Decide & act

Causal inference & uplift

promo targeting, retention offers, measuring a launch

Decide who to treat, and measure whether the treatment worked.

Potential outcomes

Hypothesis testing

Probability & Bayes

Foundations

Potential outcomes define what an effect is; hypothesis testing and power decide whether you can measure one.

Frame the problem

Frame

This is an intervention, not a prediction. The question is who responds, not who converts.

Sourcing & signal

Exploring the data

Data & labels

If an experiment ran, the treatment assignment is the most important column. If not, record every variable that influenced who got treated.

Encode & scale

Create features

Represent

Split the data

Split

Stratify on treatment so both arms appear in every fold.

Randomised data

Uplift model to pick responders.

Uplift modelling

Gradient boosting

Observational data

Propensity, matching, difference-in-differences. Be honest about confounding.

Causal inference

Aggregate data

Multi-touch attribution and marketing mix models answer 'which channel' when experiments are not possible.

Attribution & marketing mix models

Model

Nothing unusual here.

Train

Nothing unusual here.

Error analysis & significance

Evaluate offline

Qini and uplift curves, not accuracy. Placebo tests for the observational methods.

A/B testing & interleaving

Off-policy evaluation

Evaluate online

Randomise if you can, with a power calculation before launch. Guardrail metrics; novelty effects. Off-policy evaluation from logs.

Serving & release

Ship & monitor