Causal Forest

Causal Forests employ a technique called honesty, where the training data is split into two distinct parts. One subsample is used to construct the tree structure (selecting splits), and a separate, independent subsample is used to estimate the treatment effects within the leaves. This separation prevents overfitting and ensures that the estimated effects are asymptotically unbiased and have valid confidence intervals.

Unlike standard methods that estimate a single Average Treatment Effect (ATE), Causal Forests are designed to uncover Heterogeneous Treatment Effects (HTEs). The algorithm recursively partitions the feature space to identify subgroups with distinct causal responses. For example, it can reveal that a promotional discount increases sales by 20% for new customers but has a negligible 2% effect on loyal, high-frequency purchasers.

Instead of minimizing prediction error, the tree-splitting criterion maximizes the heterogeneity of treatment effects across child nodes. This is often achieved by using a gradient-based approach derived from the R-Learner framework, which directly targets the treatment effect function by residualizing both the outcome and the treatment assignment, isolating the causal signal from confounding associations.

A defining feature of the Generalized Random Forest framework is that the resulting estimates are asymptotically Gaussian and unbiased. This statistical property allows for the construction of valid confidence intervals around individual treatment effect predictions. A supply chain risk manager can therefore state with 95% confidence that switching a specific supplier reduces lead time variance by 1.5 to 3.2 days.

Causal Forests are integrated with orthogonalization techniques, such as Double Machine Learning, to control for high-dimensional confounders. By first residualizing the outcome and treatment using any flexible machine learning model, the forest operates on data where the confounding signal has been removed. This makes the treatment effect estimates robust even when hundreds of covariates influence both the intervention and the outcome.

The Causal Forest is a specific application of the broader Generalized Random Forest (GRF) framework. This framework extends tree-based methods to estimate any quantity of interest defined by a local moment condition, including conditional average treatment effects, quantile treatment effects, and instrumental variables regression. The core innovation is a gradient-based algorithm that adapts the splitting rule to the target estimand.

The foundational concept that Causal Forests are designed to estimate. Unlike the Average Treatment Effect (ATE) , which provides a single population-wide metric, HTE captures how a disruption's impact varies across different segments of the supply chain.

• Example: A port closure might have a severe negative effect on just-in-time automotive manufacturers but a negligible or even positive effect on local warehousing competitors.
• Mechanism: Causal Forests recursively partition data to identify subgroups with distinct treatment responses.

A direct application of Causal Forest logic, uplift modeling predicts the incremental impact of a specific intervention on an individual entity. It segments a population into four groups:

• Persuadables: Respond positively only because of the treatment.
• Sure Things: Would respond positively regardless.
• Lost Causes: Will not respond positively in any scenario.
• Sleeping Dogs: Would respond negatively if treated. This targeting prevents wasteful mitigation efforts on nodes that would recover autonomously.

A powerful orthogonalization framework often used alongside or as an alternative to Causal Forests. Double Machine Learning (DML) uses arbitrary ML models to partial out the effects of high-dimensional confounders.

• Process: First, predict the treatment from the confounders and the outcome from the confounders. Then, regress the residuals of the outcome on the residuals of the treatment.
• Advantage: Provides valid confidence intervals for the Average Treatment Effect even when using complex models like gradient boosting for nuisance parameter estimation.

The formal parent framework that defines the data-generating process. A Structural Causal Model (SCM) specifies the functional relationships between variables, allowing for counterfactual reasoning.

• Relation to Causal Forest: While a Causal Forest is a non-parametric estimator, an SCM provides the explicit structural equations. You can use a Causal Forest to validate or discover the heterogeneity implied by an SCM.
• Supply Chain Context: An SCM might define Inventory_Level = f(Demand_Signal, Supplier_Lead_Time, Unexplained_Disruption), where the Causal Forest estimates the shape of f.

A set of three inference rules developed by Judea Pearl that mathematically transform expressions involving the do() operator (interventions) into standard conditional probabilities.

• Purpose: Determines if a causal effect can be estimated from observational data alone, given a causal graph.
• Relevance: Before applying a Causal Forest, Do-Calculus can verify if the necessary backdoor criterion is satisfied by the available control variables, ensuring the estimated treatment effects are identifiable.

When a pre-specified causal graph is unavailable, these algorithms infer the structure directly from data. They test for conditional independencies to construct a Directed Acyclic Graph (DAG) .

• Constraint-based: Uses statistical tests (e.g., PC algorithm).
• Score-based: Searches for the graph that optimizes a fit score (e.g., GES).
• Integration: The output graph from a causal discovery algorithm provides the necessary covariate adjustment set required to run a valid Causal Forest analysis.