Ml Data Leakage Guard

Core Principle

The Golden Rule: At the exact moment of prediction in production, can I access this value from the database or compute it using only information available up to that point?

If the answer is "no" or "not completely", then data leakage exists.

Quick Reference

Critical Leakage Patterns

Pattern 1: Preprocessing Before Split

# ❌ WRONG: Leakage - fit on entire dataset
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)  # Uses test set statistics
X_train, X_test = train_test_split(X_scaled, y)

# ✅ CORRECT: Fit only on training data
X_train, X_test, y_train, y_test = train_test_split(X, y)
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)  # Fit on train only
X_test_scaled = scaler.transform(X_test)  # Transform test using train statistics

Pattern 2: Global Missing Value Imputation

# ❌ WRONG: Uses global statistics including test set
df['age'].fillna(df['age'].mean(), inplace=True)  # Global mean includes test data
X_train, X_test = train_test_split(df, y)

# ✅ CORRECT: Compute statistics on training set only
X_train, X_test, y_train, y_test = train_test_split(df, y)
train_mean = X_train['age'].mean()  # Only from training data
X_train['age'].fillna(train_mean, inplace=True)
X_test['age'].fillna(train_mean, inplace=True)  # Use train mean for test

Pattern 3: PCA/Dimensionality Reduction on Full Dataset

# ❌ WRONG: PCA learns variance structure from test set
pca = PCA(n_components=10)
X_reduced = pca.fit_transform(X)  # Includes test set variance
X_train, X_test = train_test_split(X_reduced, y)

# ✅ CORRECT: Fit PCA only on training data
X_train, X_test, y_train, y_test = train_test_split(X, y)
pca = PCA(n_components=10)
X_train_reduced = pca.fit_transform(X_train)  # Learn from train only
X_test_reduced = pca.transform(X_test)  # Apply train-learned transformation

Pattern 4: Target Encoding with Full Dataset

# ❌ WRONG: Uses target values from test set
category_means = df.groupby('category')['target'].mean()  # Includes test targets
df['category_encoded'] = df['category'].map(category_means)
X_train, X_test = train_test_split(df, y)

# ✅ CORRECT: Compute encoding only from training targets
X_train, X_test, y_train, y_test = train_test_split(df, y)
category_means = X_train.groupby('category')['target'].mean()  # Train only
X_train['category_encoded'] = X_train['category'].map(category_means)
X_test['category_encoded'] = X_test['category'].map(category_means)

Pattern 5: Feature Selection on Full Dataset

# ❌ WRONG: Feature selection sees test set
from sklearn.feature_selection import SelectKBest
selector = SelectKBest(k=10)
X_selected = selector.fit_transform(X, y)  # Uses test set for selection
X_train, X_test = train_test_split(X_selected, y)

# ✅ CORRECT: Select features using training data only
X_train, X_test, y_train, y_test = train_test_split(X, y)
selector = SelectKBest(k=10)
X_train_selected = selector.fit_transform(X_train, y_train)  # Train only
X_test_selected = selector.transform(X_test)  # Apply train-learned selection

Pattern 6: Random Split on Temporal Data

# ❌ WRONG: Random split on time series (uses future to predict past)
X_train, X_test = train_test_split(df, test_size=0.2, random_state=42)

# ✅ CORRECT: Time-based split for temporal data
split_date = '2024-01-01'
X_train = df[df['date'] < split_date]
X_test = df[df['date'] >= split_date]

Pattern 7: Future Function in Time Series Features

# ❌ WRONG: Uses future data to compute current features
df['daily_avg'] = df.groupby('date')['value'].transform('mean')  # Includes all day's data

# ✅ CORRECT: Use only past data (expanding window)
df = df.sort_values('timestamp')
df['cumulative_avg'] = df.groupby('user_id')['value'].expanding().mean().reset_index(0, drop=True)

Pattern 8: Post-Event Features

# ❌ WRONG: Feature only exists after the outcome
# Predicting loan default using "number of collection calls" as feature
# Collection calls only happen AFTER default occurs

# ✅ CORRECT: Use only pre-event features
# Use features available BEFORE the outcome: credit score, income, debt ratio, etc.

Pattern 9: Leakage in Cross-Validation

# ❌ WRONG: Preprocessing before CV split
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)
scores = cross_val_score(model, X_scaled, y, cv=5)  # Each fold sees other folds' statistics

# ✅ CORRECT: Preprocessing inside CV pipeline
from sklearn.pipeline import Pipeline
pipeline = Pipeline([
    ('scaler', StandardScaler()),
    ('model', LogisticRegression())
])
scores = cross_val_score(pipeline, X, y, cv=5)  # Scaling happens per fold

Pattern 10: Data Augmentation Leakage

# ❌ WRONG: Augment before split (test set influenced by augmented train data)
X_augmented = augment_data(X)  # Augmentation sees all data
X_train, X_test = train_test_split(X_augmented, y)

# ✅ CORRECT: Augment only training data after split
X_train, X_test, y_train, y_test = train_test_split(X, y)
X_train_augmented = augment_data(X_train)  # Augment train only
# X_test remains unchanged

Leakage Detection Checklist

After any ML preprocessing or feature engineering, verify:

1. Fit-Transform Order: Did I fit any transformer (scaler, encoder, imputer, PCA) on the full dataset before splitting?
2. Global Statistics: Did I compute any statistics (mean, std, median, mode) using the entire dataset?
3. Feature Selection: Did I select features using information from the test set?
4. Target Information: Did any feature encoding use target values from the test set?
5. Temporal Order: For time series, did I use random splits or include future information?
6. Post-Event Features: Are any features only available after the outcome occurs?
7. Cross-Validation: Did I preprocess before setting up CV folds?
8. Production Reality: At prediction time, can I actually compute this feature with available data?

