Lecture 7: Financial Applications

Learning Objectives

After completing this lecture, you should be able to:

1. Apply walk-forward validation to financial data
2. Identify and avoid common pitfalls in financial ML
3. Understand challenges specific to financial time series
4. Design features for stock prediction
5. Interpret model outputs in a trading context
6. Critically evaluate financial ML claims

Key Concepts

1. Why Financial Data is Different

Financial data presents unique challenges not found in typical ML datasets:

2. Walk-Forward Validation

Standard cross-validation is WRONG for time series because it leaks future information.

The problem with k-fold CV:

Standard k-fold: [Train|Val|Train|Val|Train]
                       ^         ^
                       Uses future data to predict past!

Walk-forward (rolling window) validation:

Period 1: [====Train====][Val]
Period 2:   [====Train====][Val]
Period 3:     [====Train====][Val]

Implementation:

1. Fix training window size (e.g., 5 years)
2. Fix test window size (e.g., 1 year)
3. Train on window, test on next period
4. Roll forward, repeat

Click chart to view Python source code

Advantages:

• Never uses future data
• Tests across multiple market regimes
• More realistic performance estimate

3. Look-Ahead Bias

Definition: Accidentally using information that wouldn’t have been available at prediction time.

Common sources:

Click chart to view Python source code

Prevention:

• Use point-in-time databases
• Strictly separate train/test by date
• Be paranoid about feature timestamps

4. Regime Changes

Markets operate in different “regimes” with different statistical properties.

Examples:

• Bull market vs bear market
• Low volatility vs high volatility
• Pre-crisis vs post-crisis

The problem: A model trained in one regime may fail completely in another.

Click chart to view Python source code

Solutions:

• Walk-forward validation (exposes regime failures)
• Include regime indicators as features
• Use ensemble of regime-specific models
• Apply strong regularization

5. Feature Engineering for Finance

Raw data is rarely predictive. Features must be engineered carefully.

Categories of features:

Feature engineering principles:

1. Stationarity: Use returns, not prices
2. Normalization: Z-score or rank within period
3. Lag appropriately: Ensure feature was available at prediction time
4. Domain knowledge: Finance experts can suggest meaningful features

6. Case Study: Stock Direction Prediction

Problem: Predict whether a stock will go up (>0% return) or down in the next month.

Architecture:

Inputs (20 features):
- 5 technical indicators
- 5 fundamental ratios
- 5 momentum features
- 5 volatility features

Hidden Layer 1: 32 neurons (ReLU)
Hidden Layer 2: 16 neurons (ReLU)
Output: 1 neuron (Sigmoid) - P(up)

Regularization: L2 (lambda=0.001), Dropout (0.3)

Training setup:

• Training: 10 years of data
• Validation: 2 years
• Test: 2 years (walk-forward)
• Batch size: 64
• Early stopping: patience=20

Expected results:

• Accuracy: 52-55% (anything above 50% is valuable!)
• Key: Small edge with many trades = significant profit

7. Transaction Costs

Real trading has costs that erode model profits:

Impact on strategy:

• A model predicting 55% accuracy with 2% average win/loss
• Before costs: Expected return = 0.552% - 0.452% = 0.2% per trade
• After 0.3% round-trip costs: 0.2% - 0.3% = -0.1% (LOSING!)

Solutions:

• Include transaction costs in loss function
• Reduce trading frequency
• Focus on larger predicted moves

8. The Efficient Market Hypothesis (EMH)

EMH states: All available information is already reflected in prices.

Three forms:

Reality for ML:

• Markets are approximately efficient
• Any predictable patterns are quickly arbitraged away
• Useful patterns exist but are small and temporary

Practical implication: Don’t expect your model to achieve 70% accuracy. Even 52% with proper risk management can be very profitable.

9. Model Interpretation in Finance

Unlike image classification, finance requires understanding WHY the model makes predictions.

