leakage
**Data Leakage** is the **most insidious problem in applied machine learning — where information from outside the training dataset "leaks" into the model, producing artificially inflated performance metrics during development that collapse catastrophically in production** — occurring when the test set contaminates training (scaling before splitting, group members in both sets), when features encode the target (using "date of loan default" to predict defaults), or when future information bleeds into the past (time series shuffling), making models appear to perform miraculously in evaluation but fail completely when deployed.
**What Is Data Leakage?**
- **Definition**: Any situation where a model has access to information during training that would not be available at prediction time — resulting in unrealistically high validation scores that don't reflect actual predictive ability.
- **Why It's Dangerous**: Leakage doesn't cause errors or warnings. The model trains fine, validation metrics look excellent, and everyone celebrates — until the model is deployed and performs no better than random. By then, months of development time and money have been wasted.
- **How Common Is It?**: Extremely common. A study found that over 20% of published ML papers in top venues had some form of data leakage.
**Types of Data Leakage**
| Type | Description | Example | Fix |
|------|------------|---------|-----|
| **Target Leakage** | Feature directly encodes the target | Using "loan_default_date" to predict if a loan will default | Remove features unavailable at prediction time |
| **Train-Test Contamination** | Test data statistics leak into training | Fitting StandardScaler on all data before splitting | Split first, then preprocess (use Pipeline) |
| **Temporal Leakage** | Future data used to predict the past | Shuffling time series data in K-Fold | Use TimeSeriesSplit |
| **Group Leakage** | Same group in train and test | Same patient's X-rays in both sets | Use GroupKFold |
| **Feature Leakage** | Feature is a proxy for the target | "Treatment received" predicts disease (because only sick people get treated) | Causal analysis of features |
**Real-World Examples**
| Scenario | Leaked Information | Observed Accuracy | Real Accuracy |
|----------|-------------------|-------------------|---------------|
| Predicting hospital readmission using "number of follow-up appointments" | Follow-ups are scheduled AFTER the outcome is known | 95% | 60% |
| Fitting PCA on entire dataset, then splitting | Test data variance structure leaked into PCA | 92% | 78% |
| Predicting fraud with "account_frozen" feature | Accounts are frozen BECAUSE of fraud | 99% | 55% |
| Patient images split randomly across train/test | Model memorizes patient-specific features | 97% | 75% |
**Prevention Checklist**
| Rule | Implementation |
|------|---------------|
| **Split first, preprocess second** | Use `sklearn.pipeline.Pipeline` to chain scaler + model |
| **Time-aware splits** | TimeSeriesSplit for temporal data, never random shuffle |
| **Group-aware splits** | GroupKFold when samples are not independent |
| **Feature audit** | For each feature, ask: "Would I have this at prediction time?" |
| **Temporal feature audit** | For each feature, ask: "Was this known BEFORE the event I'm predicting?" |
| **Holdout test set** | Final evaluation on data never seen during any development step |
**The Pipeline Solution**
```python
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.ensemble import RandomForestClassifier
# Correct: preprocessing inside pipeline (no leakage)
pipe = Pipeline([
('scaler', StandardScaler()),
('model', RandomForestClassifier())
])
pipe.fit(X_train, y_train) # Scaler fits only on train data
pipe.score(X_test, y_test) # Scaler transforms test using train statistics
```
**Data Leakage is the silent killer of machine learning projects** — producing models that appear excellent during development but fail in production because they relied on information that won't be available in the real world, preventable only through disciplined pipeline design, proper temporal/group-aware splitting, and careful auditing of every feature for temporal and causal validity.