Validation scores look perfect, but in production the model fails miserably. Almost always, there's one culprit — data leakage. Leakage occurs when a model gains access to information at training time that it wouldn't actually have at prediction time, artificially inflating evaluation scores.
MLS-C01 obsesses over leakage. Scenarios like "validation is 99% but production is 60%?" are almost always answered by leakage. Today we cover the causes, detection, and prevention of leakage, and two most common types: target leakage and time series leakage.