Feature Engineering for Trading ML

Feature engineering is the single highest-leverage activity in building ML trading models. Model selection (XGBoost vs. neural net vs. logistic regression) matters far less than the quality and diversity of input features. A simple model on great features will outperform a complex model on raw prices every time.

This skill covers constructing, validating, and selecting features from market data for use in classification (signal-classification) and regression models targeting crypto/Solana token trading.

Why Features Beat Models

Raw OHLCV data is non-stationary, noisy, and high-dimensional. Models trained directly on price series will overfit. Feature engineering transforms raw data into stationary, informative signals that capture distinct aspects of market behavior:

• Compression: Reduce thousands of price bars to dozens of descriptive statistics
• Stationarity: Convert non-stationary prices into stationary returns and ratios
• Domain knowledge: Encode trader intuition (support/resistance, volume climax) as computable quantities
• Regime awareness: Features that behave differently in trending vs. ranging markets help models adapt

Feature Categories

1. Price Features

Derived purely from OHLCV price columns. These capture trend, momentum, and volatility from the price series itself.

2. Volume Features

Volume confirms or contradicts price movements. Divergences between price and volume are among the most reliable signals in short-term trading.

3. Technical Features

Standard technical indicators computed via pandas-ta. Use the pandas-ta skill for full parameter documentation.

4. Microstructure Features

Derived from trade-level data (individual swaps/transactions). Require on-chain or DEX API data.

5. On-Chain Features

Derived from blockchain state changes. Require Helius or Solana RPC data.

6. Cross-Asset Features

Capture relationships between the target token and broader market.

7. Time Features

Cyclical encoding of calendar time. Use sin/cos encoding to preserve cyclical continuity (hour 23 is close to hour 0).

import numpy as np

hour_sin = np.sin(2 * np.pi * hour / 24)
hour_cos = np.cos(2 * np.pi * hour / 24)
day_of_week = np.sin(2 * np.pi * day / 7)

Stationarity

Non-stationary features will cause your model to fail on new data. A feature is stationary if its statistical properties (mean, variance) don't change over time.

Testing for Stationarity

Use the Augmented Dickey-Fuller (ADF) test:

from scipy.stats import adfuller

result = adfuller(feature_series.dropna())
p_value = result[1]
is_stationary = p_value < 0.05

Making Features Stationary

Rule: If a feature trends upward or downward over time, it is non-stationary. Transform it into a ratio, difference, or rate of change.

Normalization

After computing features, normalize them so that all features have comparable scales. This is critical for distance-based models (KNN, SVM) and helpful for tree models.

Critical: Use rolling statistics for normalization. Never use full-sample mean/std — that introduces lookahead bias.

# CORRECT: rolling z-score
z = (feature - feature.rolling(60).mean()) / feature.rolling(60).std()

# WRONG: full-sample z-score (lookahead bias!)
z = (feature - feature.mean()) / feature.std()

No-Lookahead Guarantee

The most dangerous bug in trading ML is lookahead bias — using future information to compute features or targets. Follow these rules absolutely:

1. Rolling calculations only: Never use .mean() or .std() on the full series. Always use .rolling(N).mean().
2. Shift targets forward, not features backward: The target is close.shift(-N) / close - 1 (future return), not close / close.shift(N) - 1 (past return used as target).
3. No future index alignment: When joining feature and target DataFrames, verify that feature row t is paired with target row t (where target already contains the forward shift).
4. Train/test split by time: Never random split. Always train = data[:split_idx], test = data[split_idx:].

Feature Selection

After computing many features, select the most predictive and least redundant:

Step 1: Remove Low-Variance Features

from sklearn.feature_selection import VarianceThreshold
selector = VarianceThreshold(threshold=0.01)
X_filtered = selector.fit_transform(X)

Step 2: Correlation Filter

Remove features with > 0.9 correlation to another feature (keep the one with higher target correlation):

corr_matrix = X.corr().abs()
upper = corr_matrix.where(np.triu(np.ones(corr_matrix.shape), k=1).astype(bool))
to_drop = [col for col in upper.columns if any(upper[col] > 0.9)]

Step 3: Feature Importance

Train a random forest and rank by importance:

from sklearn.ensemble import RandomForestClassifier
rf = RandomForestClassifier(n_estimators=100, random_state=42)
rf.fit(X_train, y_train)
importances = pd.Series(rf.feature_importances_, index=X.columns).sort_values(ascending=False)

Step 4: Mutual Information

Non-linear alternative to correlation:

from sklearn.feature_selection import mutual_info_classif
mi = mutual_info_classif(X_train, y_train, random_state=42)
mi_scores = pd.Series(mi, index=X.columns).sort_values(ascending=False)

Label Creation

Labels (targets) define what the model learns to predict.

Binary Classification

forward_return = close.shift(-N) / close - 1
label = (forward_return > threshold).astype(int)  # 1 = up, 0 = not up

Typical thresholds: 1% for 1h bars, 3% for 4h bars, 5% for daily bars.

Multi-Class Classification

label = pd.cut(forward_return,
               bins=[-np.inf, -threshold, threshold, np.inf],
               labels=[0, 1, 2])  # 0=down, 1=flat, 2=up

Regression

target = forward_return  # Predict exact return magnitude

Binary classification is recommended for initial models — it's simpler and more robust to noise.

Integration with Other Skills

• pandas-ta: Compute technical indicators that become features
• birdeye-api: Fetch OHLCV and trade data for feature computation
• helius-api: Fetch on-chain data for holder/whale features
• signal-classification: Use engineered features as model inputs
• regime-detection: Regime labels as features or for regime-conditional models
• ohlcv-processing: Clean and resample raw data before feature computation

Files

References

• references/feature_catalog.md — Complete catalog of ~40 features with formulas, lookbacks, stationarity status, and interpretation notes
• references/pitfalls.md — Common mistakes in trading feature engineering: lookahead bias, overfitting, survivorship bias, data snooping, non-stationarity

Scripts

• scripts/build_features.py — Compute 25+ features from OHLCV data with stationarity testing and quality reporting. Supports demo mode with synthetic data or live data via Birdeye API.
• scripts/feature_importance.py — Rank features by predictive power using tree-based importance and permutation importance. Identifies redundant features via correlation analysis.