ML Meets Causal Inference

The one idea: A model that predicts what will happen is not the same as a model that tells you what to do. Machine learning is a powerful ingredient for measuring cause and effect, but only when the "what causes what" reasoning comes first — otherwise a brilliant prediction model quietly points you at the wrong action.

⚠️ The trap: "Our model predicts sales (or churn) with high accuracy, so we can use it to set the price or pick who gets the offer." That feels right and it's wrong. A model that nails what will happen can point you at exactly the wrong action — you have to model the effect of the change (does raising price cut buying? does the coupon flip the decision?), not the level of the outcome.

Your data science team just shipped a model that predicts churn with 91% accuracy, and someone is about to propose targeting the highest-risk customers with a retention offer. That leap — from "we can predict who leaves" to "we know what to do about it" — is where machine learning and causal inference collide, and where most companies quietly burn money. Prediction and causation are different jobs; a model brilliant at one can be useless at the other.

But the story isn't "ML bad, causal good." The modern synthesis — laid out in Chernozhukov, Hansen, Kallus, Spindler, and Syrgkanis's CausalML book — is that ML is a spectacular ingredient in causal analysis, plugged into the right slot. You won't write this code. But you will sit in meetings where someone shows you a "Double ML" estimate or an "uplift model," and…