Building Better Inputs: The Foundation of AI Trading Success
The Best Algorithm in the World Can't Fix Bad Inputs
Here's a secret that separates amateur AI traders from professionals:
The model architecture matters far less than the features you feed it.
Feature engineering - the art of transforming raw data into meaningful inputs — is where the real edge lives.
---
What Is Feature Engineering?
Definition:
Feature engineering is the process of creating input variables (features) from raw data that help machine learning models make better predictions.
In Trading:
Transforming raw price, volume, and other data into signals that capture market behavior.
The Core Principle:
Raw data (OHLCV) contains information, but it's often hidden. Features extract and highlight that information.
---
Why Feature Engineering Matters
1. Models Learn from Features, Not Raw Data
A model seeing learns little
A model seeing learns patterns
Features encode the relationships that matter
2. Domain Knowledge Becomes Computable
"Price is extended" → Z-score feature
"Volume is unusual" → Volume ratio feature
"Trend is strong" → ADX feature
3. Reduces Noise, Amplifies Signal
Raw prices contain noise
Well-designed features filter noise
Model focuses on what matters
4. Enables Simpler Models
Good features + simple model often beats
Bad features + complex model
Interpretability improves
---
Categories of Trading Features
Category 1: Price-Based Features
Returns:
Simple returns: (Close - Previous Close) / Previous Close
Log returns: ln(Close / Previous Close)
Multi-period returns: 5-day, 20-day, 60-day
Price Relationships:
Distance from high/low
Distance from moving average
Price relative to range (where in today's range)
Candle Features:
Body size: |Close - Open|
Upper wick: High - max(Open, Close)
Lower wick: min(Open, Close) - Low
Body to range ratio
---
Category 2: Trend Features
Moving Average Features:
Price above/below MA (binary)
Distance from MA (continuous)
MA slope (trend direction)
MA crossover signals
Trend Strength:
ADX value
Consecutive higher highs/lower lows
Linear regression slope
R-squared of price trend
Trend Duration:
Bars since trend started
Bars since last MA cross
Time in current regime
---
Category 3: Momentum Features
Oscillators:
RSI value
RSI zone (oversold/neutral/overbought)
Stochastic %K and %D
CCI value
Rate of Change:
ROC over various periods
Momentum acceleration/deceleration
Momentum divergence from price
Relative Momentum:
Performance vs benchmark
Sector relative strength
Percentile rank of momentum
---
Category 4: Volatility Features
Range-Based:
ATR (Average True Range)
ATR ratio (current ATR / historical ATR)
Range expansion/contraction
Standard Deviation:
Rolling standard deviation
Bollinger Band width
Z-score of price
Volatility Regime:
High/low volatility classification
Volatility percentile
Volatility trend (increasing/decreasing)
---
Category 5: Volume Features
Volume Ratios:
Volume / Average volume
Volume trend (increasing/decreasing)
Relative volume by time of day
Price-Volume Relationships:
Up volume vs down volume
Volume on up days vs down days
OBV (On-Balance Volume)
Volume-price trend
Volume Patterns:
Volume spike detection
Volume climax signals
Accumulation/distribution
---
Category 6: Time Features
Calendar Features:
Day of week
Month of year
Quarter
Days until/since events (earnings, FOMC)
Session Features:
Time of day
Session (Asian, European, US)
Minutes since open/until close
Cyclical Encoding:
Sin/cos transformation for cyclical features
Preserves cyclical relationships
Day of week: sin(2π × day/7), cos(2π × day/7)
---
Category 7: Cross-Asset Features
Correlation Features:
Rolling correlation with benchmark
Correlation regime changes
Beta to market
Relative Features:
Spread between assets
Ratio between assets
Relative performance
Market Context:
VIX level and change
Sector performance
Market breadth
---
Feature Engineering Best Practices
Practice 1: Normalize Features
Raw values vary wildly across assets and time.
Methods:
Z-score: (value - mean) / std
Min-max scaling: (value - min) / (max - min)
Percentile ranking
Why: Models work better with normalized inputs.
---
Practice 2: Handle Look-Ahead Bias
Features must only use data available at prediction time.
Common Mistakes:
Using future data in calculations
Normalizing with full dataset statistics
Including target information in features
Solution: Always use rolling/expanding windows.
---
Practice 3: Create Interaction Features
Combine features to capture relationships.
Examples:
RSI × Trend direction
Volume ratio × Price change
Volatility × Momentum
Why: Captures conditional relationships.
---
Practice 4: Lag Features Appropriately
Include historical values of features.
Examples:
RSI 1 bar ago, 5 bars ago, 20 bars ago
Return over last 1, 5, 20, 60 periods
Volatility change over time
Why: Captures temporal patterns.
---
Practice 5: Test Feature Importance
Not all features help. Some hurt.
Methods:
Correlation with target
Feature importance from tree models
Ablation studies (remove and test)
Why: Reduces overfitting, improves interpretability.
---
AI-Powered Feature Engineering
1. Automated Feature Generation
AI can generate thousands of feature combinations:
Mathematical transformations
Interaction terms
Lagged versions
2. Feature Selection
AI identifies which features actually help:
Removes redundant features
Identifies most predictive features
Optimizes feature set for model
3. Dynamic Feature Weighting
AI adjusts feature importance over time:
Some features work better in certain regimes
Adaptive weighting based on recent performance
Regime-specific feature sets
4. Deep Learning Feature Extraction
Neural networks can learn features automatically:
Convolutional layers for pattern detection
Recurrent layers for sequence patterns
Attention mechanisms for importance weighting
---
Feature Engineering Mistakes
Too Many Features - More features ≠ better model. Overfitting risk increases. Curse of dimensionality. Start simple, add complexity only if needed.
Highly Correlated Features - Multiple features measuring the same thing. Redundancy without benefit. Check correlation matrix, remove duplicates.
Unstable Features - Features that change dramatically with small data changes. Unreliable in live trading. Test stability across time periods.
Ignoring Domain Knowledge - Letting AI generate features without trading logic. May find spurious patterns. Combine AI generation with human curation.
Not Testing Out-of-Sample - Features that work in-sample may fail out-of-sample. Always validate on unseen data.
---
Building Your Feature Library
Step 1: Start with Fundamentals
Returns (multiple timeframes)
Volatility (ATR, std dev)
Trend (MA relationships)
Momentum (RSI, ROC)
Volume (ratios, trends)
Step 2: Add Domain-Specific Features
What do you look for when trading?
Encode your analysis into features
Test if they add predictive value
Step 3: Create Derived Features
Combinations of base features
Regime indicators
Cross-asset relationships
Step 4: Continuously Refine
Monitor feature performance
Remove degraded features
Add new features as markets evolve
---
Key Takeaways
Feature engineering transforms raw data into meaningful model inputs
Good features often matter more than model complexity
Categories include price, trend, momentum, volatility, volume, time, and cross-asset
Always normalize, avoid look-ahead bias, and test out-of-sample
AI can automate feature generation and selection, but domain knowledge guides the process
---
Your Turn
What features do you find most predictive in your trading?
Have you experimented with creating custom features?
Share your feature engineering insights below 👇
Education