The "Prediction Time" Test

For every feature and preprocessing step, ask:

When the model is deployed and receives a new data point at time T:
- Can I query this value from the database?
- Can I compute this statistic using only data available before time T?
- Does this feature require knowing the outcome I'm trying to predict?

If any answer is "NO", you have data leakage.

Examples

Example 1: Detecting Normalization Leakage

Input: Code that normalizes data before train-test split

from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()
X_normalized = scaler.fit_transform(X)
X_train, X_test, y_train, y_test = train_test_split(X_normalized, y, test_size=0.2)

Leakage Detection:

• ❌ LEAKAGE DETECTED: fit_transform on entire dataset X
• The scaler learned mean and std from the full dataset, including test set
• At prediction time, you won't have access to test set statistics
• Test set performance is artificially inflated

Corrected Code:

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)
scaler = StandardScaler()
X_train_normalized = scaler.fit_transform(X_train)  # Fit on train only
X_test_normalized = scaler.transform(X_test)  # Transform using train statistics

Example 2: Detecting Missing Value Imputation Leakage

Input: Code that fills missing values with global mean

df = pd.read_csv('data.csv')
df['income'].fillna(df['income'].mean(), inplace=True)
X_train, X_test, y_train, y_test = train_test_split(df, y, test_size=0.2)

Leakage Detection:

• ❌ LEAKAGE DETECTED: Global mean computed on entire dataset
• The mean includes test set values
• At prediction time, you won't know the test set mean
• This inflates test performance

Corrected Code:

df = pd.read_csv('data.csv')
X_train, X_test, y_train, y_test = train_test_split(df, y, test_size=0.2)
train_mean = X_train['income'].mean()  # Compute mean from training data only
X_train['income'].fillna(train_mean, inplace=True)
X_test['income'].fillna(train_mean, inplace=True)  # Use training mean for test

Example 3: Detecting Time Series Leakage

Input: Stock price prediction with random split and rolling average feature

df['rolling_avg_7d'] = df.groupby('stock')['price'].rolling(7, center=True).mean()
X_train, X_test = train_test_split(df, test_size=0.2, random_state=42)

Leakage Detection:

• ❌ LEAKAGE DETECTED - Multiple Issues:
  1. center=True in rolling window uses future prices (t+3 days) to compute feature at time t
  2. Random split on temporal data means using "tomorrow's data" to predict "yesterday"
  3. At prediction time, you can't access future prices

Corrected Code:

# Fix 1: Use backward-looking rolling window (no center=True)
df = df.sort_values(['stock', 'date'])
df['rolling_avg_7d'] = df.groupby('stock')['price'].rolling(7, min_periods=1).mean().reset_index(0, drop=True)

# Fix 2: Time-based split instead of random
split_date = '2024-01-01'
X_train = df[df['date'] < split_date]
X_test = df[df['date'] >= split_date]

Example 4: Detecting Target Encoding Leakage

Input: Category encoding using target mean from full dataset

category_target_mean = df.groupby('category')['target'].mean()
df['category_encoded'] = df['category'].map(category_target_mean)
X_train, X_test, y_train, y_test = train_test_split(df.drop('target', axis=1), df['target'])

Leakage Detection:

• ❌ LEAKAGE DETECTED: Target encoding uses test set labels
• The encoding directly uses the answer (target) from test set
• This is severe leakage - you're literally using test labels as features
• At prediction time, you don't know the target value

Corrected Code:

X_train, X_test, y_train, y_test = train_test_split(df.drop('target', axis=1), df['target'])
# Compute target mean only from training data
train_df = X_train.copy()
train_df['target'] = y_train
category_target_mean = train_df.groupby('category')['target'].mean()

X_train['category_encoded'] = X_train['category'].map(category_target_mean)
X_test['category_encoded'] = X_test['category'].map(category_target_mean)

Example 5: Detecting Post-Event Feature Leakage

Input: Predicting customer churn using "number of retention calls" as feature

features = ['account_age', 'monthly_spend', 'support_tickets', 'retention_calls_count']
X = df[features]
y = df['churned']

Leakage Detection:

• ❌ LEAKAGE DETECTED: Post-event feature
• "retention_calls_count" only exists AFTER customer shows churn signals
• Company only makes retention calls to customers who are already churning
• This is a consequence of churn, not a predictor
• At prediction time (before churn), this feature doesn't exist or is always 0

Corrected Code:

# Remove post-event features, use only pre-event features
features = ['account_age', 'monthly_spend', 'support_tickets', 'login_frequency',
            'feature_usage_decline', 'payment_delays']
X = df[features]
y = df['churned']

Leakage Severity Levels

CRITICAL (Model is completely invalid):

• Target encoding with test labels
• Post-event features
• Using future data in time series

HIGH (Significantly inflated performance):

• Preprocessing before train-test split
• Feature selection on full dataset
• Global imputation statistics

MEDIUM (Moderate performance inflation):

• Incorrect cross-validation setup
• Subtle temporal leakage in feature engineering

References

• references/leakage-patterns.md: Comprehensive catalog of leakage patterns
• references/temporal-leakage.md: Time series specific leakage issues
• references/detection-strategies.md: How to detect leakage in existing code

Maintenance

• Sources: ML best practices, Kaggle competition lessons, production ML experience
• Last updated: 2026-01-19
• Known limits: Cannot detect domain-specific leakage without context (e.g., whether a feature is post-event in a specific business context)