Turning a business need into a task.
WHAT TO ESTABLISH
What decision is being made Who or what makes it What information is available at that moment What outcome would count as success
WHY THE AVAILABLE-INFORMATION QUESTION MATTERS MOST
It determines which features are legitimate, and prevents leakage.
WHAT TO CHOOSE
The prediction target, precisely.
WHY PRECISION MATTERS
Predicting whether a customer will churn requires defining churn, and the definition changes the problem entirely.
WHAT TO DECIDE
- The prediction window: how far ahead
- The population: who is scored
- The frequency: how often
WHAT A PROXY TARGET IS
Predicting something measurable that stands in for what you actually care about.
WHY THAT IS RISKY
The model optimises the proxy, and the gap between proxy and intent becomes the model's behaviour.
WHAT AN EXAMPLE LOOKS LIKE
Optimising clicks when you care about satisfaction.
WHAT TO DO ABOUT IT
Choose proxies carefully, and monitor the outcome you actually care about.
WHAT TO AGREE BEFORE BUILDING
The metric, the threshold for usefulness, and how it will be evaluated.
WHY BEFORE
Agreeing afterwards invites moving the target.
