Leakage quick reference
Recall common leakage patterns and the response that prevents each one.
Common objection This is our strongest predictor, so removing it will hurt the model. A teammate is defending cancellation_reason or another late workflow field. Your line If it is strongest because it sees the future or the label path, the lift is not usable. Let us build the earliest safe version of the same signal. Arguing about usefulness instead of prediction-time availability. It preserves the hypothesis while rejecting unfair evidence. Leak vs proxy Which representation can the model know before scoring? Replace post-outcome fields with earlier observable proxies. Core rule What is the fit-on-train leakage rule? Any transform that learns…
Sign up free — one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in