Background: Time Series Cross-Validation (Purged CV)

"Using standard K-Fold on financial data is like using tomorrow's newspaper to predict today's stock price."

Why Standard Cross-Validation Fails in Finance?

Standard K-Fold Assumption: Samples are mutually independent.

Financial Data Reality:

• Today's return is highly correlated with yesterday's (autocorrelation)
• Features used to predict day 100 may include information from days 99, 98
• Labels (returns) often involve multi-day windows

Result: Information "leaks" from test set to training set, causing severe overfitting.

A Concrete Leakage Case

Scenario: Predicting 5-day future returns

Data: Days 1-100
Label: ret_5d[t] = (close[t+5] - close[t]) / close[t]

Sample 95's label uses: Days 95-100 prices
Sample 96's label uses: Days 96-101 prices (partial overlap!)

Standard K-Fold might:

• Training set includes sample 95 (label involves days 95-100)
• Test set includes sample 96 (label involves days 96-101)
• Days 96-100 information is in both training and testing!

Three Time Series CV Methods

Method 1: Simple Time Split (Walk-Forward)

Fold 1: Train [1-60]  -> Test [61-70]
Fold 2: Train [1-70]  -> Test [71-80]
Fold 3: Train [1-80]  -> Test [81-90]
Fold 4: Train [1-90]  -> Test [91-100]

Pros: Simple, no future information leakage Cons: Training set grows larger, early data may be stale

Method 2: Rolling Window

Fold 1: Train [1-60]   -> Test [61-70]
Fold 2: Train [11-70]  -> Test [71-80]
Fold 3: Train [21-80]  -> Test [81-90]
Fold 4: Train [31-90]  -> Test [91-100]

Pros: Fixed training set size, uses most recent data Cons: Low sample utilization

Method 3: Purged K-Fold (When Labels Overlap)

When observations have label intervals rather than point labels:

1. Fold definition: Define test folds from time intervals appropriate to the deployment question
2. Purge: Remove training samples whose labels overlap with test set
3. Embargo: Add safety buffer beyond the purge zone

Purging addresses one specific overlap channel. It does not repair revised data, features computed with future availability, universe survivorship, target leakage, repeated researcher access, or an unrealistic execution model.

Purged CV Explained

Problem Setup:

• Feature window: Past 20 days
• Label window: Future 5 days
• Data: Days 1-100

Steps:

1. Split Folds (assume 5 folds)
   Fold 3 test set: Days 41-60

2. Identify Leakage Zone
   Test labels involve: Days 41-65 (41+5-1 to 60+5-1)
   Training features involve: Days 21-40 may also affect test

3. Purge
   Remove training samples whose labels overlap with test set
   Remove: Days 36-40 (labels involve 36-45, overlaps with 41-65)

4. Embargo
   Remove N additional samples after the Purge boundary
   If Embargo = 5 days, remove training samples from days 41-45

Visualization:

The Role of Embargo

Why is Embargo Needed?

Even after purging label overlap, there may still be:

• Feature autocorrelation (today's MA20 and tomorrow's MA20 are nearly identical)
• Information propagation delay (news impact lasts several days)
• Market state persistence (trends don't disappear overnight)

Suggested Embargo Length:

Rule of Thumb: Embargo ≈ 0.5 x Label Window

Practical Calculation Example

Setup:

• Data: 1000 samples (4 years daily)
• Label: 10-day future return
• K = 5 folds
• Embargo = 5 days

Fold Allocation:

Note: Each fold loses about 15 samples to prevent leakage.

Comparison with Other Methods

Multi-Agent Perspective

In multi-agent systems, different Agents need different CV strategies:

Signal Agent (predicting 5-day returns):
  - Purge: 5-day label window
  - Embargo: 3 days
  - Conservative model performance estimate

Regime Agent (identifying market states):
  - Purge: Usually not needed (state is current)
  - Embargo: Longer (state transitions have inertia)
  - Focus on accuracy during state transitions

Risk Agent (predicting volatility):
  - Purge: Volatility window (e.g., 20 days)
  - Embargo: 5 days
  - Volatility clustering requires longer Embargo

Common Misconceptions

Misconception 1: Using Purged CV prevents overfitting

Wrong. Purged CV only prevents information leakage, it cannot prevent:

• Overfitting from too many features
• Data snooping (repeatedly testing until finding good results)
• Excessive model complexity

Misconception 2: Longer Embargo is always better

Not entirely true. Too long Embargo:

• Wastes effective training samples
• May make training data too stale
• Increases computational cost

Misconception 3: Only need Purged CV for final testing

Wrong. Hyperparameter tuning must also use Purged CV, otherwise you'll select overfitting parameters.

Practical Recommendations

1. Check if Purging is Needed

Need Purging when:
- Labels involve multi-day windows (e.g., future N-day returns)
- Features involve long windows (e.g., 60-day moving average)
- Samples have overlap

Less Need for Purging when:
- Labels are instantaneous (e.g., next tick direction)
- Samples are completely independent (e.g., cross-section of different stocks)

2. Validate Purge Effectiveness

Comparison Experiment:
1. Train with standard K-Fold, record test accuracy
2. Train with Purged K-Fold, record test accuracy
3. Larger difference indicates more severe original leakage

3. Reserve Completely Independent Test Set

Illustrative allocation (choose boundaries from the problem, not this percentage):
- Development period: leakage-safe folds for model selection and tuning
- Final chronological holdout: sealed before research begins and opened only after the contract is frozen

"Use once" is a governance rule, not a mathematical property. If the result influences another model change, the set has become development data and a new final evaluation is required.

Summary