Feature Engineering: Creating Predictive Inputs

Feature Engineering: The Art of Prep

Machine learning models are only as good as the data you feed them. Feature engineering is the process of creating useful input variables from raw price data.

Price-Based Features

• Returns: (P_t - P_{t-1}) / P_{t-1} (simple percentage change)
• Log Returns: ln(P_t / P_{t-1}) (mathematically cleaner for modeling)
• Rolling Mean: Average price over last N periods (trend)
• Rolling Std: Price volatility over last N periods
• Price vs. SMA: How far price is from its average (reversion signal)
• High/Low Range: (High - Low) / Close (intraday volatility)

import pandas as pd
import numpy as np

df = pd.read_csv('SPY.csv')

# Create features
df['Return'] = df['Close'].pct_change()
df['Log_Return'] = np.log(df['Close'] / df['Close'].shift(1))
df['SMA_20'] = df['Close'].rolling(20).mean()
df['STD_20'] = df['Close'].rolling(20).std()
df['Price_vs_SMA'] = (df['Close'] - df['SMA_20']) / df['SMA_20']
df['Range'] = (df['High'] - df['Low']) / df['Close']

Technical Indicator Features

• RSI: Momentum 0-100
• MACD: Trend-following indicator
• ATR: Volatility measure
• OBV: On-Balance Volume (accumulation/distribution)
• Bollinger Bands: Overbought/oversold zones

Important: All features should be normalized to the same scale (usually [-1, 1] or [0, 1]) so the model doesn't overweight large-magnitude features.

from sklearn.preprocessing import StandardScaler

# Normalize features to mean=0, std=1
scaler = StandardScaler()
features_scaled = scaler.fit_transform(df[['RSI', 'MACD', 'ATR']])
df['RSI_scaled'] = features_scaled[:, 0]
df['MACD_scaled'] = features_scaled[:, 1]
df['ATR_scaled'] = features_scaled[:, 2]

Volume Features

• Volume Ratio: Today's volume / 20-day average
• Volume Trend: Is volume increasing or decreasing?
• VWAP Deviation: How far from volume-weighted average price

Time Features (Seasonality)

• Day of Week: Monday, Tuesday, etc. (evidence of different momentum)
• Month: Certain months have higher returns (January Effect)
• Quarter End: End-of-period portfolio rebalancing effects

Label Creation: What Are We Predicting?

• Binary Classification: Will price be higher in 5 days? (1=yes, 0=no)
• Multi-Class: Up/Down/Flat over next 5 days
• Regression: Predict the exact return (%, continuous number)
• Direction Only: Will it go up or down? (ignore magnitude)

# Create binary label: will price be higher in 5 days?
df['Return_5d'] = df['Close'].shift(-5) / df['Close'] - 1
df['Label'] = (df['Return_5d'] > 0).astype(int)  # 1 if up, 0 if down

# Shift features so we're only using data available at decision time
for col in feature_columns:
    df[col] = df[col].shift(1)  # Shift to prevent lookahead bias

# Remove NaN rows
df = df.dropna()