P26Machine LearningAdvancedDiagnose18 min

Detect Data Leakage Before Deployment

Identify a feature that is highly predictive only because it is created after the outcome.

#leakage#time-splits#deployment#validation

Scenario

A churn model reaches validation AUC 0.97. One feature is cancellation_reason, recorded when an account closes.

Problem statement

Explain why this is leakage and how to redesign the feature set and validation split.

Your Task

  1. Define what information exists at prediction time.
  2. Remove or time-shift post-outcome features.
  3. Use a split that matches deployment chronology.

Related Concepts

Temporal validationFeature contractsDeployment realism

Ready to count this problem?

Mark it complete when you have reasoned through the solution in your own words.