Causal ML and Uplift Modeling

The one idea: You don't have to hand-pick segments to find CATEs anymore — causal machine learning (meta-learners, causal forests) lets the data search for who the treatment helps most, across every feature at once. But the model has to be built to target the effect, not the outcome, or it will happily learn something else entirely.

⚠️ The trap: "Our model predicts who'll buy with 90% accuracy, so we'll target the treatment at whoever it scores highest." That feels right and it's wrong. **A model trained to predict the outcome and a model trained to predict the treatment effect are answering different questions and can rank customers in opposite orders — the strongest predictor of "will buy" is routinely unrelated, or even inversely related, to "will buy more because of the treatment."**

The treatment-effects lesson ended with a warning: don't slice into ever-smaller segments looking for CATEs without a substantive reason, because manual slicing burns statistical error fast and risks carving noise into shapes. Causal machine learning is the answer to that tension — instead of you picking two or three segments by hand, a model searches across every available feature for where the treatment effect actually varies, using enough data-driven discipline to avoid just finding noise. This lesson covers how those models work, and the trap that catches teams who bring an ordinary ML habit into a causal problem.

From CATE-by-segment to CATE-by-model

Recall CATE: E[Y1 − Y0 | X = x], the average effect within some slice defined by covariates X. The onboarding-flow example split by one variable (power user vs. casual) chosen by hand. Real targeting problems…