Data leakage usually happens when future or target-related information unintentionally enters the training data, making the model look better than it actually is.

To detect it, watch for unusually high validation performance, check features for any target-derived signals, and use time-based splits where relevant.

To mitigate it, build end-to-end pipelines so preprocessing happens only on training data, keep strict train/test separation, and validate that all features reflect what would be available in production.