Leakage happens when training data includes information unavailable at prediction time.

Leakage creates unrealistically strong offline metrics and weak production performance because the model learned future or target-derived signals.