Why interpretation matters:

• Regulatory requirements (explainability)
• Risk management (understand exposures)
• Model debugging (detect look-ahead bias)
• Strategy confidence (logical reasoning)

Interpretation techniques:

1. Feature importance: Which inputs drive predictions?
2. Partial dependence: How does output change with one input?
3. SHAP values: Contribution of each feature to each prediction
4. Attention weights: (for attention-based models)

10. Realistic Expectations

What neural networks CAN do in finance:

• Find subtle patterns humans miss
• Process many features simultaneously
• Adapt to changing conditions (with retraining)
• Provide consistent, unemotional signals

What neural networks CANNOT do:

• Predict the unpredictable (true random events)
• Maintain edge indefinitely (competition erodes alpha)
• Work without proper validation (easy to overfit)
• Replace domain expertise (garbage in, garbage out)

Expected performance:

• Directional accuracy: 50-55%
• Sharpe ratio improvement: 0.1-0.3
• Alpha generation: 1-5% annually (before costs)

Key Formulas

Sharpe Ratio

Sharpe = (Return - RiskFreeRate) / Volatility

Maximum Drawdown

MDD = max(Peak - Trough) / Peak

Walk-Forward Return

Total Return = Product(1 + r_i) - 1
where r_i is return in period i

Finance Application: Complete Workflow

Step-by-step process:

1. Data Collection
   • Price data, fundamentals, alternative data
   • Ensure point-in-time accuracy
2. Feature Engineering
   • Calculate technical indicators
   • Normalize features
   • Remove look-ahead bias
3. Train/Test Split
   • Walk-forward methodology
   • Multiple non-overlapping test periods
4. Model Training
   • Architecture design
   • Regularization
   • Early stopping on validation
5. Evaluation
   • Accuracy, Sharpe, drawdown
   • Transaction cost analysis
   • Regime analysis
6. Paper Trading
   • Test on live data without real money
   • Verify execution is feasible
7. Live Trading (Carefully)
   • Start with small position sizes
   • Monitor for degradation

Practice Questions

Mathematical Understanding

Q1: A model has 54% accuracy predicting daily stock direction. The average win is 1% and average loss is 1%. What is the expected daily return (ignoring costs)?

Q2: Transaction costs are 0.1% per trade. How does this affect the strategy from Q1?

Conceptual Understanding

Q3: Why is standard k-fold cross-validation inappropriate for stock prediction?

Q4: Why might a model work well in backtests but fail in live trading?

Q5: A trading model has 60% accuracy but loses money. How is this possible?

Application

Q6: Design features for predicting the next month’s return of S&P 500 stocks. List 5 features and explain why each might be predictive.

Q7: You’ve built a model that achieves 58% accuracy on your test set. What additional analysis would you do before trusting this result?

Reading List

Essential Reading

• Lopez de Prado (2018) - “Advances in Financial Machine Learning” - The definitive reference
• Aronson (2006) - “Evidence-Based Technical Analysis” - Rigorous methodology

Walk-Forward Validation

• Bailey et al. (2014) - “The Probability of Backtest Overfitting”
• Harvey et al. (2016) - “…and the Cross-Section of Expected Returns” (multiple testing)

Financial Machine Learning

• Dixon, Halperin & Bilokon (2020) - “Machine Learning in Finance”
• Heaton et al. (2016) - “Deep Learning for Finance: Deep Portfolios”

Market Efficiency

• Fama (1970) - “Efficient Capital Markets: A Review”
• Lo (2004) - “The Adaptive Markets Hypothesis”

Practical Guides

• Chan (2013) - “Algorithmic Trading: Winning Strategies and Their Rationale”

Summary

This lecture covered:

1. Financial data challenges - Non-stationarity, low signal, regime changes
2. Walk-forward validation - Proper time series validation methodology
3. Look-ahead bias - The silent killer of financial ML
4. Regime changes - Markets behave differently over time
5. Feature engineering - Domain knowledge is essential
6. Transaction costs - Can turn winners into losers
7. Realistic expectations - Small edges are still valuable

Key Takeaway: Financial ML is harder than other ML domains. Success requires combining ML expertise with deep finance knowledge and rigorous validation.

Next Lecture: Modern Networks and Future - We’ll explore modern architectures and future directions.