Orthogonal Random Forests

Ordinary causal forests estimate how a treatment effect differs across individuals by growing trees that split the data to find groups with similar effects. The problem is that if the models used to control for confounding variables have any bias — and with flexible machine-learning models they almost always do — that bias can leak straight into the effect estimate and doesn't shrink away even with more data.

Orthogonal Random Forests fix this by building "orthogonality" into the forest itself: at every split, the effect estimate within each leaf is computed using a bias-correcting formula (based on the double machine learning idea of residualizing out nuisance effects) rather than a naive average. This makes the estimate much less sensitive to small errors in the auxiliary models used to control for confounders, at the cost of extra computation per split.

Worked example

A researcher wants to know how a trading signal's effect on returns varies with a stock's volatility regime, controlling for several correlated confounders (liquidity, sector, size) using flexible gradient-boosted models. A standard causal forest's leaf-level effect estimates would inherit the small residual biases from those boosted confounder models. An Orthogonal Random Forest instead residualizes the treatment and outcome against the confounder models first at each split, so that even imperfect confounder models don't distort the final estimate of how the signal's effect differs across volatility regimes.