A causal forest is a tree ensemble designed to estimate how a conditional average treatment effect varies with observed features. For a quant evaluating an execution intervention, it can distinguish conditions in which the intervention improves outcomes from conditions in which it makes them worse—even when the overall average is positive. The causal interpretation comes from the study design and identification assumptions.
The average can conceal the question
Imagine a fictional randomised experiment comparing a new execution rule, , with the existing rule, . The outcome is an execution score in basis points, measured over a fixed evaluation window; higher is better. Two groups, defined using pre-treatment features, each represent half the target population. Their stipulated population means are:
| Feature group | Mean under existing rule | Mean under new rule | Conditional effect |
|---|---|---|---|
| A | 5 bp | 7 bp | +2 bp |
| B | 5 bp | 4 bp | −1 bp |
The average treatment effect is bp. That summary conceals a negative effect in group B. A causal forest targets the variation behind the summary, learning feature-based neighbourhoods rather than requiring these two groups to be specified in advance.
Original stipulated population example, not forest estimates. Opposite conditional effects average to a small positive effect; no sample size or uncertainty is supplied.
These numbers illustrate the target, not the output of a fitted forest. They are conditional means, not jointly observed potential outcomes for individual orders. No standard error follows from this table.
What makes the comparison causal?
With potential outcomes and , the target is
This is an average for units sharing features , not recovery of a particular unit's unobserved individual treatment effect. Only one potential outcome is observed per unit. See Athey and Wager, §1.1.
For the usual binary-treatment setting, identification requires consistency with a well-defined intervention, no interference between units, and an assignment mechanism that makes treated and control units comparable conditional on pre-treatment . Randomisation can supply that comparability. In observational data it is the unconfoundedness assumption, . Overlap also matters: both assignments must be possible at the feature values being studied. Under these conditions,
An unrecorded urgency signal driving both rule choice and execution quality can invalidate this interpretation. Splitting trees more finely cannot establish the missing assumption. See Athey and Wager, §1.1, and Wager and Athey, §§2.1–2.3.
Three parts of the machinery
An ordinary random forest searches for features useful in predicting outcomes. A causal forest uses splits aimed at differences in effects. High prediction importance for volatility alone says nothing about the causal effect of changing volatility.
Honesty separates split selection from leaf estimation within each tree: one subsample determines the partition; a different subsample estimates the effects within it. The roles can change across trees. This limits reuse of a lucky outcome pattern in both discovering and measuring a subgroup. It does not remove unmeasured confounding. In the generalized-random-forest construction, shared leaves produce adaptive observation weights, which enter a local estimating equation. Prediction need not be the simple mean of each tree's effect estimate. The official GRF reference describes these steps.
Orthogonalisation addresses a different problem. Let and . A residual-based local moment is
Estimating the outcome and treatment models separately, using out-of-sample or out-of-bag nuisance predictions, helps isolate effect variation from baseline outcome and assignment variation. Orthogonality reduces first-order sensitivity to nuisance estimation error under its conditions; it does not make arbitrary nuisance fits harmless. Athey, Tibshirani and Wager, §§6.1.1–6.2 develop this distinction.
For the ensemble background, Bagging and Random Forests (membership; public preview) explains feature subsampling and averaging.
An effect map still needs uncertainty
A fitted surface can look precise while local treated or control information is scarce. Wager and Athey's pointwise inference result assumes independent identically distributed observations, identification and overlap, plus fixed-dimensional feature-support, smoothness, moment, tree-regularity and subsample-growth conditions. Honesty alone is insufficient. Adding trees reduces algorithmic randomness; it does not create independent observations. A confidence interval at one fixed feature value also does not automatically cover every subgroup selected after searching the surface. Dependent orders, shared market shocks or interference require a design and inference procedure that account for them; the iid theorem does not cover them automatically. See Theorem 11 and its assumptions.
Sources
- Athey and Wager, Estimating Treatment Effects with Causal Forests: An Application (2019), §1.1. Primary manuscript.
- Wager and Athey, Estimation and Inference of Heterogeneous Treatment Effects using Random Forests (JASA, 2018; linked July 2017 manuscript), §§2.1–2.4 and Theorem 11. Primary manuscript.
- Athey, Tibshirani and Wager, Generalized Random Forests (Annals of Statistics, 2019; linked April 2018 manuscript), §§2.1 and 6.1.1–6.2. Primary manuscript.
- GRF authors, The GRF Algorithm, training, prediction, honesty and orthogonalization; accessed 30 October 2026.