Machine Learning Overview (page 13 of 13)

Target leakage occurs when a model is trained based on features or information that will not be known at prediction time. For instance, if we want to predict whether we should grant a customer a line of credit to make a purchase, then we cannot use a binary feature indicating that they made their first payment on-time on the loan. That would be target leakage because payment history is not known at the time when the loan officer is deciding whether to authorize the line of credit.