Framing the prediction problem
Machine learning succeeds when the question is precise. What exactly is being predicted, how far ahead, for whom, and what will someone do with the answer? A churn score nobody acts on has no value, so we define the decision and the cost of wrong calls first, and choose metrics such as precision at a threshold or forecast error that match. Stakeholder interviews at this stage also reveal constraints, such as regulators needing explanations for every decision.
We also check whether ML is warranted. A transparent rule set or a regression on a handful of features sometimes performs as well and is easier to explain to auditors and customers. When we do recommend ML, we begin with a simple baseline so every improvement has to earn its added complexity. A baseline might be last year's value, a moving average or a logistic regression, and beating it honestly is the first milestone.
