Data · Established · Intermediate

Data Leakage

When information unavailable at prediction time leaks into training, producing offline scores that collapse in production.

What Data Leakage is

Leakage takes many forms: a feature computed after the outcome, duplicate records spanning train and test, or target-encoded categories fitted on the full dataset.

How it works

It is prevented with time-based splits, group-aware splitting, pipelines that fit transformations inside cross-validation folds, and suspicion of any feature that looks too predictive.

Why it matters

It is the most common reason a model that scored 0.99 offline delivers no value live, and it wastes entire project cycles when caught late.

Common uses

→Model validation reviews
→Audit of feature pipelines

Watch for

✓Silent until deployment

Continue exploring

Model Evaluation Overfitting Feature Engineering Cross-Validation