Frame the prediction

A model starts with a decision, not an algorithm.

8 minutes. Mission: turn business data into a useful prediction.

Start at prediction time

A sales team wants to call customers likely to accept an offer. One row is one eligible customer. The target is whether the customer accepts. Features must be available before the call. Information recorded afterward cannot help a real prediction.

FrameUnit: eligible customer. Target: accepts within 7 days. Prediction time: before first call. Action: prioritize limited calls.

This is supervised learning: examples pair inputs with known outcomes so a model can learn a mapping used on new examples.^[1]

Earn the right to model

A baseline sets the minimum useful score. For classification, predict the most frequent class. If 9% accept, always predicting “no” gets 91% accuracy and zero useful leads. High accuracy can still be worthless.

Write before revealing: for “predict invoice late payment,” name one row, target, prediction time, action, and a feature unavailable at prediction time.

Check retrieval

Win: you can reject a vague “use ML” request until its prediction contract is testable.

Primary source

Read the opening module of the scikit-learn MOOC. Keep the workflow sheet beside later lessons.

StartNext: split before learning