A hospital builds a model to predict which admitted patients will be readmitted within 30 days. Candidate features include age, diagnosis, length of stay, number of prior admissions, discharge destination (home / rehab / hospice), and number of follow-up visits in the 30 days after discharge.
Which feature is target leakage?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Follow-up visits are recorded during the outcome window — a readmitted patient necessarily shows up in the hospital again. The feature is available in the historical table but not at prediction time (discharge), so it makes the model look brilliant offline and useless in production. The leakage test is always: would I know this value at the moment I need the prediction?
Prior admissions are known at discharge and legitimately predictive. Leakage is about timing, not strength.
Length of stay is final at discharge, before the 30-day window begins. It is a valid feature.
Discharge destination is decided at discharge — the prediction moment — so it is known in time. Hospice does correlate with outcomes, but that is signal, not leakage.