Machine learning predicts what will happen. Many business questions ask what would happen if we acted — which needs causal reasoning.
Prediction Isn't Causation
A model may find that customers who receive discounts churn more — because discounts are offered to at-risk customers. Using that model to conclude discounts cause churn would be wrong.
Gold Standard: Experiments
Randomised experiments (A/B tests) remove confounding and estimate causal effects directly.
Observational Methods
When experiments aren't possible:
- Matching and propensity scores: compare similar treated and untreated units.
- Difference-in-differences: compare changes over time between groups.
- Regression discontinuity: exploit sharp cut-offs in treatment rules.
- Instrumental variables: use factors that affect treatment but not outcomes directly.
- Causal graphs: make assumptions explicit and identify what to control for.
Uplift Modelling
Predict the effect of an action on each individual — who will respond because of a campaign, not just who will buy anyway.
Caution
Observational methods depend on assumptions that can't always be tested. State them clearly and check sensitivity.