-level market data with many microstructure covariates

DML naturally extends to estimating conditional average treatment effects (CATE) —how the causal impact varies across subpopulations defined by covariates.

• The linear DML score can be interacted with context variables to discover effect modifiers
• Identifies which market regimes or asset characteristics amplify or dampen a trading signal's causal impact
• Integrates with causal forests for fully nonparametric heterogeneity analysis

A principle that DML operationalizes: an estimator is doubly robust if it remains consistent when either the propensity score model OR the outcome regression model is correctly specified—not necessarily both.

• Provides two chances to get the model right
• DML's orthogonal score is constructed to satisfy this property
• Contrast with inverse probability weighting (IPW) which fails if propensity scores are misspecified
• Contrast with outcome regression which fails if the response surface is misspecified
• Essential for financial applications where perfect model specification is impossible