Target leakage happens when training set contains a feature carrying target information unavailable at inference time

It should be obvious that we remove the target variable from the training set’s feature list when feeding data into the model, but it’s a pretty common user error.

A less obvious example is having a feature that’s not the target variable itself, but an information ‘from the future’, that is unavailable at inference time. For example, a column called insurance_claim_filed when trying to predict car accidents.

Also called a label leakage.