Causal Forest

A core innovation of the Causal Forest is honest estimation. Each tree is built using a random subsample of the data, which is split into two parts:

• Splitting subsample: Used exclusively to construct the tree structure and determine the recursive partitions.
• Estimation subsample: Used exclusively to calculate the treatment effect within each leaf. This separation prevents overfitting and ensures that the estimated treatment effects are asymptotically normal and unbiased, enabling valid confidence intervals.

Unlike standard regression which estimates a single average treatment effect, Causal Forests automatically discover effect modifiers. The algorithm recursively partitions the covariate space to maximize the difference in treatment effects between child nodes:

• Identifies subgroups where a financial policy or medical intervention is most effective.
• Captures non-linear interactions without pre-specification.
• Reveals segments like 'high-volatility assets' or 'illiquid small-cap stocks' where a trading signal has a distinct causal impact.

The Causal Forest is a specific instance of the Generalized Random Forest (GRF) framework. It extends the idea of a random forest from predicting outcomes to estimating any quantity defined by a local moment condition. Key properties include:

• Adaptive kernel weighting: Each test point receives a weighted set of training neighbors based on how often they share a leaf.
• Gradient-based splitting: Trees are grown using pseudo-outcomes derived from the gradient of the estimating equation, targeting the causal parameter directly. This framework also supports instrumental variables regression and quantile treatment effects.

To estimate causal effects from observational data, Causal Forests employ double machine learning principles. The algorithm orthogonalizes the estimation by first fitting two separate models:

• Propensity model: Estimates the probability of receiving treatment given covariates.
• Outcome model: Estimates the baseline outcome given covariates. The residuals from these models are then used to grow the forest, making the final treatment effect estimates robust to regularization bias and confounding errors in the nuisance components.

A critical advantage for financial decision-making is the ability to provide valid pointwise confidence intervals. Leveraging the asymptotic normality derived from honest estimation and the infinitesimal jackknife, a Causal Forest outputs:

• The predicted treatment effect for any given observation.
• The standard error of that prediction. This allows a quantitative strategist to test if a predicted alpha signal is statistically different from zero for a specific asset before committing capital, moving beyond point predictions to risk-aware inference.

Causal Forests are designed to handle a large number of potential confounders and effect modifiers without manual feature selection. The algorithm performs an implicit regularized search:

• Splits are chosen from a random subset of variables at each node, ensuring diversity.
• Irrelevant variables are naturally ignored as they do not reduce the loss function.
• This makes it suitable for alternative data environments where hundreds of weak signals (e.g., sentiment scores, macro indicators) might interact to cause a price movement.