Data Leakage
When information from outside the training set sneaks into it, producing a model that looks brilliant in testing and fails in the real world — the most common way ML projects fool their own builders.
Reviewed July 16, 2026Stable
When not to use it
- (Not applicable — leakage is a failure to avoid, not a technique to use. The "when" is: always guard against it.)
Reach for something else instead
- Strict pipeline discipline — fit all preprocessing inside the training fold only.
- Forward-chaining validation for time series, so you never train on the future.
- Feature auditing — check every feature for availability at prediction time.
Sources & further reading
- Kaufman et al. (2012), Leakage in Data Mining — the definitive formulation and taxonomy.
- Kapoor & Narayanan (2023), Leakage and the Reproducibility Crisis in ML-based Science — how pervasive leakage undermines published results.
- Google, Rules of Machine Learning — practitioner guidance on avoiding training/serving skew.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Preprocessing (scaling, imputing) before the train/test split, so test statistics leak in.
- Including a feature that's a proxy for or caused by the target.
- Random-splitting time-series data, training the model on information from the future.
At a glance
FieldMachine Learning
What it isoutside info contaminating training
Symptomresults too good to be true
Fixsplit first, audit every feature
DifficultyIntermediate